Paper deep dive
MADDIE: Multi-Agent Drug Discovery Engine
Ye Liu, Reza Averly, Avery Meyer, Frazier N. Baker, Ruoxi Gao, Daniel Adu-Ampratwum, Xia Ning
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Drug discovery remains a resource-intensive and time-consuming process, involving billions of dollars and a decade of development to advance a drug to market. While many in silico tools have accelerated this process, operating them still demands substantial time, effort, and expertise from medicinal chemists. Autonomous systems built on Large Language Models (LLMs) have emerged as a promising solution, yet current designs limit their performance. Existing single-agent systems scale poorly to long-horizon, multi-step, and realistic tasks, whereas multi-agent systems typically execute predefined workflows that cannot capture the iterative, non-linear nature of a real drug discovery campaign. Here, we present MADDIE, a multi-agent drug discovery engine. It comprises a central agent, Orchestrator, and four specialized worker agents: HitIdentifier, LeadOptimizer, Synthesizer, and Evaluator, each representing a critical step in a drug discovery campaign. Under Orchestrator's orchestration and coordination, these worker agents are composed into flexible workflows, enabling MADDIE to adapt its strategy and navigate the drug discovery pipeline end-to-end. Across 50 protein targets, MADDIE produces molecules satisfying key property requirements and their synthesis routes, outperforming existing agentic systems, generative models, and general-purpose LLMs by a wide margin. Our analysis reveals how these agents collaborate and that MADDIE can employ different strategies across targets to achieve the campaign's goal, while still being efficient in time and tokens. Together, they demonstrate MADDIE as a low-cost, effective, and efficient autonomous engine for accelerating drug discovery. Code and dataset is available via this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.16975v1
- Canonical: https://arxiv.org/abs/2610.16975v1
Trouble viewing inline? Open PDF directly →
Full Text
173,862 characters extracted from source content.
Expand or collapse full text
MADDIE MADDIE : Multi-Agent Drug Discovery Engine Ye Liu Email: liu.12989@osu.edu Affiliation: Biomedical Informatics, The Ohio State University Reza Averly Email: averly.1@osu.edu Affiliation: Computer Science and Engineering, The Ohio State University Avery Meyer Email: meyer.1938@osu.edu Affiliation: Biomedical Informatics, The Ohio State University Frazier N. Baker Email: baker.3239@osu.edu Affiliation: Computer Science and Engineering, The Ohio State University Ruoxi Gao Email: gao.2396@osu.edu Affiliation: Computer Science and Engineering, The Ohio State University Daniel Adu-Ampratwum Email: adu-ampratwum.1@osu.edu Affiliation: Medicinal Chemistry and Pharmacognosy, The Ohio State University Xia Ning Email: ning.104@osu.edu Affiliation: Biomedical Informatics, The Ohio State University Affiliation: Computer Science and Engineering, The Ohio State University Affiliation: Medicinal Chemistry and Pharmacognosy, The Ohio State University Affiliation: Translational Data Analytics Institute, The Ohio State University Abstract Drug discovery remains a resource-intensive and time-consuming process, involving billions of dollars and a decade of development to advance a drug to market. While many in silico tools have accelerated this process, operating them still demands substantial time, effort, and expertise from medicinal chemists. Autonomous systems built on Large Language Models (LLMs) have emerged as a promising solution, yet current designs limit their performance. Existing single-agent systems scale poorly to long-horizon, multi-step, and realistic tasks, whereas multi-agent systems typically execute predefined workflows that cannot capture the iterative, non-linear nature of a real drug discovery campaign. Here, we present MADDIE MADDIE , a multi-agent drug discovery engine. It comprises a central agent, Orchestrator Orchestrator , and four specialized worker agents: HitIdentifier HitIdentifier , LeadOptimizer LeadOptimizer , Synthesizer Synthesizer , and Evaluator Evaluator , each representing a critical step in a drug discovery campaign. Under Orchestrator Orchestrator ’s orchestration and coordination, these worker agents are composed into flexible workflows, enabling MADDIE MADDIE to adapt its strategy and navigate the drug discovery pipeline end-to-end. Across 50 protein targets, MADDIE MADDIE produces molecules satisfying key property requirements and their synthesis routes, outperforming existing agentic systems, generative models, and general-purpose LLMs by a wide margin. Our analysis reveals how these agents collaborate and that MADDIE MADDIE can employ different strategies across targets to achieve the campaign’s goal, while still being efficient in time and tokens. Together, they demonstrate MADDIE MADDIE as a low-cost, effective, and efficient autonomous engine for accelerating drug discovery. Code and dataset is available via https://github.com/ninglab/MADDIE. †equal-contributors: These authors contributed equally to this work.†equal-contributors: These authors contributed equally to this work. 1 Introduction Drug discovery is a complex, capital-intensive, and time-consuming process: bringing one approved drug to the market typically costs more than one billion of dollars and takes more than a decade (Wouters et al., 2020; Sertkaya et al., 2024; Mulcahy et al., 2025; Singh et al., 2023). To expedite this process, many in silico tools have been developed (Trott and Olson, 2010b; Yu et al., 2023; Swanson et al., 2024; Chen et al., 2025; Dey et al., 2025; Gao et al., 2026), each designed for a specific task in the pipeline (e.g., predicting binding affinity). Using them, however, has traditionally fallen to medicinal chemists. Autonomous systems capable of intelligently using these tools have gained considerable attention recently (Averly et al., 2025; Liu et al., 2025; Pan et al., 2025; Gao et al., 2025b; Wang et al., 2025; Cao et al., 2026; Kantorow et al., 2026; Vichentijevikj et al., 2026; Ock et al., 2026), substantially reducing the time and effort demanded of medicinal chemists while showing promising results. These systems are typically built upon Large Language Models (LLMs), reasoning engines capable of intelligent decision-making and planning over complex tasks (Singh et al., 2025). LIDDIA (Averly et al., 2025) is one of the first such systems: a single-agent capable of autonomously navigating drug discovery in silico. Following LIDDIA, several systems have been introduced, with more attention towards multi-agent systems (Ock et al., 2026; Liu et al., 2025; Wang et al., 2025; Gao et al., 2025b; Vichentijevikj et al., 2026). Some of them are designed for a specific task under the drug discovery pipeline, while others for end-to-end drug discovery (Gao et al., 2025b; Vichentijevikj et al., 2026). While promising, these agentic systems leave room for improvements. Single-agent systems such as LIDDIA (Averly et al., 2025) and AgentD (Ock et al., 2026) demonstrate how LLMs can intelligently coordinate different in silico tools for drug discovery, yet their performance does not scale well to more realistic and complex tasks (Section 4.1). Multi-agent systems such as PharmAgents (Gao et al., 2025b), Prompt-to-Pill (Vichentijevikj et al., 2026) and F.A.D.E. (Kantorow et al., 2026) are designed for complex tasks, but they typically assume a predefined workflow. This assumption does not reflect real-world drug discovery, which is a highly non-linear process governed by many requirements. For instance, successful drugs must bind well to the target, while maintaining favorable pharmacokinetic and physicochemical property profiles (D. Segall, 2012). However, identifying such molecules is highly difficult, requiring numerous iterations of searching the expansive chemical space (Bohacek et al., 1996). Other multi-agent systems have also been proposed that enable more flexible workflows but are still limited in adaptability (Pan et al., 2025; Cao et al., 2026). Building upon these insights, we propose MADDIE MADDIE , a Multi-Agent Drug DIscovery Engine. As illustrated in Figure 1, the system is composed of five agents, a central agent called (1) Orchestrator Orchestrator , and four specialized worker agents: (2) HitIdentifier HitIdentifier ; (3) LeadOptimizer LeadOptimizer ; (4) Synthesizer Synthesizer ; and (5) Evaluator Evaluator . Orchestrator Orchestrator serves as an intelligent orchestration component that governs the whole campaign. Unlike static workflows, Orchestrator Orchestrator dynamically determines the workflow through its actions, sustains a long-horizon campaign using its memory module and monitors the progress by maintaining its evolving repository of candidate molecules. Under its orchestration, MADDIE MADDIE ’s worker agents are composed into flexible workflows, enabling MADDIE MADDIE to adapt its strategy as a campaign unfolds, mirroring the iterative, decision-driven nature of real-world drug discovery. Each worker agent carries out a step critical in advancing a drug discovery campaign—from identifying hits and optimizing them, to proposing synthesis routes and evaluating the quality of molecules. MADDIE MADDIE ’s worker agents are equipped with state-of-the-art in silico tools, including generative models for de novo molecular generations, thus allowing MADDIE MADDIE to explore beyond known chemical space and propose novel molecules designed for each campaign’s objectives. Figure 1: MADDIE MADDIE architecture, consisting of Orchestrator Orchestrator , workspace and four worker agents: HitIdentifier HitIdentifier , LeadOptimizer LeadOptimizer , Synthesizer Synthesizer and Evaluator Evaluator . We rigorously benchmark MADDIE MADDIE with a wide range of methods designed to accelerate drug discovery. We highlight several insights from MADDIE MADDIE here: (1) MADDIE MADDIE outperforms existing baselines by a large margin, generating molecules satisfying key pharmaceutical properties required for successful drugs across 50 targets, along with providing their synthesis routes (Section 4.1). This is consistent with our motivation for the design of MADDIE MADDIE : an autonomous system that can scale to realistic tasks and adapt throughout a campaign to generate satisfying molecules across targets; (2) MADDIE MADDIE ’s performance arises from the collaborative efforts of each agent, emphasizing their importance in contributing to the satisfying molecules (Section 4.2). Each worker agent makes a distinct contribution, while Orchestrator Orchestrator contributes the most. Notably, Orchestrator Orchestrator tracks the campaign through its evolving repository of molecules and directs MADDIE MADDIE ’s effort towards the most promising ones, substantially reducing the number of molecules MADDIE must explore to reach its objective; (3) MADDIE MADDIE demonstrates its capability to adapt its workflow during campaigns for varying targets, employing distinct strategies required to produce the satisfying molecules (Section 4.3). Even on a target with no existing marketed FDA-approved drugs (Zuo and Chen, 2024), MADDIE MADDIE can identify more than 100 promising molecules through its intelligent adaptation, emphasizing its capability; (4) MADDIE MADDIE generates one promising molecule for PSEN1, a critical target associated with Alzheimer’s Disease, with the molecule’s in silico properties exceeding those of the reference ligand and the associated FDA-approved drug (Section 4.4); and (5) MADDIE MADDIE serves as a low-cost and efficient autonomous drug discovery system, highlighting its potential to accelerate drug discovery (Section 4.5). Together, these findings indicate that dynamic orchestration of specialized worker agents through a central agent offers an effective design for autonomous drug discovery, and MADDIE MADDIE can serve as a promising system in this domain. 2 Related Works 2.1 Agents for Drug Discovery Agents powered by LLMs are capable of coordinating in silico tools across broader drug discovery workflows. Existing systems operate at different levels of scope and autonomy, from task-specific assistants to end-to-end drug discovery. These systems are typically classified into two categories: (1) single-agent system and (2) multi-agent system. CACTUS (McNaughton et al., 2024) focuses on assistive cheminformatics tasks such as estimating molecular properties, retrieving structurally similar molecules, and evaluating drug-likeness. TxGemma (Wang et al., 2025) also proposes a conversational assistant for therapeutics predictions and question answering, allowing better interaction with medicinal chemists. AgentD (Ock et al., 2026) spans a wider range of the early-stage computational pipeline in drug discovery, from target and ligand information retrieval, to molecular generation, property prediction, and refinement. LIDDIA LIDDIA (Averly et al., 2025) provides an autonomous workflow where LLM-based reasoning guides the iterative process of finding satisfying molecules that fit a specific profile. These examples demonstrate that LLMs can coordinate in silico tools toward a larger drug discovery goal. However, their performance is typically limited and does not scale well for more complex and realistic tasks (Section 4.1). Multi-agent systems distribute their tasks across specialized agents, enabling them to tackle complex problems, such as drug discovery, that can exceed the reach of any single-agent system. Their system designs and assigned roles for the agents vary among systems. DrugAgent (Liu et al., 2025) coordinates planner and instructor agents to develop, implement, and refine machine-learning solutions using drug discovery domain knowledge. Robin (Ghareeb et al., 2026) assigns literature-search and data-analysis to separate agents, using their output to loop through hypothesis formulation, experiment planning, and result interpretation. More directly focused on target binding site identification and candidate generation, F.A.D.E. (Kantorow et al., 2026) uses a hierarchical multi-agent setup to produce ranked candidates and demonstrates its effectiveness through case studies on targets EGFR and CRBP1. Prompt-to-Pill (Vichentijevikj et al., 2026) and PharmAgents (Gao et al., 2025b) work at a broader scope, coordinating agents across drug discovery phases, from preclinical stage to clinical-trial simulation. Mozi (Cao et al., 2026) proposes a dual-layer multi-agent architecture for drug discovery, a layer for control and governance and another one for workflow. FROGENT (Pan et al., 2025) shows an example of a decentralized multi-agent system, using an orchestrator agent that decomposes the task into a high-level plan, and the other agents to execute each subtask in the plan. Together, these systems demonstrate the overall capability of collaboration between different specialized agents on drug discovery tasks. However, their coordination generally centers on predefined workflows, and their system evaluations often focus on individual case studies. This is in contrast to MADDIE MADDIE , which enables flexible workflow adaptation across different campaigns (Section 4.3). 2.2 Generative Models for Drug Design Generative models have recently been employed for de novo drug design to accelerate drug discovery. These models generally fall under two categories: (1) structure-based drug design (SBDD) and (2) ligand-based drug design (LBDD). SBDD methods generate ligand molecules conditioned on the three-dimensional structures of target binding pockets. Early approaches such as Pocket2Mol (Peng et al., 2022) use autoregressive sampling to generate molecular atoms sequentially, while more recent approaches are diffusion-based (Guan et al., 2023; Gu et al., 2024). Notably, conDitar-dev (Gao et al., 2026) decouples protein encoding from the diffusion process through a separately pretrained pocket encoder and performs inference-time noise optimization to improve molecular properties beyond binding affinity. Several works augment SBDD models with LLM-based refinement (Gao et al., 2025a; Hu et al., 2026b; Hu et al., 2026a). ELILLM (Hu et al., 2026a), one of the state-of-the-art methods, encodes the initial molecules into the LLM latent space and performs Bayesian optimization to identify promising molecular representations. LBDD generates molecules conditioned on the shape of the ligand that binds to the target structure’s pocket. Previous approaches include SQUID (Adams and Coley, 2023), using a fragment-based generative model to decode fragments and a scoring model to adjust the angles of rotatable bonds between fragments, and DESERT (Long et al., 2022), an image-to-sequence model with an encoder-decoder architecture that generates molecular structures conditioned on voxelized shapes. DiffSMol (Chen et al., 2025), one of state-of-the-art methods for LBDD, encodes the ligand shape into pretrained embeddings and uses these embeddings to condition a diffusion-based molecular generation model. Together, these generative models provide effective means of finding promising molecules. Generation alone, however, is only the first step of drug discovery, and the molecules these models produce often demand substantial manual downstream filtering and selection (Zhavoronkov et al., 2019; Gao et al., 2026). MADDIE MADDIE addresses this gap by integrating the filtering and selection into the system itself: HitIdentifier HitIdentifier and LeadOptimizer LeadOptimizer propose the molecules, while Evaluator Evaluator and Synthesizer Synthesizer assess them—all under Orchestrator Orchestrator ’s coordination. 3 Experimental Settings Dataset We curate a dataset to evaluate MADDIE MADDIE . Concretely, we collect (1) a set of 50 protein targets, along with their 3D binding pocket structures, and (2) a set of drugs approved by U.S. Food and Drug Administration (FDA), each is associated with a target in our dataset. These drugs are used to define what molecular property requirements MADDIE MADDIE ’s generated molecules should meet (Table 1). We provide the full protein targets, their annotations, the curated FDA-approved drugs, and the protocol to collect them in this paper (Appendix 8). The targets span major therapeutic areas including autoimmune disease, cancer, cardiovascular disease, diabetes, infectious disease, and neurological conditions. The dataset includes 42 targets from existing datasets: (1) LIDDIA (Averly et al., 2025) (12 targets) and (2) CrossDocked2020 (Francoeur et al., 2020) (30 targets), and additional 8 targets from OpenTargets (Buniello et al., 2025). We ensure that these targets do not exist in CrossDocked2020 training set, thus allowing MADDIE MADDIE to be compared against methods trained on CrossDocked2020. We use the experimentally validated 3D structures available (Berman et al., 2000) and preprocess them to extract the binding pockets, similar to LIDDIA’s dataset preprocessing (Averly et al., 2025). For each target, we identify whether it has drugs that are (1) small molecules, (2) FDA-approved, and (3) currently marketed as of July 2026. Out of 50 targets, 22 targets are associated with at least one drug. For the remaining targets (28), no drug associations are identified, indicating opportunities for drug development for these targets. We use these drugs to calculate the property requirements for our experiment (Appendix 9). Evaluation Metrics We use several metrics to evaluate the performance of MADDIE MADDIE . These metrics are based on (1) the molecular properties often considered in real-world drug discovery, (2) the synthesizability of the molecule, and (3) whether the molecule can satisfy multiple property and synthesizability requirements simultaneously. We evaluate the generated molecules on several key properties critical for a successful drug. We first consider properties commonly used in the literature (Gao et al., 2026): (1) binding affinity, measured by Vina docking score (Trott and Olson, 2010b) ( Vina ); (2) drug-likeness, or desirability of a molecule using quantitative estimate of drug-likeness (Bickerton et al., 2012) ( QED ); and (3) synthesizability estimate, using synthetic accessibility score (Ertl and Schuffenhauer, 2009) ( SAS ). In addition to common properties, we further assess the toxicities of these molecules. Concretely, we measure the probability that a molecule may (1) induce mutations in DNA (Xu et al., 2012) ( Mut ), (2) fail clinical trials owing to toxicity (Gayvert et al., 2016) ( CTx ), and (3) inhibit the hERG potassium channel (Wang et al., 2016) ( hERG ). Finally, we apply Lilly medicinal filters (Bruns and Watson, 2012) ( Lly ) to flag whether the molecules have undesirable traits, such as interfering with assay measurements (e.g., fluorescence) or lacking druggability (e.g., molecules lacking both oxygen and nitrogen). We assess the synthesizability of a molecule by determining whether a synthesis route can be generated ( Syn ). We employ the same approach as Synthesizer Synthesizer in MADDIE MADDIE to measure the synthesizability of the molecule (Section 7.2.3). This is a more rigorous assessment than synthesizability estimate ( SAS ; Section 7.2.4). For every molecular property and synthesizability defined previously, we identify a threshold as a requirement for a molecule to meet. These thresholds are determined using existing FDA-approved drugs curated in our dataset as references (Appendix 9), ensuring that they are realistic rather than arbitrary cutoffs. We define a molecule as satisfying only if it jointly meets all requirements. We report the performance of MADDIE MADDIE using several metrics: (1) Ntout $N_tout$ , the number of targets for which at least one output molecule is produced, as a measure of MADDIE MADDIE ’s coverage in delivering molecules for the targets; (2) Nout $N_out$ and Nsat $N_sat$ , the average number of output and satisfying molecules per target, respectively; (3) %Pass Pass , the average percentage of satisfying molecules among the output molecules per target; and (4) the per-requirement pass rates (% Vina , % QED , % SAS , % Mut , % CTx , % hERG , % Lly , % Syn ), by measuring the average percentage of output molecules satisfying each respective requirement per target. Baselines Table 1: The thresholds used in the experiments and whether each method can measure the corresponding properties and synthesizability during generation (✓ indicates it can). The thresholds are calculated using marketed FDA-approved drugs in our dataset (Appendix 9). Method Vina QED SAS Mut CTx hERG Lly Syn <<-6.53 >>0.36 <<5.35 <<0.51 <<0.71 <<0.96 True True Agents for Drug Discovery MADDIE MADDIE (ours) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ LIDDIA LIDDIA (Averly et al., 2025) ✓ ✓ ✓ – – – – – Prompt-to-Pill (Vichentijevikj et al., 2026) ✓ ✓ – ✓ – ✓ – – AgentD (Ock et al., 2026) – ✓ – – – – – – Generative Models for Drug Design Pocket2Mol (Peng et al., 2022) – – – – – – – – TargetDiff (Guan et al., 2023) – – – – – – – – conDitar (Gao et al., 2026) – – – – – – – – DiffSMol (Chen et al., 2025) – – – – – – – – ELILLM (Hu et al., 2026a) – – – – – – – – Generalist LLMs Nemotron-3-Ultra (Blakeman et al., 2026) – – – – – – – – Qwen3-30B (Yang et al., 2025) – – – – – – – – To evaluate the effectiveness of MADDIE MADDIE , we compare it with a wide range of baselines, each cast as a generate-then-filter pipeline: a molecular generation method followed by a filtering stage. The molecule generation methods span three categories: (1) Agents for Drug Discovery, (2) Generative Models for Drug Design and (3) Generalist LLMs. The molecules generated from each method are then filtered against each requirement, if the method cannot assess the corresponding property or synthesizability during generation. Table 1 presents which property or synthesizability each method can assess during its generation. Every molecule from generative models for drug design and generalist LLMs is filtered against every requirement after generation. In contrast, molecules from agents for drug discovery baselines are filtered only against the requirements that cannot be assessed by the method during generation. Only molecules that pass the filtering are considered as the output molecules of that method. Coupling a generation method with a filtering stage enables a fair comparison with MADDIE MADDIE , while also being a widely adopted paradigm in drug discovery (Gao et al., 2026; Zhavoronkov et al., 2019). We use a variety of methods for each category of the baselines. (1) For Agents for Drug Discovery, we compare against two state-of-the-art single-agent systems, LIDDIA (Averly et al., 2025) and AgentD (Ock et al., 2026), and one multi-agent system, Prompt-to-Pill (Vichentijevikj et al., 2026). These agents are designed for end-to-end drug discovery workflow, similar to ours, and serve as direct points of comparison for MADDIE MADDIE . They are also selected since their implementations are publicly accessible (Appendix 10). (2) For Generative Models for Drug Design, we compare against a wide variety of models: Pocket2Mol (Peng et al., 2022), TargetDiff (Guan et al., 2023), conDitar (Gao et al., 2026), DiffSMol (Chen et al., 2025) and ELILLM (Hu et al., 2026a). These models include both SBDD (Pocket2Mol, TargetDiff, conDitar, ELILLM) and LBDD (DiffSMol) approaches. (3) For Generalist LLMs, we compare against two open-source LLMs, Qwen3-30B (Yang et al., 2025) and Nemotron-3-Ultra (Blakeman et al., 2026). We focus our comparison of MADDIE MADDIE to open-source LLMs for reproducibility and accessibility: their weights are publicly available and are less likely to be discontinued than proprietary LLMs. Each LLM is competitive on general reasoning and scientific benchmarks (Yang et al., 2025; Blakeman et al., 2026), making them strong baselines. Qwen3-30B is also used as the reasoning engine for MADDIE MADDIE , while Nemotron-3-Ultra (550B) is among the largest and state-of-the-art open-source LLMs available. We only select two due to our experimentation budget. Each method is instructed to produce the same fixed number of output molecules per target as MADDIE MADDIE for a fair comparison. Methods that cannot be instructed in this way, including the five generative models for drug design and AgentD, are instead instructed to produce as many intermediate molecules as MADDIE MADDIE produces, per target. Specifically, we define an intermediate molecule as any molecule that HitIdentifier HitIdentifier , LeadOptimizer LeadOptimizer , and Synthesizer Synthesizer produce during a campaign, whether or not they serve as an output molecule. The number of intermediate molecules is used as a basis for how many molecules the generative models and AgentD should generate, and also for analysis (Section 4.2). This ensures that these baselines receive the same opportunities as MADDIE MADDIE to explore the chemical space (by generating the same number of molecules), thus supporting a fair comparison. Implementation Details In MADDIE MADDIE , we adopt Qwen3-30B (Yang et al., 2025) as the backbone LLM for Orchestrator Orchestrator and all four worker agents, an open-source LLM with strong performance across varying tasks while also small enough to be deployed locally. The model is deployed with vLLM (Kwon et al., 2023) and decoded with a temperature of 0.7, a top-p of 0.8, and a context window of 64K tokens. These are the recommended parameters for Qwen3-30B (Yang et al., 2025). We also provide implementation details for the baselines (Appendix 11), along with more details on MADDIE MADDIE (Appendix 12). 4 Experimental Results 4.1 Overall Performance Table 2: Main experimental results of MADDIE MADDIE against baselines. Method Ntout $N_tout$ Nout $N_out$ Nsat $N_sat$ %Pass Pass % Vina % QED % SAS % Mut % CTx % hERG % Lly % Syn Agents for Drug Discovery MADDIE MADDIE 50 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 LIDDIA LIDDIA 49 36.0 28.0 80.0 83.1 96.1 100.0 100.0 100.0 100.0 100.0 100.0 Prompt-to-Pill 31 14.2 2.9 20.9 55.0 88.5 100.0 41.0 100.0 98.1 100.0 100.0 AgentD 35 44.2 44.2 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 Generative Models for Drug Design Pocket2Mol 50 33.1 33.1 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 TargetDiff 48 15.8 15.8 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 conDitar 49 56.7 56.7 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 DiffSMol 27 24.7 24.7 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 ELILLM 27 4.5 4.5 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 Generalist LLMs Nemotron-3-Ultra 38 23.2 23.2 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 Qwen3-30B 3 8.0 8.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 Table 2 reports the performance of MADDIE MADDIE and the baselines. For each target, every method is instructed to produce 100 output molecules, a number large enough for further exploration (e.g., downstream selection for wet-lab validation) and small enough to keep evaluation tractable. In this setting, we compute the metrics only on targets for which a method produces at least one output molecule (e.g., generated molecules after filtering for baselines), ensuring fairness across methods. We also compare the property values of molecules generated by MADDIE MADDIE against baselines, the reference ligands, and our curated FDA-approved drugs (Appendix 13.1). Only MADDIE MADDIE delivers the requested 100 satisfying molecules on all 50 targets. The closest baselines fall well short: conDitar delivers 56.7 satisfying molecules over the 49 targets it covers, Pocket2Mol 33.1 over 50, and AgentD 44.2 over 35. This demonstrates MADDIE MADDIE ’s advantage in consistently finding satisfying molecules across targets. Such consistency arises from MADDIE MADDIE ’s multi-agent design: MADDIE MADDIE can adaptively generate, optimize, and evaluate according to what the current molecules lack, using its Orchestrator Orchestrator to decide which worker agent to delegate next. This forms a dynamic decision-making loop that improves molecules over time, thereby succeeding on a variety of targets. Agentic baselines employ evaluation tools during generation but still leave requirements unsatisfied. As illustrated in Table 1, each agent is capable of measuring some properties for the requirements: LIDDIA measures Vina , QED and SAS during its campaign, Prompt-to-Pill measures Vina , QED , Mut and hERG , and AgentD measures QED . For LIDDIA and Prompt-to-Pill, however, the per-requirement pass rates still fall short on exactly these properties, 83.1% on Vina for LIDDIA LIDDIA and 41.0% on Mut for Prompt-to-Pill. These failures likely arise from the reasoning engine powering these agents: the LLMs themselves. LLMs may misinterpret the tool outputs or hallucinated, leading it to accept a molecule that does not satisfy the requirements—a well-known weakness of LLM-based agent systems (Lin et al., 2025). In contrast, MADDIE MADDIE addresses this weakness by accepting a molecule only after it has been evaluated by Evaluator Evaluator and Synthesizer Synthesizer (Section 7). Generative models for drug design achieve perfect passing rates by filtering, but retain far fewer molecules than requested. For these baselines, filtering is applied against every requirement after generation, and molecules failing any requirement are discarded, leading to 100% pass rates for every target (with at least one output molecule). Filtering, however, cannot improve on what the methods produce: conDitar and Pocket2Mol retain only 56.7 and 33.1 satisfying molecules on average, respectively, both well below the requested 100. When the method rarely produces molecules meeting the requirements, filtering may retain none of them, as with DiffSMol for 23 of the targets. This exposes a fundamental limitation for methods without feedback from tools during generation: the pass rate can always be enforced to 100% through filtering, but the effective yield, or the number of satisfying molecules cannot be improved. Generalist LLMs produce very few satisfying molecules. For instance, Nemotron-3-Ultra, one of the state-of-the-art open-source LLMs (Blakeman et al., 2026), produces on average only 23.2 satisfying molecules on 38 targets. Qwen3-30B, the LLM that drives every agent in MADDIE MADDIE , covers only 3 targets on its own, emphasizing that the gap to the 50 reached by MADDIE MADDIE is attributable to its design rather than solely to the capability of the underlying language model. 4.2 Ablation Study We conduct ablation experiments to assess the effectiveness of the different agents in the MADDIE MADDIE framework. Specifically, we remove Synthesizer Synthesizer (Section 7.2.3), LeadOptimizer LeadOptimizer (Section 7.2.2) and Orchestrator Orchestrator (Section 7.1) individually, resulting in three variants of MADDIE MADDIE : (1) w/o Synthesizer Synthesizer , without the capability to assess synthesizability; (2) w/o LeadOptimizer LeadOptimizer , which cannot optimize any property; and (3) w/o Orchestrator Orchestrator , which only allows a fixed workflow during a campaign, respectively. In MADDIE MADDIE w/o Orchestrator Orchestrator , we set the workflow such that each agent is invoked sequentially in this order: HitIdentifier HitIdentifier , LeadOptimizer LeadOptimizer , Evaluator Evaluator , then Synthesizer Synthesizer . The workflow for MADDIE MADDIE w/o Orchestrator Orchestrator is repeated until the requested number of satisfying molecules is obtained. We do not remove HitIdentifier HitIdentifier as it serves as the starting point of a campaign and Evaluator Evaluator as every candidate molecule in MADDIE MADDIE is required to pass that worker agent. We measure Nitm $N_itm$ , the average number of intermediate molecules generated per target, and we provide the results in Table 3. Synthesizer Synthesizer is responsible for more than half of MADDIE MADDIE ’s satisfying molecules. MADDIE MADDIE w/o Synthesizer Synthesizer generates only 40.2 instead of 100 satisfying molecules, mostly because they fail the synthesis requirement (% Syn ). Synthesizer Synthesizer also generates synthesizable analogs for candidate molecules without a route, and removing it causes HitIdentifier HitIdentifier and LeadOptimizer LeadOptimizer to make up the shortfall by generating and optimizing more candidates. This is shown by the number of intermediate molecules generated: 1,554.8 molecules as opposed to 1,043.8 with Synthesizer Synthesizer on average. LeadOptimizer LeadOptimizer reduces the number of intermediate molecules MADDIE MADDIE generates by nearly 30%. Without it, the number increases from 1,043.8 to 1,327.7 as presented in Table 3. In MADDIE MADDIE w/o LeadOptimizer LeadOptimizer , a candidate molecule that fails a requirement cannot be improved further, forcing this variant to rely on HitIdentifier HitIdentifier for generation. In addition, for challenging targets that HitIdentifier HitIdentifier struggles with (e.g., ABCD1_ALD), MADDIE MADDIE w/o LeadOptimizer LeadOptimizer cannot deliver the requested molecules, resulting in only covering 49 instead of 50 targets. Table 3: Ablation study. Method Ntout $N_tout$ Nout $N_out$ Nsat $N_sat$ %Pass Pass % Syn Nitm $N_itm$ MADDIE MADDIE 50 100.0 100.0 100.0 100.0 1,043.8 w/o Synthesizer Synthesizer 50 100.0 40.2 40.2 40.2 1,554.8 w/o LeadOptimizer LeadOptimizer 49 100.0 100.0 100.0 100.0 1,327.7 w/o Orchestrator Orchestrator 50 100.0 100.0 100.0 100.0 1,743.8 Orchestrator Orchestrator greatly decreases the number of molecules explored to find the satisfying molecules. A campaign without Orchestrator Orchestrator requires generating 1,743.8 molecules on average, around 70% more than with Orchestrator Orchestrator , to reach the same goal of 100 molecules satisfying all requirements. This gap reflects the role of Orchestrator Orchestrator : rather than generating molecules blindly, Orchestrator Orchestrator intelligently uses feedback from its prior decisions and accumulated molecules to decide what to generate next, concentrating its effort on molecules likely to satisfy the requirements. As a result, fewer molecules are required to explore, leading to a more efficient and effective method. Taken together, these results show that every agent contributes by reducing the amount of exploration a campaign requires. They differ, however, in their role of reducing the intermediate molecules: (1) Synthesizer Synthesizer generates synthesizable analogs, allowing other agents to focus on the other requirements; (2) LeadOptimizer LeadOptimizer improves existing candidates, providing a more efficient exploration for MADDIE MADDIE ; and (3) Orchestrator Orchestrator intelligently adapts to what existing molecules require, enabling MADDIE MADDIE to focus on the most promising candidate molecules. 4.3 Campaign Dynamics Analysis We analyze how a campaign unfolds in MADDIE MADDIE : first through the transition probabilities across the worker agents aggregated over all the targets (Section 4.3.1), then through the flow of molecules from the worker agent that produces them to the output (Section 4.3.2), and finally through the trajectories of the worker agents on representative targets (Section 4.3.3). Each transition below is a delegation made by Orchestrator Orchestrator , and the agent itself is excluded in these figures as it serves as the central agent to which all worker agents must report. Only the four specialized worker agents are presented. 4.3.1 Overall Transition Probabilities Figure 2: Transition probabilities among worker agents during a campaign. Red color represents the transition with the highest probability out of each worker agent. Figure 2 shows the empirical transition probabilities among the worker agents, calculated from the delegation decisions Orchestrator Orchestrator makes over the 50 targets. Each edge represents a transition probability. Evaluator Evaluator is the most frequently involved worker agent during a campaign. After being invoked, HitIdentifier HitIdentifier goes to Evaluator Evaluator with probability 0.97, LeadOptimizer LeadOptimizer with probability 1.00 and Synthesizer Synthesizer with probability 0.82. Every molecule that is generated, either by HitIdentifier HitIdentifier or Synthesizer Synthesizer (e.g., synthesizable analogs), or optimized is therefore assessed before it is processed any further. Any decision made in MADDIE MADDIE on which molecules to progress is attributed to the evaluated properties from Evaluator Evaluator , emphasizing its role as a quality checkpoint for MADDIE MADDIE . MADDIE MADDIE adapts its campaign based on feedback from Evaluator Evaluator . In Figure 2, transitions out of Evaluator Evaluator (except to itself) are almost equal in probability across worker agents. Following Evaluator Evaluator , the campaign returns to HitIdentifier HitIdentifier to seek new molecules with probability 0.22, moves to LeadOptimizer LeadOptimizer to optimize one or more deficient properties with probability 0.20, or moves to Synthesizer Synthesizer to search for a synthesis route with probability 0.32. Since no transition probability dominates (out of Evaluator Evaluator ), a typical campaign in MADDIE MADDIE does not follow a fixed trajectory after evaluation. Instead, MADDIE MADDIE adapts, invoking the next worker agent according to Evaluator Evaluator ’s feedback throughout a campaign. Every campaign is guardrailed by either Evaluator Evaluator or Synthesizer Synthesizer before it finishes. A campaign never finishes until either of them is called at the end. This behavior follows similarly to real-world drug discovery, since the last step of a campaign is typically an assessment of properties and synthesizability, ensuring that the molecules are grounded in evidence. 4.3.2 Molecule Flow Across Agents (a) Molecules from HitIdentifier HitIdentifier (b) Molecules from LeadOptimizer LeadOptimizer (c) Molecules from Synthesizer Synthesizer Figure 3: Flow of molecules from the worker agent at which they originate to the campaign output. Each subfigure traces the molecules originating from one worker agent; each edge gives the number of those molecules subsequently passed to another worker agent or reaching the output, averaged over the 50 targets. Figure 3 traces the molecules produced by HitIdentifier HitIdentifier , LeadOptimizer LeadOptimizer , and Synthesizer Synthesizer during a campaign. The figure illustrates how each worker agent contributes to the satisfying molecules, and how these worker agents collaborate to achieve those molecules. HitIdentifier HitIdentifier generates the most molecules but contributes the fewest satisfying ones. HitIdentifier HitIdentifier sends 541.4 molecules to Evaluator Evaluator on average (Figure 3(a)), more than twice as many as LeadOptimizer LeadOptimizer and Synthesizer Synthesizer , which send 241.5 (Figure 3(b)) and 260.9 (Figure 3(c)), respectively. Yet only 22.5 of them end up satisfying all the requirements, compared with 33.4 and 72.8 for the other two agents, respectively. This contrast reflects a division of effort among these agents: HitIdentifier HitIdentifier explores, providing a wide range of candidates other agents can use, while LeadOptimizer LeadOptimizer and Synthesizer Synthesizer exploit, using these candidates to generate molecules more likely to meet the requirements. It is through this coupling of exploration and exploitation that MADDIE MADDIE can produce satisfying molecules in its campaign. The satisfying molecules are drawn from all three agents, with Synthesizer Synthesizer contributing the largest share. The three contributions are not independent: LeadOptimizer LeadOptimizer optimizes the molecules that HitIdentifier HitIdentifier generates, and Synthesizer Synthesizer generates its analogs from molecules that HitIdentifier HitIdentifier and LeadOptimizer LeadOptimizer produce, after being assessed by Evaluator Evaluator . Synthesizer Synthesizer produces the most satisfying molecules because it often acts last, generating analogs from candidate molecules that already satisfy most requirements. Each agent thus contributes to the final molecules, either directly or through another agent. 4.3.3 Agent Execution Across Campaigns (a) ABCD1_ALD (b) BCHE (c) ADAM10 (d) ADRB1 (e) IDH2 (f) PAK4 Figure 4: Trajectories of MADDIE MADDIE over six representative targets: (a) ABCD1_ALD, (b) BCHE, (c) ADAM10, (d) ADRB1, (e) IDH2, and (f) PAK4. Rows represent the agents in MADDIE MADDIE and columns represent one delegation or step during the campaign. The number on the arrow denotes the number of molecules passed to the next agent. The number on the box denotes the number of molecules stored in each repository after the delegation. The arrow width (worker agents) and box size (Orchestrator Orchestrator ) are proportional to the number of molecules. We further examine campaigns for different targets in MADDIE MADDIE to examine how they unfold. We also provide the number of molecules stored in P and S under Orchestrator Orchestrator (Section 7), representing the number of molecules satisfying (1) all requirements except synthesizability, and (2) all requirements. During a campaign, MADDIE MADDIE can generate more satisfying molecules than requested; in that case, it returns only the first satisfying molecules generated, up to the number requested. MADDIE MADDIE adaptively employs different strategies throughout a campaign. Figure 4(a) follows a campaign against ABCD1_ALD, a challenging target (Zuo and Chen, 2024) with no marketed FDA-approved drugs as of July 2026. Over successive rounds of delegation (by Orchestrator Orchestrator ), generation (by HitIdentifier HitIdentifier ), optimization (by LeadOptimizer LeadOptimizer ), and assessment (by Evaluator Evaluator and Synthesizer Synthesizer ), MADDIE MADDIE identifies a total of 136 satisfying molecules, demonstrating its capability. We highlight several patterns of MADDIE MADDIE in this campaign: (1) the most common strategy is generate-then-optimize, which invokes HitIdentifier HitIdentifier followed by LeadOptimizer LeadOptimizer , with Evaluator Evaluator assessing the molecules in between generations. MADDIE MADDIE applies this strategy twice in the first half of the campaign and once near the end; (2) Orchestrator Orchestrator adapts its strategy throughout the campaign. Most notably, Orchestrator Orchestrator employs generate-then-optimize during the first half, observes the results (from the molecules stored in P and S), then adopts a new strategy, multi-generate-then-optimize, which calls HitIdentifier HitIdentifier multiple times before optimizing via LeadOptimizer LeadOptimizer at the end (with Evaluator Evaluator in between). This strategy dominates the second half of the campaign, accumulating 86 property satisfying molecules in P; (3) Evaluator Evaluator is called after every generation—whether by HitIdentifier HitIdentifier , LeadOptimizer LeadOptimizer or Synthesizer Synthesizer —emphasizing its importance; (4) Synthesizer Synthesizer is the least frequently called agent (only twice). However, it is only after Synthesizer Synthesizer does S increase: Synthesizer Synthesizer ’s synthesis analog generation drives the sharp increase in satisfying molecules, from 44 to 136, thus ending the campaign. This highlights Synthesizer Synthesizer as a critical agent despite its infrequent use. Together, the campaign illustrates how worker agents, coordinated through Orchestrator Orchestrator ’s feedback-driven decisions, can solve a difficult task—generating 100 satisfying molecules for a target with no existing marketed FDA-approved drugs. MADDIE MADDIE dynamically adapts its campaign to different targets. We present five more campaigns, each of which MADDIE MADDIE approaches with distinct strategies. (1) The campaign on BCHE (Darvesh, 2016; Pajk et al., 2020), presented in Figure 4(b), is notable in that LeadOptimizer LeadOptimizer is never called: MADDIE MADDIE relies solely on HitIdentifier HitIdentifier to generate molecules, with Synthesizer Synthesizer invoked only at the end to produce synthesizable analogs. (2) For target ADAM10 (Smith et al., 2020), shown in Figure 4(c), MADDIE MADDIE pursues two kinds of strategies: HitIdentifier HitIdentifier followed by LeadOptimizer LeadOptimizer and then Synthesizer Synthesizer , and HitIdentifier HitIdentifier followed directly by Synthesizer Synthesizer . Each strategy is interleaved with Evaluator Evaluator between successive generations. The two strategies differ in whether LeadOptimizer LeadOptimizer is invoked, suggesting that Orchestrator Orchestrator observes whether the molecules require optimization and adapts accordingly. (3) ADRB1, IDH2, and PAK4, each illustrated in Figure 4(d), 4(e), and 4(f) respectively, follow shorter trajectories with a similar strategy, all beginning with HitIdentifier HitIdentifier and then LeadOptimizer LeadOptimizer , but diverging after: ADRB1 proceeds directly to Synthesizer Synthesizer and finishes, while IDH2 performs an additional round of optimization, and PAK4 an additional round of generation, before each concludes with a final Synthesizer Synthesizer call. Together, these campaigns show that MADDIE MADDIE does not execute a fixed workflow, but instead, it adapts, mirroring the adaptive workflows a medicinal chemist typically employs in real-world drug discovery. We discuss additional representative trajectories on two targets, ABO_BGAT and HTR2A (Appendix 13.2). 4.4 Case Study on PSEN1 (a) PSEN1-01 (b) PSEN1 reference ligand (c) Nirogacestat (d) PSEN1-01 and PSEN1 pocket binding interaction (surface representation) (e) PSEN1-01 and PSEN1 pocket binding interaction (cartoon representation). Hydrogen bonds and hydrophobic interactions are denoted as solid blue and dashed gray lines, respectively. Figure 5: PSEN1 case study. To assess MADDIE MADDIE ’s generated molecules beyond computational predictions, we ask a domain expert to evaluate them. We focus on three targets associated with Alzheimer’s disease: PSEN1 (Bagaria et al., 2022), ACHE (Hung et al., 2025), and BCHE (Jasiecki and Wasąg, 2019). We screen the molecules according to two criteria before expert evaluation. The criteria are (1) whether their in silico properties exceed those of reference ligands and at least one of the associated FDA-approved drugs for the target, and (2) whether their structures are chemically unstable through manual validation. We discuss the results for PSEN1 here, and the remaining targets are illustrated in Appendix 14. On PSEN1, MADDIE MADDIE proposes a molecule that surpasses both the co-crystallized ligand and the approved drug across all six in silico properties. PSEN1 encodes the catalytic subunit of Gamma-secretase, an enzyme complex linked to Alzheimer’s disease and certain tumors because it processes multiple critical membrane proteins, including the amyloid precursor protein (APP) and the cancer-related signaling protein (De Strooper et al., 2012). Despite decades of medicinal chemistry (Sherrington et al., 1995), there is only one marketed FDA-approved drug (as of July 2026) associated with PSEN1, Nirogacestat—originally developed for Alzheimer’s disease, though currently it is used for the treatment of patients with desmoid tumors (Gounder et al., 2023). Against this target, MADDIE MADDIE proposes 16 molecules with in silico properties better than either the co-crystallized reference ligand or Nirogacestat on binding affinity ( Vina ), drug-likeness ( QED ), synthesizability ( SAS ), and toxicity ( Mut , CTx , hERG ). We illustrate all 16 molecules in Figure A2. Notably, PSEN1-01 (Figure 5(a)) surpass both the reference ligand (Figure 5(b)) and Nirogacestat (Figure 5(c)) across all properties in silico, demonstrating its potential for this target. PSEN1-01 interacts favorably withing the binding site of PSEN1. Figure 5(d) and 5(e) illustrates the binding site and how PSEN1-01 interacts with it. PSEN1-01 buries the majority of its molecular surface into the site, with the 4-hydroxyphenyl ring left partially exposed. Notably, the molecule engages with the two sequence motifs of PSEN1: the hydroxyl on its pyrrolidine ring accepts hydrogen bonds with A434 and L435 of the PAL motif (Sato et al., 2008), while the acetamide carbonyl accepts from G382 of the GxGD motif (Yamasaki et al., 2006). These hydrogen bonds are complemented by nine hydrophobic contacts that is distributed over two distinct subsites: the pendant phenyl against L85, V379, T421 and L425, and the benzoyl phenyl against L268, L271, V272 and I287 (2 contacts). Together, these interactions account for the binding improvement of PSEN1-01 over both the reference ligand and Nirogacestat, and show that it can satisfy the shape complementarity this binding site demands while still satisfying other key properties. 4.5 Computational Efficiency and Cost We assess the efficiency of MADDIE MADDIE along two dimensions: (1) the number of intermediate molecules generated to achieve the requested number of satisfying molecules (Section 4.5.1), and (2) the computational cost of a campaign, measured by running time and token consumption (Section 4.5.2). 4.5.1 Intermediate Molecules Explored by MADDIE MADDIE Table 4: Number of targets grouped by the number of intermediate molecules explored. Nitm $N_itm$ Targets Share (%) <750<750 10 20.0 750750–1,0001,000 18 36.0 1,0001,000–1,2501,250 9 18.0 1,2501,250–1,5001,500 8 16.0 ≥1,500≥ 1,500 5 10.0 Table 4 reports the number of intermediate molecules produced over the 50 targets. The targets are grouped into several categories based on the intermediate molecules, and we observe how many targets exist in these categories. A majority (>>50%) of the targets require less than 1,000 intermediate molecules to achieve the requested satisfying molecules. Averaged over all targets, MADDIE MADDIE generates only 1,043.8 molecules to deliver 100 satisfying molecules, indicating that it explores the chemical space intelligently rather than through exhaustive enumeration. For the hardest 10% of targets, MADDIE MADDIE generates more than 1,500 molecules (up to 2,601 molecules), yet still delivers the requested satisfying molecules, emphasizing its robustness on even challenging targets. 4.5.2 Computational Cost We evaluate the computational cost of MADDIE MADDIE by measuring (1) the running time and output token consumption, (2) the number of actions (e.g., process and delegate; Section 7.1) taken by Orchestrator Orchestrator , and (3) the number of delegations per worker agent. We also include the computational cost for the shortest and longest campaigns, and present the results in Table 5. MADDIE MADDIE serves as a fast and low-cost autonomous system for drug discovery. Across the 50 targets, MADDIE MADDIE finishes a campaign in 3.3 hours and consumes 21,125 output tokens on average. MADDIE MADDIE takes at least about one hour on ADAMTS5 (Jiang et al., 2021) and almost half a day on HDAC8 (Kim et al., 2022), each represents the shortest and longest campaign of MADDIE MADDIE , respectively. Relative to the satisfying molecules delivered, the cost of generating one satisfying molecules is roughly 2 minutes and 211 output tokens, a cost negligible against conventional drug discovery (Mulcahy et al., 2025). Half of the delegations are made to Evaluator Evaluator . Out of 10.0 delegations issued per target on average, 5.10 go to Evaluator Evaluator , while HitIdentifier HitIdentifier , Synthesizer Synthesizer and LeadOptimizer LeadOptimizer receive 2.24, 1.66 and 1.02, respectively. Evaluator Evaluator is thus delegated the same amount of time as the remaining three agents combined, consistent with the ablation study presented previously (Section 4.2). The number of delegations, however, does not determine the campaign duration: ADAMTS5 and HDAC8 both takes 1.2 and 11.6 hours, respectively, yet they differ by only two delegations (4 and 6). This indicates that campaign cost is determined by the work performed within each delegation rather than the number of delegations issued. Table 5: Computational cost of MADDIE MADDIE over 50 targets. Metric Cost Action Delegation per agent Time (h) Out. token Total Process Delegate HI LO EV SY Average 3.3 21,125 24.9 14.9 10.0 2.2 1.0 5.1 1.7 Shortest (ADAMTS5) 1.2 7,508 9 5 4 1 0 2 1 Longest (HDAC8) 11.6 30,448 44 38 6 1 1 3 1 5 Discussions We discuss the current limitations of MADDIE MADDIE and highlight several potential directions for future work. In vitro validation for MADDIE MADDIE ’s generated molecules MADDIE MADDIE is an autonomous multi-agent system for in silico drug discovery— with dedicated agents equipped with state-of-the-art in silico tools to generate, optimize, evaluate, and plan the synthesis of molecules. MADDIE MADDIE has demonstrated its ability to generate molecules that satisfy requirements grounded in these tools. However, such computational tools remain imperfect proxies for laboratory synthesis and in vitro experiments, which are indispensable next steps to validate the real-world potential of molecules generated by MADDIE MADDIE . Despite their importance, both synthesis and in vitro experimentation remain challenging to perform systematically and at scale Chen et al. (2025). This motivates future research into how to better scale and incorporate in vitro validation into MADDIE MADDIE . In anticipation of such validation, MADDIE MADDIE provides synthesis routes, predicted binding affinity, and ADMET profiling for molecules it generates, providing extensive information needed to prioritize which candidates to advance. Furthermore, the results from in vitro experimentation can also be provided back to MADDIE MADDIE , enabling the system to adapt its workflow and revise its strategy to produce molecules more likely to advance in a real-world drug discovery campaign. Human-in-the-loop design for MADDIE MADDIE Domain experts are essential for successful drug discovery—a domain where reliable safety assessments is the highest priority and in-depth understanding is required. In this study, we couple MADDIE MADDIE ’s evaluation with assessments from a domain expert, serving primarily to verify the quality of the generated molecules, rather than to drive the generation process itself. A natural extension for MADDIE MADDIE is integrating domain experts’ feedback directly into the system, enabling an interactive framework in which expert judgments can steer the generation process toward more desirable regions of chemical space. The current design for MADDIE MADDIE is well-positioned to integrate this feedback: because Orchestrator Orchestrator already maintains an evolving repository of candidates and decides which to pursue, expert judgments can be supplied as additional information for Orchestrator Orchestrator to shape its decision. For instance, domain experts can flag undesirable substructures and scaffolds on grounds that current in silico tools do not capture, or indicate preferences among a set of candidates, thus steering the campaign towards more desirable molecules. More importantly, such feedback need not be time-consuming: a domain expert can provide occasional feedback while still meaningfully shaping the drug discovery campaign. Through a human-in-the-loop design, MADDIE MADDIE can improve its reliability of producing molecules likely to advance in a real-world drug discovery campaign (e.g., in vitro validation), while serving as a collaborator for domain experts (rather than a system that can only be inspected). MADDIE MADDIE as a fully autonomous and self-driving system with laboratory integration A more ambitious direction is to integrate automated laboratory platforms directly into MADDIE MADDIE , enabling a fully autonomous, self-driving drug discovery system. Such a system allows MADDIE MADDIE to synthesize molecules on automated synthesis platforms and submit the synthesized molecules for in vitro validation, closing the loop between in silico design and in vitro validation. Synthesizer Synthesizer and Evaluator Evaluator are the natural integration points, by equipping both agents with automated synthesis and in vitro validation platforms, respectively. Orchestrator Orchestrator can also serve as another point of integration. Since synthesis is costly and time-consuming, this makes candidate selection a resource allocation problem (e.g., selecting only the most promising molecules to synthesize under a fixed budget), a decision that requires a component with a global-level view of a campaign, which Orchestrator Orchestrator provides. Only molecules Orchestrator Orchestrator deems satisfactory should proceed to synthesis, with synthetic feasibility from Synthesizer Synthesizer and predicted properties from Evaluator Evaluator grounding that decision. Several safeguards must be implemented before a fully autonomous self-driving system can be deployed responsibly. A system that acts in the physical world introduces safety concerns absent from a system that works purely in silico. For instance, a molecule deemed unsatisfactory in silico can simply be discarded, whereas a synthesized molecule must be handled, stored, and disposed of responsibly. Safeguards are also needed to ensure molecules generated by MADDIE MADDIE are not hazardous, whether due to error or misuse of the system. Such safeguards may include structural screening against molecules known to be toxic and hazardous and including medicinal chemists in the critical step of laboratory testing. Even so, the potential benefits are substantial. Grounding MADDIE MADDIE ’s decisions in laboratory feedback can let the system steer its workflow toward real-world outcomes, and the resulting laboratory measurements can accumulate into data well suited to improving MADDIE MADDIE . More broadly, such a system can shift the role of agents in drug discovery from informing medicinal chemists to driving the drug discovery process itself, with the potential to substantially reduce the time and cost of bringing new medicine to patients. 6 Conclusions We present MADDIE MADDIE , an autonomous multi-agent system for in silico drug discovery capable of generating molecules with satisfying properties for drug development. Rather than executing a fixed workflow, MADDIE MADDIE decides which worker agent to invoke through Orchestrator Orchestrator , drawing on its memory module and evolving repositories of candidates for its decision-making. Across 50 targets, MADDIE MADDIE outperforms state-of-the-art baselines spanning generative models for drug design, agents for drug discovery, and generalist LLMs, successfully generating 100 satisfying molecules per target that meet key molecular property requirements while also providing their synthesis routes. Our ablation reveals that this performance is attributed to MADDIE MADDIE ’s collaboration between agents, where each worker agent contributes distinctly, while Orchestrator Orchestrator contributes the most by concentrating effort on the most promising candidates that greatly reduce the chemical space MADDIE MADDIE must explore to find the satisfying molecules. Consistent with its design, MADDIE MADDIE exhibits strong adaptability across campaigns, effectively employing different strategies as each target demands. Notably, MADDIE MADDIE demonstrates intelligent strategy adaptations for a challenging target with no existing marketed FDA-approved drugs. On PSEN1, a critical target associated with Alzheimer’s disease, MADDIE MADDIE generates PSEN1-01, a promising molecule with in silico properties outperforming those of the reference ligand and Nirogacestat, while also demonstrating favorable binding interaction with the challenging site of the target. MADDIE MADDIE achieves strong performance at low cost and time, with one campaign roughly producing 21K output tokens and taking 3.3 hours on average. We envision MADDIE MADDIE as a foundation for future autonomous drug discovery systems that can carry a molecule in silico to potential drug candidates, ultimately accelerating the timeline and cost of bringing new drugs to patients. 7 Methods We introduce MADDIE MADDIE , a multi-agent system for in silico drug discovery. It is composed of a central agent, Orchestrator Orchestrator , and four specialized worker agents. Given a target protein and a set of desired molecular properties, MADDIE MADDIE aims to identify molecules that bind to the target and satisfy all desired properties. MADDIE MADDIE performs the task by distributing subtasks across worker agents, with Orchestrator Orchestrator deciding which subtask to pursue next. This design enables MADDIE MADDIE to decompose a complex, multi-objective problem typical in drug discovery into manageable subtasks, enabling each agent to collectively advance the drug discovery campaign. Orchestrator Orchestrator (Section 7.1) is responsible for the entire workflow of MADDIE MADDIE , deciding how a drug discovery campaign proceeds by delegating to a worker agent and processing the molecules. Under its coordination, the worker agents perform distinct subtasks critical to a successful drug discovery campaign (Hughes et al., 2011): (1) HitIdentifier HitIdentifier (Section 7.2.1) serves as the entry point of a campaign, providing candidate molecules from which a campaign proceeds; (2) LeadOptimizer LeadOptimizer (Section 7.2.2) serves as the refinement stage, advancing existing candidate molecules towards those with satisfying properties; (3) Synthesizer Synthesizer (Section 7.2.3) bridges the gap between molecules designed in silico and those that can be made in practice, proposing synthesis routes and finding synthesizable analogs; and (4) Evaluator Evaluator (Section 7.2.4) functions as a quality control layer, assessing the key properties that define a successful drug, including binding affinity, drug-likeness, and safety-related properties. Acting in concert, these components allow MADDIE MADDIE to emulate the iterative, collaborative nature of a real-world campaign within a single autonomous system (Hughes et al., 2011). Each agent (Orchestrator Orchestrator and the four worker agents) maintains its own workspace, each is used to store relevant artifacts required for their tasks (Table 6). The workspaces are dynamically changing and growing continuously, with more artifacts (e.g., molecules) generated as the campaign progresses. Each worker agent operates within its own workspace, with no access to the others. This prevents any alteration of artifacts from worker agents outside their own workspace, preventing error propagation throughout MADDIE MADDIE . Meanwhile, Orchestrator Orchestrator alone has access to each worker agent’s workspace, granting it a global view required to lead the campaign. 7.1 Orchestrator Orchestrator At the center of MADDIE MADDIE is Orchestrator Orchestrator , an agent responsible for the decision-making of the entire drug discovery campaign. Orchestrator Orchestrator decides how a campaign proceeds: which worker agent to call next and which candidate molecules to focus on. Each decision is conditioned on two sources of information: (1) the existing candidate molecules generated throughout the campaign, which captures the progress of the campaign, and (2) the profiles of the worker agents (Appendix 12.2), which describe the capabilities that each worker agent can contribute to the campaign. Together, these information allow Orchestrator Orchestrator to reason jointly about what the campaign needs and which worker agent is best suited to provide it. Furthermore, this information enables dynamic decision-making for Orchestrator Orchestrator , choosing which drug discovery step to focus on next and calling the worker agent dedicated to that step. This design makes MADDIE MADDIE ’s campaign non-linear and adaptive, mirroring the iterative, feedback-driven nature of real-world drug discovery (Hughes et al., 2011). Operationally, Orchestrator Orchestrator makes two kinds of decisions. It can choose to advance the campaign by issuing a delegate delegate action to a worker agent, or to collect information (e.g., the properties of existing candidate molecules) by issuing a process process action. (1) Using a delegate action, Orchestrator Orchestrator sends a worker agent a prompt consisting of a task description and references to the artifacts stored in each worker agent’s workspace (Section 7.2). In return, it receives a structured report of what the worker agent has done. These reports, along with the information gathered from process process action(s) and the profiles of worker agents will determine which worker agent should be called next via delegate delegate . (1) Using a process action, Orchestrator Orchestrator reads the artifacts produced from the worker agent it has just called. What it takes from them depends on the worker agent: the molecules generated by HitIdentifier HitIdentifier or modified by LeadOptimizer LeadOptimizer , the property values computed by Evaluator Evaluator , or the synthesis routes found by Synthesizer Synthesizer . From these artifacts it updates its view of which candidate molecules meet the requirements and which of them still fall short. With these two actions, Orchestrator Orchestrator drives the campaign forward (through delegate) while continually grounding its decision (through process). Carrying these actions out over many steps rests on two elements of its design: (1) the memory that keeps its own record tractable and (2) the satisfying candidate repositories in which it keeps the molecules worth pursuing, stored in the workspace. Memory module A defining feature of Orchestrator Orchestrator is its capability to sustain a long-horizon campaign for MADDIE MADDIE (Section 4.3.3). This is enabled by a dedicated memory module that dynamically stores the most relevant information. The module holds the records of the campaign itself—the information of the artifacts produced, the reports the worker agents return and Orchestrator Orchestrator ’s own prior decisions—kept in full. Other information is stored and pruned to the most recent entries, such as system-level messages generated throughout the campaign (e.g., Orchestrator Orchestrator calling a worker agent with incorrect format). When the memory module exceeds what the underlying LLM can hold in its context, the module compresses its records by summarizing them, selectively storing and pruning information based on its recency and relevance. The module thus provides Orchestrator Orchestrator a compact, decision-relevant context, allowing it to condition each decision on the salient history of the campaign while remaining within the effective reasoning capacity of the underlying LLM. Satisfying candidate repositories Over the course of a campaign, Orchestrator Orchestrator maintains two repositories of molecules. The first repository P holds the candidate molecules that satisfy every property evaluated by Evaluator Evaluator . Concretely, these are molecules from HitIdentifier HitIdentifier and LeadOptimizer LeadOptimizer that pass each property requirement. The second repository S holds candidate molecules that both satisfy every property and are also synthesizable as verified by Synthesizer Synthesizer . This set has two sources: (1) molecules from P that are synthesizable, and (2) their synthesizable analogs (Section 7.2.3) found by Synthesizer Synthesizer that satisfy every property as evaluated by Evaluator Evaluator . Orchestrator Orchestrator takes both repositories into account when choosing each subsequent action. These repositories serve several purposes: P tracks which candidate molecules are worth considering for synthesis assessments, while S tracks which of them are both satisfying and also synthesizable. They also inform Orchestrator Orchestrator on which worker agent it should call next. For instance, a large P means there are many candidate molecules with satisfying properties, which informs Orchestrator Orchestrator that it is better to call Synthesizer Synthesizer to verify their synthesizability next. Orchestrator Orchestrator is allowed a budget of actions per campaign, and stops once S holds the requested number of molecules or this budget is exhausted. At each step Orchestrator Orchestrator is informed of how many actions remain, and once a few of them are left it stops searching for new molecules and delegates those in P to Synthesizer Synthesizer , ensuring that the candidates accumulated so far can still enter S before the budget runs out. The combination of these capabilities in Orchestrator Orchestrator is what distinguishes MADDIE MADDIE from prior approaches. Together they allow MADDIE MADDIE to operate autonomously, to accelerate drug discovery process (Section 4.5.2), and to achieve strong performance in our benchmark (Table 2). Critically, they allow MADDIE MADDIE to complete long-horizon campaigns on varying targets (Section 4.3.3). 7.2 Worker Agents Each worker agent carries out one specialized subtask in a drug discovery campaign, each critical for a successful campaign. Each worker agent maintains a static profile, containing a brief summary of its capabilities and what subtask they support (Appendix 12.2). These profiles are used by Orchestrator Orchestrator during its decision-making. Each worker agent receives from Orchestrator Orchestrator a prompt containing a task description with references to the artifacts in its workspace, executes the subtask with its own tools and stores the artifacts it produces in its workspace. The worker agent then issues a report by providing a structured summary of what it has done back to Orchestrator Orchestrator . Table 6 lists the artifacts each worker agent requires and produces in its workspace and the tools used. The subsections below describe the strategy with which each worker agent approaches its subtask. Table 6: Artifacts required and produced by each worker agent, along with their respective tools. Worker agent Input Output Tools HitIdentifier HitIdentifier Target pocket P. Candidate molecules. conDitar (Gao et al., 2026) for SBDD; DiffSMol (Chen et al., 2025) for LBDD; AutoDock (Trott and Olson, 2010a) for predicting binding affinity. Evaluator Evaluator Candidate molecules. Property values. UniDock (Yu et al., 2023), QED (Bickerton et al., 2012), SAScore (Ertl and Schuffenhauer, 2009), ADMET-AI (Swanson et al., 2024) and Lilly filter (Bruns and Watson, 2012). LeadOptimizer LeadOptimizer Candidate molecules that fail one or more requirements. Optimized candidate molecules. GeLLMO-C (Dey et al., 2025); MolOpt algorithms (Gao et al., 2022); RDKit and Evaluator Evaluator -backed oracles for the target properties. Synthesizer Synthesizer Candidate molecules. Synthesis routes or synthesizable analogs. MEEA∗ (Zhao et al., 2024) and PrexSyn (Luo and Coley, 2026). 7.2.1 HitIdentifier HitIdentifier HitIdentifier HitIdentifier serves as the entry point of MADDIE MADDIE ’s campaign, supplying the candidate molecules from which the rest of the campaign proceeds. It is the only worker agent that conditions the molecule generation directly on the target structure, enabling MADDIE MADDIE to propose molecules likely to bind the target. Similar to hit identification step in drug discovery, HitIdentifier HitIdentifier prioritizes binding affinity, a key property for molecules to have the intended therapeutic effects (Sawada et al., 2024). Every molecule MADDIE MADDIE proposes is derived (e.g., optimized) from what HitIdentifier HitIdentifier finds, emphasizing its importance. HitIdentifier HitIdentifier first receives a subtask from Orchestrator Orchestrator containing the references to artifacts in its workspace (e.g., target structure), a number of molecules to return and a budget of iterations to find them. HitIdentifier HitIdentifier then finds these molecules over a series of iterations using two strategies: structure-based drug design (SBDD) (Batool et al., 2019) and ligand-based drug design (LBDD) (Shim and MacKerell, 2011). At each iteration, it selects a strategy, invokes the corresponding tool, and accumulates the resulting molecules alongside those from previous iterations. The molecules are then scored using their predicted binding affinities, stored in its workspace, and accumulated to guide the selection of the next strategy. HitIdentifier HitIdentifier stops once the molecules satisfying the constraints reaches or exceeds the requested number from Orchestrator Orchestrator or it exhausts the budget of iterations. Structure-based drug design SBDD generates molecules conditioned on the three-dimensional structure of the target pocket. It enables HitIdentifier HitIdentifier to design molecules by exploiting the binding pocket’s spatial configurations, offering a mechanistic approach to molecular discovery grounded in pocket-ligand interactions. SBDD further provides a principled starting point for ligand-based modeling, or LBDD, and thus complements each other. Ligand-based drug design LBDD instead generates molecules conditioned on the shape of one that already binds to the pocket. Where SBDD generates molecules from the binding pocket, LBDD expands around the known binding molecules. Since it relies on different prior information (e.g., shape-conditioned instead of pocket-conditioned), LBDD broadens HitIdentifier HitIdentifier ’s exploration when combined with SBDD and thus, further increases the likelihood to find more and stronger binding molecules. 7.2.2 LeadOptimizer LeadOptimizer LeadOptimizer LeadOptimizer is the molecular-refinement component of MADDIE MADDIE . It complements the molecular generative stages as done by HitIdentifier HitIdentifier (e.g., generates candidates from a binding pocket) and Synthesizer Synthesizer (e.g., generates synthesizable analogs from existing candidates). Specifically, their candidate molecules are used as starting points by LeadOptimizer LeadOptimizer to produce derivatives with improved properties. LeadOptimizer LeadOptimizer addresses the different property requirements a campaign requires through two different families of optimizer: (1) pretrained controllable multi-property optimizer (Dey et al., 2025) and (2) oracle-guided single-property optimizer (Gao et al., 2022), each is distinct and complements each other. They provide both adaptability and control over the property improvements to whatever a campaign demands, enabling LeadOptimizer LeadOptimizer to transform molecules in MADDIE MADDIE to those with satisfying properties. LeadOptimizer LeadOptimizer performs a subtask by first formulating a plan and then selecting the most suitable optimizer. Concretely, after Orchestrator Orchestrator identifies which candidate molecules are needed to improve, what properties require improvement, and the number of derivatives to produce for each molecule, LeadOptimizer LeadOptimizer turns the request into a structured plan with objectives. It contains both properties to be improved and the required values or ranges. The plan then guides LeadOptimizer LeadOptimizer to select the best optimizer from one of two families, based on which properties to optimize, how many of them, and the performance of that optimizer, when such records are stored by LeadOptimizer LeadOptimizer . Once an optimizer is selected, LeadOptimizer LeadOptimizer applies it to each candidate molecule, producing molecule derivatives that are then validated, de-duplicated, stored in its workspace, and reported back to Orchestrator Orchestrator . If the initial attempt doesn’t achieve the requested property improvement(s) or produce enough valid derivatives, LeadOptimizer LeadOptimizer may run follow-up attempts including retrying or switching optimizers, subject to an execution budget. Pretrained controllable multi-property optimizer The pretrained controllable multi-property optimizer family is well suited when the subtask requires improving multiple properties while attempting to preserve other desirable molecular properties. This family offers LeadOptimizer LeadOptimizer greater control over which properties to maintain and improve. In contrast to using multiple single-property optimizers (the second family), a pretrained multi-property optimizer provides more balanced improvement (e.g., improving while maintaining) and efficiency (e.g., one instead of multiple runs). However, because this family is often pretrained on specific properties (Dey et al., 2025), it limits their adaptability for the wide range of properties a drug discovery campaign may require, necessitating LeadOptimizer LeadOptimizer to use another family of optimizers. Oracle-guided single-property optimizer Oracle-guided single-property optimizers refine molecules using an in silico property predictor available to LeadOptimizer LeadOptimizer (Table 6). This family is primarily used when the properties to optimize are not supported by the first family. LeadOptimizer LeadOptimizer has access to more than one optimizer in this family. To select the best one, it uses its available property predictors, prior optimization results, and historical performance records of each tool as guidance (Appendix 12.3). 7.2.3 Synthesizer Synthesizer Synthesizer Synthesizer investigates the synthesis of molecules, assessing whether the candidate molecules in MADDIE MADDIE can be made in practice. This is critical for in silico drug discovery, as synthesizability is a major bottleneck to development of AI-generated molecules (Blass, 2021). Synthesizer Synthesizer assesses synthesizability based on whether it can generate a synthesis route for a given molecule, grounding its assessments in interpretable reaction chemistry rather than relying solely on estimated synthetic accessibility scores (Ertl and Schuffenhauer, 2009). Synthesizer Synthesizer uses two complementary strategies to generate synthesis routes: (1) retrosynthesis planning (Zhao et al., 2024), and (2) synthesizable analog generation (Luo and Coley, 2026). This two-strategy approach mirrors conventional practice in drug discovery, where molecules that are difficult to synthesize directly are used to identify analogs from established libraries of synthesizable molecules. Retrosynthesis planning Synthesizer Synthesizer uses retrosynthesis planning to generate synthesis routes by starting from a given candidate molecule and working through the reverse logic of synthesis reactions to reach purchasable starting materials. Using this approach, Synthesizer Synthesizer can identify synthesis routes to a specific molecule, mimicking the retrosynthetic analyses performed by chemists in conventional drug discovery. If this planning process can identify a route for a given molecule, Synthesizer Synthesizer uses the route as the evidence to conclude that the molecule is synthesizable and terminates. However, if no route is found, Synthesizer Synthesizer investigates further using the synthesizable analog generation strategy. Synthesizable analog generation In synthesizable analog generation, Synthesizer Synthesizer identifies analogs by constructing synthesis routes to molecules that are structurally similar to the given candidate molecule. Unlike retrosynthesis planning, the route generation starts from purchasable starting materials and works through the forward logic of synthesis reactions to produce analogs. This forward-synthesis approach naturally complements retrosynthesis planning, and can uncover synthesis routes that retrosynthesis planning may struggle to identify. In some cases, the product of the generated synthesis route is not an analog, but the given candidate molecule itself. When this occurs, Synthesizer Synthesizer concludes that the candidate molecule is synthesizable. Otherwise, Synthesizer Synthesizer concludes that the candidate molecule is unsynthesizable and provides the analogs and their routes to Orchestrator Orchestrator as potential alternatives to the candidate molecule. Thus, Synthesizer Synthesizer goes beyond synthesizability assessment, converting unsynthesizable molecules into potentially promising new leads to further the drug discovery campaign. 7.2.4 Evaluator Evaluator Evaluator Evaluator serves as the quality control mechanism in MADDIE MADDIE . Every decision made by Orchestrator Orchestrator regarding the candidate molecules is grounded in Evaluator Evaluator ’s assessments, thus shaping the trajectory of MADDIE MADDIE ’s campaign. Evaluator Evaluator performs a comprehensive assessment of each candidate molecule and supplies the property values on which the rest of a campaign depends. The properties Evaluator Evaluator supports include binding affinity, drug-likeness, synthesizability, ADMET and structural liabilities. Each is critical to a successful drug and computed with its own respective tool (Table 6). Its synthesizability assessment differs from Synthesizer Synthesizer since Evaluator Evaluator uses a rapid heuristic scorer (Ertl and Schuffenhauer, 2009) rather than finding an entire synthesis route from purchasable starting molecules to the candidate molecule. The score complements the synthesis route—it can convey information on the ease of synthesis (Ertl and Schuffenhauer, 2009) as a quantitative number that can help in prioritizing molecules. Acknowledgements This project was made possible, in part, by support from the National Science Foundation grant no. IIS-2435819 (X.N., D.A.-A.), the National Library of Medicine grant no. R01LM014385 (X.N.), and the National Institute of Aging grant no. R01AG082044 (X.N.). Any opinions, findings, conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the funding agency. Conflict of interests The authors claim no conflict of interests. Ethics Statement MADDIE MADDIE is an autonomous system for drug discovery, enabling users to generate molecules that meet the specified property requirements prompted in natural language. We acknowledge that not all generated molecules are safe, and MADDIE MADDIE can potentially be used to generate harmful molecules. We take a step to minimize the potential negative impact and design MADDIE MADDIE to operate purely in silico, with no capability for the system to execute synthesis and produce physical molecules. Nonetheless, we highly encourage supervision by responsible domain experts when using MADDIE MADDIE . All users of MADDIE MADDIE are advised to exercise discretion and to comply with applicable safety guidelines, regulations, and established professional best practices. Code Availability All the source code and implementation for MADDIE MADDIE is available in https://github.com/ninglab/MADDIE. Data Availability The preprocessed 3D structures of the protein targets, their annotations, and the associated FDA-approved drugs are available in https://github.com/ninglab/MADDIE. Author Contributions X.N. conceived the research and conducted the project administration. X.N. investigated the research, obtained funding and resources for the research, and supervised the student and postdoctoral authors. Y.L., R.A. and X.N. designed the computational methodologies of MADDIE MADDIE . Y.L. and R.A. designed and implemented the computational methodologies of Orchestrator Orchestrator . R.G. designed and implemented the computational methodologies of HitIdentifier HitIdentifier . A.M. designed and implemented the computational methodologies of LeadOptimizer LeadOptimizer . F.B. designed and implemented the computational methodologies of Synthesizer Synthesizer . Y.L. designed and implemented the computational methodologies of Evaluator Evaluator . Y.L. conducted the main experiments. R.A., A.M., F.B. and R.G. conducted the baseline experiments. R.A. and A.M. conducted data curation. Y.L. and R.A. conducted formal analysis and result analysis. D.A.-A. contributed medicinal chemistry expertise to the result analysis. Y.L. conducted visualization; R.A. and A.M. contributed to visualization. R.A. drafted the original manuscript. Y.L., A.M., F.B., R.G. and X.N. conducted the manuscript reviewing and editing. All the authors reviewed the final paper. References Adams and Coley (2023) K. Adams and C. W. Coley Equivariant shape-conditioned generation of 3d molecules for ligand-based drug design. In The Eleventh International Conference on Learning Representations, Cited by: §2.2. Averly et al. (2025) R. Averly, F. N. Baker, I. A. Watson, and X. Ning LIDDIA: language-based intelligent drug discovery agent. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 12004–12028. External Links: Document, ISBN 979-8-89176-332-6 Cited by: §1, §1, §10, §13.1, §2.1, §3, §3, Table 1, §8. Avram et al. (2023) S. Avram, T. B. Wilson, R. Curpan, L. Halip, A. Borota, A. Bora, C. G. Bologa, J. Holmes, J. Knockel, J. J. Yang, and T. I. Oprea DrugCentral 2023 extends human clinical data and integrates veterinary drugs. Nucleic Acids Research 51 (D1), p. D1276–D1287. External Links: ISSN 1362-4962, Link, Document Cited by: §8. Bagaria et al. (2022) J. Bagaria, E. Bagyinszky, and S. S. A. An Genetics, functions, and clinical impact of presenilin-1 (psen1) gene. International Journal of Molecular Sciences 23 (18). External Links: Link, ISSN 1422-0067, Document Cited by: §4.4. Batool et al. (2019) M. Batool, B. Ahmad, and S. Choi A structure-based drug discovery paradigm. International Journal of Molecular Sciences 20 (11). External Links: Link, ISSN 1422-0067, Document Cited by: §7.2.1. Berman et al. (2000) H.M. Berman, Z. F. J. Westbrook, G. Gilliland, T.N. Bhat, H. Weissig, I.N. Shindyalov, and P.E. Bourne The protein data bank. Nucleic Acids Research 28 (1), p. 235–242. External Links: ISSN 1362-4962, Link, Document Cited by: §3, §8. Bickerton et al. (2012) G. R. Bickerton, G. V. Paolini, J. Besnard, S. Muresan, and A. L. Hopkins Quantifying the chemical beauty of drugs. Nature chemistry 4 (2), p. 90–98. Cited by: §3, Table 6. Birks et al. (2015) J. Birks, L. Chong, and J. Grimley Evans Rivastigmine for alzheimer’s disease. Cochrane Database of Systematic Reviews (9). External Links: ISSN 1465-1858, Document, Link Cited by: §14. Blakeman et al. (2026) A. Blakeman, A. Thomas, A. Jhunjhunwala, A. Gupta, A. Khattar, A. Rajfer, A. Renduchintala, A. Asif, A. Vavre, A. F. Miranda, et al. Nemotron 3 ultra: open, efficient mixture-of-experts hybrid mamba-transformer model for agentic reasoning. arXiv preprint arXiv:2606.15007. Cited by: §3, Table 1, §4.1. Blass (2021) B. E. Blass Basic principles of drug discovery and development. Second edition edition, Academic Press, London (eng). External Links: ISBN 978-0-12-817214-8 978-0-12-817215-5 Cited by: §7.2.3. Bohacek et al. (1996) R. S. Bohacek, C. McMartin, and W. C. Guida The art and practice of structure-based drug design: a molecular modeling perspective. Medicinal research reviews 16 (1), p. 3–50. Cited by: §1. Bruns and Watson (2012) R. F. Bruns and I. A. Watson Rules for identifying potentially reactive or promiscuous compounds. Journal of Medicinal Chemistry 55 (22), p. 9763–9772. External Links: ISSN 0022-2623, Document, Link Cited by: §3, Table 6. Buniello et al. (2025) A. Buniello, D. Suveges, C. Cruz-Castillo, M. B. Llinares, H. Cornu, I. Lopez, K. Tsukanov, J. M. Roldán-Romero, C. Mehta, L. Fumis, G. McNeill, J. D. Hayhurst, R. E. Martinez Osorio, E. Barkhordari, J. Ferrer, M. Carmona, P. Uniyal, M. J. Falaguera, P. Rusina, I. Smit, J. Schwartzentruber, T. Alegbe, V. W. Ho, D. Considine, X. Ge, S. Szyszkowski, Y. Tsepilov, M. Ghoussaini, I. Dunham, D. G. Hulcoop, E. M. McDonagh, and D. Ochoa Open targets platform: facilitating therapeutic hypotheses building in drug discovery. Nucleic Acids Research 53 (D1), p. D1467–D1475. External Links: ISSN 1362-4962, Document, Link Cited by: §11.1, §3, §8. Cao et al. (2026) H. Cao, S. Liu, F. Zhang, Z. Liu, H. Li, B. Feng, S. Bai, L. Chen, K. Xie, and Y. Li Mozi: governed autonomy for drug discovery llm agents. arXiv preprint arXiv:2603.03655. Cited by: §1, §1, §2.1. Chen et al. (2025) Z. Chen, B. Peng, T. Zhai, D. Adu-Ampratwum, and X. Ning Generating 3d small binding molecules using shape-conditioned diffusion models with guidance. Nature machine intelligence 7 (5), p. 758–770. Cited by: §1, §11.2, §2.2, §3, Table 1, §5, Table 6. D. Segall (2012) M. D. Segall Multi-parameter optimization: identifying high quality compounds with a balance of properties. Current Pharmaceutical Design 18 (9), p. 1292–1310. External Links: ISSN 1381-6128, Link, Document Cited by: §1. Darvesh (2016) S. Darvesh Butyrylcholinesterase as a diagnostic and therapeutic target for alzheimer’s disease.. Current Alzheimer Research 13 (10), p. 1173–7. External Links: Document, Link Cited by: §4.3.3. De Strooper et al. (2012) B. De Strooper, T. Iwatsubo, and M. S. Wolfe Presenilins and γ-secretase: structure, function, and role in alzheimer disease. Cold Spring Harbor perspectives in medicine 2 (1), p. a006304. Cited by: §4.4. Dey et al. (2025) V. Dey, X. Hu, and X. Ning Large language models for controllable multi-property multi-objective molecule optimization. In Findings of the Association for Computational Linguistics: EMNLP 2025, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 20996–21023. External Links: Document Cited by: §1, §7.2.2, §7.2.2, Table 6. Ertl and Schuffenhauer (2009) P. Ertl and A. Schuffenhauer Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of cheminformatics 1 (1), p. 8. Cited by: §3, §7.2.3, §7.2.4, Table 6. Francoeur et al. (2020) P. G. Francoeur, T. Masuda, J. Sunseri, A. Jia, R. B. Iovanisci, I. Snyder, and D. R. Koes Three-dimensional convolutional neural networks and a cross-docked data set for structure-based drug design. Journal of Chemical Information and Modeling 60 (9), p. 4200–4215. External Links: ISSN 1549-9596, Document, Link Cited by: §3, §8. Gao et al. (2025a) B. Gao, Y. Huang, Y. Liu, W. Xie, B. He, H. Tan, W. Ma, Y. Zhang, and Y. Lan CIDD: collaborative intelligence for structure-based drug design empowered by llms. In Advances in Neural Information Processing Systems, D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen (Eds.), Vol. 38, Main Conference, p. 13651–13699. External Links: Document Cited by: §2.2. Gao et al. (2025b) B. Gao, Y. Huang, Y. Liu, W. Xie, W. Ma, Y. Zhang, and Y. Lan Pharmagents: building a virtual pharma with large language model agents. arXiv preprint arXiv:2503.22164. Cited by: §1, §1, §2.1. Gao et al. (2026) R. Gao, J. Peng, Z. Chen, F. N. Baker, D. C. Kombo, J. L. K. Jr., A. A. Scholte, Y. Li, M. J. LaMarche, L. I. Iconaru, H. Biemann, M. Hong, and X. Ning Generating developable 3d molecules via pocket-conditioned diffusion and property-aware optimization. External Links: 2607.12349, Link Cited by: §1, §11.2, §2.2, §2.2, §3, §3, §3, Table 1, Table 6. Gao et al. (2022) W. Gao, T. Fu, J. Sun, and C. Coley Sample efficiency matters: a benchmark for practical molecular optimization. In Advances in Neural Information Processing Systems, Vol. 35, p. 21342–21357. Cited by: §12.3, §7.2.2, Table 6. Gayvert et al. (2016) K. M. Gayvert, N. S. Madhukar, and O. Elemento A data-driven approach to predicting successes and failures of clinical trials. Cell chemical biology 23 (10), p. 1294–1301. Cited by: §3. Ghareeb et al. (2026) A. E. Ghareeb, B. Chang, L. Mitchener, A. Yiu, C. J. Szostkiewicz, D. Shved, G. J. Gyimesi, J. M. Laurent, S. M. Wright, M. T. Razzak, A. D. White, S. C. Finnemann, M. M. Hinks, and S. G. Rodriques A multi-agent system for automating scientific discovery. Nature 655 (8122), p. 497–505. External Links: ISSN 1476-4687, Link, Document Cited by: §2.1. Gounder et al. (2023) M. Gounder, R. Ratan, T. Alcindor, P. Schöffski, W. T. Van Der Graaf, B. A. Wilky, R. F. Riedel, A. Lim, L. M. Smith, S. Moody, et al. Nirogacestat, a γ-secretase inhibitor for desmoid tumors. New England Journal of Medicine 388 (10), p. 898–912. Cited by: §4.4. Gu et al. (2024) S. Gu, M. Xu, A. Powers, W. Nie, T. Geffner, K. Kreis, J. Leskovec, A. Vahdat, and S. Ermon Aligning target-aware molecule diffusion models with exact energy optimization. In Advances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37, p. 44040–44063. External Links: Document Cited by: §2.2. Guan et al. (2023) J. Guan, W. W. Qian, X. Peng, Y. Su, J. Peng, and J. Ma 3D equivariant diffusion for target-aware molecule generation and affinity prediction. In International Conference on Learning Representations, Cited by: §11.2, §2.2, §3, Table 1. Hu et al. (2026a) X. Hu, A. Li, Q. Xing, J. Ji, H. Tuo, and B. Yang Empowering llms for structure-based drug design via exploration-augmented latent inference. In Proceedings of the ACM Web Conference 2026, W ’26, New York, NY, USA, p. 4244–4255. External Links: ISBN 9798400723070, Document Cited by: §11.2, §2.2, §3, Table 1. Hu et al. (2026b) X. Hu, H. Tuo, J. Ji, A. Li, Q. Xing, and B. Yang Steering large language models through the DMTA cycle: structure-based drug design via knowledge-driven bi-level thompson sampling. In Proceedings of the 43rd International Conference on Machine Learning, T. Zhang, M. Dudik, M. Jaggi, A. Agarwal, S. Li, D. Schuurmans, J. Zhu, F. Berkenkamp, H. Dong, and A. Bietti (Eds.), Proceedings of Machine Learning Research, Vol. 306, p. 45703–45730. Cited by: §2.2. Hughes et al. (2011) J. P. Hughes, S. Rees, S. B. Kalindjian, and K. L. Philpott Principles of early drug discovery. British journal of pharmacology 162 (6), p. 1239–1249. Cited by: §7.1, §7. Hung et al. (2025) L. Hung, K. Y. Sanbonmatsu, R. F. Williams, and J. C.-H. Chen Acetylcholinesterase: structure, dynamics, and interactions with organophosphorus compounds. Protein Science 34 (10), p. e70297. External Links: Document, Link, https://onlinelibrary.wiley.com/doi/pdf/10.1002/pro.70297 Cited by: §4.4. Jasiecki and Wasąg (2019) J. Jasiecki and B. Wasąg Butyrylcholinesterase protein ends in the pathogenesis of alzheimer’s disease—could bche genotyping be helpful in alzheimer’s therapy?. Biomolecules 9 (10). External Links: Link, ISSN 2218-273X, Document Cited by: §4.4. Jiang et al. (2021) L. Jiang, J. Lin, S. Zhao, J. Wu, Y. Jin, L. Yu, N. Wu, Z. Wu, Y. Wang, and M. Lin ADAMTS5 in osteoarthritis: biological functions, regulatory network, and potential targeting therapies. Frontiers in Molecular Biosciences Volume 8 - 2021. External Links: Link, Document, ISSN 2296-889X Cited by: §4.5.2. Kantorow et al. (2026) J. Kantorow, N. Mani, N. R. Mohanraj, and X. Zong F.a.d.e. (fully agentic drug engine): a conversational ai platform for drug discovery. bioRxiv. External Links: Document, Link Cited by: §1, §1, §2.1. Kass-Hout et al. (2016) T. A. Kass-Hout, Z. Xu, M. Mohebbi, H. Nelsen, A. Baker, J. Levine, E. Johanson, and R. A. Bright OpenFDA: an innovative platform providing access to a wealth of fda’s publicly available data. Journal of the American Medical Informatics Association 23 (3), p. 596–600. External Links: ISSN 1067-5027, Document, Link Cited by: §8. Kim et al. (2022) J. Y. Kim, H. Cho, J. Yoo, G. W. Kim, Y. H. Jeon, S. W. Lee, and S. H. Kwon Pathological role of hdac8: cancer and beyond. Cells 11 (19). External Links: Link, ISSN 2073-4409, Document Cited by: §4.5.2. Kwon et al. (2023) W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, Cited by: §3. Lilienfeld (2002) S. Lilienfeld Galantamine — a novel cholinergic drug with a unique dual mode of action for the treatment of patients with alzheimer’s disease. CNS Drug Reviews 8 (2), p. 159–176. External Links: Document, Link, https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1527-3458.2002.tb00221.x Cited by: §14. Lin et al. (2025) X. Lin, Y. Ning, J. Zhang, Y. Dong, Y. Liu, Y. Wu, X. Qi, N. Sun, Y. Shang, K. Wang, et al. Llm-based agents suffer from hallucinations: a survey of taxonomy, methods, and directions. arXiv preprint arXiv:2509.18970. Cited by: §4.1. Liu et al. (2025) S. Liu, Y. Lu, S. Chen, X. Hu, J. Zhao, Y. Lu, and Y. Zhao DrugAgent: automating ai-aided drug discovery programming through llm multi-agent collaboration. External Links: 2411.15692, Link Cited by: §1, §2.1. Long et al. (2022) S. Long, Y. Zhou, X. Dai, and H. Zhou Zero-shot 3d drug design by sketching and generating. In Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35, p. 23894–23907. External Links: Document Cited by: §2.2. Luo and Coley (2026) S. Luo and C. W. Coley Efficient and programmable exploration of synthesizable chemical space. Transactions on Machine Learning Research. Note: External Links: ISSN 2835-8856, Link Cited by: §7.2.3, Table 6. McNaughton et al. (2024) A. D. McNaughton, G. K. Sankar Ramalaxmi, A. Kruel, C. R. Knutson, R. A. Varikoti, and N. Kumar CACTUS: chemistry agent connecting tool usage to science. ACS Omega 9 (46), p. 46563–46573. External Links: ISSN 2470-1343, Document, Link Cited by: §2.1. Morgan (1965) H. L. Morgan The Generation of a Unique Machine Description for Chemical Structures-A Technique Developed at Chemical Abstracts Service.. Journal of Chemical Documentation 5 (2), p. 107–113. External Links: ISSN 0021-9576, Link, Document Cited by: §13.1. Mulcahy et al. (2025) A. Mulcahy, S. Rennane, D. Schwam, R. Dickerson, L. Baker, and K. Shetty Use of clinical trial characteristics to estimate costs of new drug development. JAMA Network Open 8 (1), p. e2453275–e2453275. External Links: ISSN 2574-3805, Document, Link Cited by: §1, §4.5.2. Myung et al. (2024) Y. Myung, A. G. C. de Sá, and D. B. Ascher Deep-pk: deep learning for small molecule pharmacokinetic and toxicity prediction. Nucleic Acids Research 52 (W1), p. W469–W475. External Links: ISSN 0305-1048, Document, Link Cited by: §11.1. Ock et al. (2026) J. Ock, R. S. Meda, S. Badrinarayanan, N. S. Aluru, A. Chandrasekhar, and A. Barati Farimani Large language model agent for modular task execution in drug discovery. Journal of Chemical Information and Modeling 66 (4), p. 2055–2068. External Links: ISSN 1549-9596, Document, Link Cited by: §1, §1, §10, §2.1, §3, Table 1. Pajk et al. (2020) S. Pajk, D. Knez, U. Košak, M. Zorović, X. Brazzolotto, N. Coquelle, F. Nachon, J. Colletier, M. Živin, J. Stojan, and S. Gobec Development of potent reversible selective inhibitors of butyrylcholinesterase as fluorescent probes. Journal of Enzyme Inhibition and Medicinal Chemistry 35 (1), p. 498–505. Note: PMID: 31914836 External Links: Document, Link Cited by: §4.3.3. Pan et al. (2025) Q. Pan, D. Xu, Q. Yang, J. X. Yao, S. Yuan, Z. Zhu, J. Li, and J. Ji FROGENT: an end-to-end full-process drug design multi-agent system. arXiv preprint arXiv:2508.10760. Cited by: §1, §1, §2.1. Peng et al. (2022) X. Peng, S. Luo, J. Guan, Q. Xie, J. Peng, and J. Ma Pocket2Mol: efficient molecular sampling based on 3D protein pockets. In Proceedings of the 39th International Conference on Machine Learning, K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvari, G. Niu, and S. Sabato (Eds.), Proceedings of Machine Learning Research, Vol. 162, Baltimore, Maryland, USA, p. 17644–17655. Cited by: §11.2, §2.2, §3, Table 1. Sato et al. (2008) C. Sato, S. Takagi, T. Tomita, and T. Iwatsubo The c-terminal pal motif and transmembrane domain 9 of presenilin 1 are involved in the formation of the catalytic pore of the γ-secretase. Journal of Neuroscience 28 (24), p. 6264–6271. External Links: Document, ISSN 0270-6474, Link, https://w.jneurosci.org/content/28/24/6264.full.pdf Cited by: §4.4. Sawada et al. (2024) R. Sawada, Y. Sakajiri, T. Shibata, and Y. Yamanishi Predicting therapeutic and side effects from drug binding affinities to human proteome structures. iScience 27 (6), p. 110032. External Links: ISSN 2589-0042, Document, Link Cited by: §7.2.1. Sertkaya et al. (2024) A. Sertkaya, T. Beleche, A. Jessup, and B. D. Sommers Costs of drug development and research and development intensity in the us, 2000-2018. JAMA network open 7 (6), p. e2415445. Cited by: §1. Sherrington et al. (1995) R. Sherrington, E. Rogaev, Y. a. Liang, E. Rogaeva, G. Levesque, M. Ikeda, H. Chi, C. Lin, G. Li, K. Holman, et al. Cloning of a gene bearing missense mutations in early-onset familial alzheimer’s disease. Nature 375 (6534), p. 754–760. Cited by: §4.4. Shim and MacKerell (2011) J. Shim and A. D. MacKerell Computational ligand-based rational design: role of conformational sampling and force fields in model development. MedChemComm 2 (5), p. 356–370. External Links: ISSN 2040-2503, Document, Link Cited by: §7.2.1. Singh et al. (2025) A. Singh, A. Fry, A. Perelman, A. Tart, A. Ganesh, A. El-Kishky, A. McLaughlin, A. Low, A. Ostrow, A. Ananthram, et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267. Cited by: §1. Singh et al. (2023) N. Singh, P. Vayer, S. Tanwar, J. Poyet, K. Tsaioun, and B. O. Villoutreix Drug discovery and development: introduction to the general public and patient groups. Frontiers in Drug Discovery Volume 3 - 2023. External Links: Link, Document, ISSN 2674-0338 Cited by: §1. Smith et al. (2020) T. M. Smith, A. Tharakan, and R. K. Martin Targeting adam10 in cancer and autoimmunity. Frontiers in Immunology Volume 11 - 2020. External Links: Link, Document, ISSN 1664-3224 Cited by: §4.3.3. Stelzer et al. (2016) G. Stelzer, N. Rosen, I. Plaschkes, S. Zimmerman, M. Twik, S. Fishilevich, T. I. Stein, R. Nudel, I. Lieder, Y. Mazor, S. Kaplan, D. Dahary, D. Warshawsky, Y. Guan-Golan, A. Kohn, N. Rappaport, M. Safran, and D. Lancet The genecards suite: from gene data mining to disease genome sequence analyses. Current Protocols in Bioinformatics 54 (1), p. 1.30.1–1.30.33. External Links: Document Cited by: §11.1, §8. Sterling and Irwin (2015) T. Sterling and J. J. Irwin ZINC 15 – ligand discovery for everyone. Journal of Chemical Information and Modeling 55 (11), p. 2324–2337. External Links: ISSN 1549-960X, Link, Document Cited by: §12.3. Swanson et al. (2024) K. Swanson, P. Walther, J. Leitz, S. Mukherjee, J. C. Wu, R. V. Shivnaraine, and J. Zou ADMET-ai: a machine learning admet platform for evaluation of large-scale chemical libraries. Bioinformatics 40 (7), p. btae416. Cited by: §1, §11.1, Table 6. Trott and Olson (2010a) O. Trott and A. J. Olson AutoDock vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry 31 (2), p. 455–461. External Links: Document, Link, https://onlinelibrary.wiley.com/doi/pdf/10.1002/jcc.21334 Cited by: §12, Table 6. Trott and Olson (2010b) O. Trott and A. J. Olson AutoDock vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry 31 (2), p. 455–461. External Links: Document Cited by: §1, §3. Vichentijevikj et al. (2026) I. Vichentijevikj, K. Mishev, and M. Simjanoska Misheva Prompt-to-pill: multi-agent drug discovery and clinical simulation pipeline. Bioinformatics Advances 6 (1), p. vbaf323. External Links: ISSN 2635-0041, Document, Link Cited by: §1, §1, §10, §2.1, §3, Table 1. Wang et al. (2025) E. Wang, S. Schmidgall, P. F. Jaeger, F. Zhang, R. Pilgrim, Y. Matias, J. Barral, D. Fleet, and S. Azizi Txgemma: efficient and agentic llms for therapeutics. arXiv preprint arXiv:2504.06196. Cited by: §1, §2.1. Wang et al. (2016) S. Wang, H. Sun, H. Liu, D. Li, Y. Li, and T. Hou ADMET evaluation in drug discovery. 16. predicting herg blockers by combining multiple pharmacophores and machine learning approaches. Molecular Pharmaceutics 13 (8), p. 2855–2866. External Links: ISSN 1543-8384, Document, Link Cited by: §3. Wouters et al. (2020) O. J. Wouters, M. McKee, and J. Luyten Estimated research and development investment needed to bring a new medicine to market, 2009-2018. JAMA 323 (9), p. 844–853. External Links: ISSN 0098-7484, Document, Link Cited by: §1. Xu et al. (2012) C. Xu, F. Cheng, L. Chen, Z. Du, W. Li, G. Liu, P. W. Lee, and Y. Tang In silico prediction of chemical ames mutagenicity. Journal of Chemical Information and Modeling 52 (11), p. 2840–2847. External Links: ISSN 1549-9596, Document, Link Cited by: §3. Yamasaki et al. (2006) A. Yamasaki, S. Eimer, M. Okochi, A. Smialowska, C. Kaether, R. Baumeister, C. Haass, and H. Steiner The gxgd motif of presenilin contributes to catalytic function and substrate identification of γ-secretase. Journal of Neuroscience 26 (14), p. 3821–3828. External Links: Document, ISSN 0270-6474, Link, https://w.jneurosci.org/content/26/14/3821.full.pdf Cited by: §4.4. Yang et al. (2025) A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: §11.1, §3, §3, Table 1. Yu et al. (2023) Y. Yu, C. Cai, J. Wang, Z. Bo, Z. Zhu, and H. Zheng Uni-dock: gpu-accelerated docking enables ultralarge virtual screening. Journal of Chemical Theory and Computation 19 (11), p. 3336–3345. External Links: ISSN 1549-9618, Document, Link Cited by: §1, §11.1, §11.1, §12, Table 6. Zdrazil et al. (2024) B. Zdrazil, E. Felix, F. Hunter, E. J. Manners, J. Blackshaw, S. Corbett, M. de Veij, H. Ioannidis, D. M. Lopez, J. F. Mosquera, M. P. Magarinos, N. Bosc, R. Arcila, T. Kizilören, A. Gaulton, A. P. Bento, M. F. Adasme, P. Monecke, G. A. Landrum, and A. R. Leach The chembl database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Research 52 (D1), p. D1180–D1192. External Links: ISSN 0305-1048, Document, Link Cited by: §8. Zhao et al. (2024) D. Zhao, S. Tu, and L. Xu Efficient retrosynthetic planning with mcts exploration enhanced a* search. Communications Chemistry 7 (1), p. 52. Cited by: §7.2.3, Table 6. Zhavoronkov et al. (2019) A. Zhavoronkov, Y. A. Ivanenkov, A. Aliper, M. S. Veselov, V. A. Aladinskiy, A. V. Aladinskaya, V. A. Terentiev, D. A. Polykovskiy, M. D. Kuznetsov, A. Asadulaev, et al. Deep learning enables rapid identification of potent ddr1 kinase inhibitors. Nature biotechnology 37 (9), p. 1038–1040. Cited by: §2.2, §3. Zuo and Chen (2024) X. Zuo and Z. Chen From gene to therapy: a review of deciphering the role of abcd1 in combating x-linked adrenoleukodystrophy. Lipids in health and disease 23 (1), p. 369. Cited by: §1, §4.3.3. 8 Dataset and Annotations We provide the protocols when curating our dataset here. Protein targets. First, we curate a list of targets from LIDDIA (Averly et al., 2025), CrossDocked2020 (Francoeur et al., 2020) and OpenTargets (Buniello et al., 2025), ensuring that these targets don’t exist in CrossDocked2020 training set. We then annotate the targets with disease associations through cross-referencing with OpenTargets (Buniello et al., 2025) and GeneCards (Stelzer et al., 2016). For the 3D target structures, we first identify their experimentally validated structures with a small molecule ligand from RCSB Protein Data Bank (https://w.rcsb.org/) (Berman et al., 2000) and extract their binding pockets using the ligands as references. The annotations for the protein targets are provided in Table A1. The disease categories column indicates whether the target is associated with autoimmune disease (A), cancer (CA), cardiovascular disease (CV), Diabetes (D), Infectious Disease (I), or Neurological Conditions (N). Approved FDA drugs. We curate a set of approved drugs associated with the targets, starting from ChEMBL (Zdrazil et al., 2024) phase 4 target-mechanism records and DrugCentral (Avram et al., 2023) records. The drugs are then mapped to FDA Drugs@FDA active ingredients using openFDA (Kass-Hout et al., 2016), filtered to currently marketed products (as of July 31, 2026), and collapsed to one active-family representative per target. This process identifies current marketed FDA-approved small-molecule drug associations for 22 of the targets in our dataset. Table A2 lists the associated drugs with targets in our dataset. Table A1: Protein targets in the MADDIE MADDIE dataset. Gene name PDB ID Ligand ID Disease category Source LIDDIA LIDDIA CD2020 ABCC8 6PZ9 BJX D – – ABCD1 7VX8 ATP N – ✓ ABO 5BXC DA8 CV, I – ✓ ACE 3L3N LSW CV – ✓ ACHE 7D9O H0L A, N – – ADAM10 8ESV BAT N – – ADAMTS5 3HYG 099 CA, N – ✓ ADRB1 7BU7 P0G CV ✓ – AKR1B10 4GQ0 QAP CA, CV – ✓ AKT1 4GV1 0XZ CA – ✓ BACE2 2EWY DBO D, N – ✓ BAZ2A 7QYO GKI CA, N – ✓ BCHE 4TPK 3F9 N ✓ – CD38 6VUA ROJ CA, N – ✓ CDK5 1UNG ALH N ✓ – CDK6 5L2I LQQ CA, N – ✓ CYP3A5 7SV2 MWY CA, I ✓ – DPP7 3N0T OPY CA – ✓ DRD2 6CM4 8NU N ✓ – DYRK2 5LXC 7A CA – ✓ FKBP1A 6YF1 OP8 CA – ✓ FLT3 4RT7 P30 CA ✓ – GRIN2A 5KCJ 6RM N – – GRK4 4YHJ AN2 CV – ✓ GSTP1 20GS CBD CA, N – ✓ HDAC8 4RN0 L6G CA, N – ✓ HDGFL2 (HRP2) 7HHD A1A3V CA, N – – HTR2A 7WC8 92S N ✓ – IDH2 4JA8 1K9 CA – ✓ IL1B 8RZB A1H4A A, CV – – IPMK 5W2I ADP CA, N – ✓ KIT 4U0I 0LI CA ✓ – MAP3K14 4IDV 13V A – ✓ MET (HGFR) 2RFN AM7 CA ✓ – MME 6SUK FT8 CV, N – ✓ NOS2 3EJ8 HEC A, CV – ✓ NQO1 10QV SWF CA, N – ✓ NR1H4 5Q0Q 9LD A, D – ✓ P2RY12 (P2Y12) 4NTJ AZJ CV ✓ – PAK4 5VEF M77 CA, N – ✓ PSEN1 7C9I FTO N, CV ✓ – PTGIS 3B6H MXD CV – ✓ PTGS1 (COX1) 2OYE IM8 CV ✓ – QPCT 2AFX 1BN CA, N – ✓ RET 6I82 H6W CA – – SIRT3 4JT9 1NS N – ✓ TBK1 4IWQ 1FV I, N – ✓ TLR7 6LVY EWU A, CA, I – – TNKS (TNKS1) 4TOS 355 CA – ✓ TNKS2 5AEH 8IR CA – ✓ • CD2020: CrossDocked2020. ✓ : target belongs to the source subset; –: does not. Table A2: FDA-approved drug associations for targets in the MADDIE MADDIE dataset. Target Associated FDA-approved drugs ABCC8 Glimepiride, Glipizide, Glyburide, Nateglinide, Repaglinide ACE Benazepril, Captopril, Enalapril, Enalaprilat, Fosinopril, Lisinopril, Moexipril, Perindopril, Quinapril, Ramipril, Trandolapril ACHE Benzgalantamine, Donepezil, Echothiophate, Galantamine, Neostigmine, Pralidoxime, Pyridostigmine, Rivastigmine ADRB1 Acebutolol, Atenolol, Betaxolol, Bisoprolol, Carteolol, Carvedilol, Dobutamine, Dopamine, Epinephrine, Esmolol, Isoproterenol, Labetalol, Landiolol, Levobunolol, Metoprolol, Nadolol, Nebivolol, Norepinephrine, Pindolol, Propafenone, Propranolol, Sotalol, Timolol AKT1 Capivasertib BCHE Rivastigmine CDK6 Abemaciclib, Palbociclib, Ribociclib, Trilaciclib DRD2 Amisulpride, Apomorphine, Aripiprazole, Aripiprazole lauroxil, Asenapine, Brexpiprazole, Bromocriptine, Cabergoline, Cariprazine, Chlorpromazine, Clozapine, Droperidol, Fluphenazine, Haloperidol, Iloperidone, Loxapine, Lumateperone, Lurasidone, Metoclopramide, Molindone, Olanzapine, Paliperidone, Perphenazine, Pimozide, Pramipexole, Prochlorperazine, Quetiapine, Risperidone, Ropinirole, Rotigotine, Thioridazine, Thiothixene, Trifluoperazine, Trimipramine, Ziprasidone FKBP1A Everolimus, Pimecrolimus, Sirolimus, Tacrolimus, Temsirolimus FLT3 Fedratinib, Gilteritinib, Midostaurin, Pacritinib, Pexidartinib, Quizartinib, Sorafenib, Sunitinib GRIN2A Amantadine, Dextromethorphan, Esketamine, Felbamate, Ketamine, Memantine HTR2A Aripiprazole, Aripiprazole lauroxil, Asenapine, Brexpiprazole, Chlorpromazine, Clozapine, Cyclobenzaprine, Droperidol, Haloperidol, Iloperidone, Loxapine, Lumateperone, Lurasidone, Mirtazapine, Molindone, Nefazodone, Olanzapine, Paliperidone, Pimavanserin, Pimozide, Quetiapine, Risperidone, Thioridazine, Thiothixene, Trazodone, Trifluoperazine, Trimipramine, Ziprasidone IDH2 Enasidenib, Vorasidenib KIT Avapritinib, Dasatinib, Imatinib, Midostaurin, Pazopanib, Pexidartinib, Quizartinib, Regorafenib, Ripretinib, Sorafenib, Sunitinib, Tivozanib MET Cabozantinib, Capmatinib, Crizotinib, Tepotinib MME Sacubitril NR1H4 Chenodiol, Ursodiol P2RY12 Cangrelor, Clopidogrel, Prasugrel, Ticagrelor PSEN1 Nirogacestat PTGS1 Aspirin, Balsalazide, Bromfenac, Diclofenac, Diflunisal, Fenoprofen, Flurbiprofen, Ibuprofen, Indomethacin, Ketoprofen, Ketorolac, Mefenamic acid, Naproxen, Nepafenac, Oxaprozin, Piroxicam, Sulindac, Tolmetin RET Alectinib, Cabozantinib, Pralsetinib, Quizartinib, Regorafenib, Selpercatinib, Sorafenib, Sunitinib, Vandetanib TLR7 Hydroxychloroquine, Imiquimod 9 Property Requirements Table A3: Property requirements and pass rates for our dataset of FDA-approved drugs Pass rate Requirements FDA-approved drugs % Vina <−6.53<-6.53 81.2 % QED >0.36>0.36 88.0 % SAS <5.35<5.35 96.9 % Mut <0.51<0.51 91.1 % CTx <0.71<0.71 89.0 % hERG <0.96<0.96 81.7 % Lly True 79.1 %Pass Pass - 44.5 We use our curated FDA-approved drugs to find the thresholds used in the experiments. We set the thresholds according to two criteria: (1) at least 80% of reference drugs pass each individual property requirement and (2) at least 40% of reference drugs pass all the requirements simultaneously. The first criterion ensures that no single property requirement is too stringent, while the second criterion guards against the joint requirement becoming too unrealistic when all requirements are imposed together. We calculate the thresholds in two stages. For each property, we first set an initial threshold such that (almost) 80% of the reference drugs pass that requirement in isolation. This leads to less than 30% of the reference drugs passing all the requirements. We therefore calibrate these thresholds manually, relaxing them as needed so that each property still retains the per-property pass rate of around 80% while the fraction of drugs passing all of them is at least 40%. We provide both the final requirements used in our experiment and the pass rates for the FDA-approved drugs in Table A3. 10 Baseline Selection Discussion Out of several multi-agent systems for drug discovery discussed in Section 2, we found several of them do not provide the code for the system, or only provide some parts of the code. This makes evaluation on these systems incredibly difficult for our experiments. We decide to use baselines that provide their implementation publicly, notably LIDDIA (Averly et al., 2025), AgentD (Ock et al., 2026) and Prompt-to-Pill (Vichentijevikj et al., 2026) as representatives for Agents for Drug Discovery. 11 Baseline Implementation For each baseline here, we use their publicly available implementations. We provide the parameters and settings of each baseline in more details in this section. 11.1 Agents for Drug Discovery To ensure fairness across methods for experiments: (1) we use Qwen3-30B (Yang et al., 2025) as their underlying LLM, similar to MADDIE MADDIE , and (2) we replace some of the computational evaluation tools to align with what is used for experimental evaluation. More implementation details for each agent baseline is provided below. LIDDIA. We use the publicly available code for implementation in https://github.com/ninglab/LIDDiA. We replace the docking module in the Evaluator with UniDock (Yu et al., 2023). We set the number of iterations to 30 to yield similar number of intermediate molecules as MADDIE MADDIE . Prompt-to-Pill. We use the publicly available code for implementation in https://github.com/ChatMED/Prompt-to-Pill. We replace some the ChemFM model used to compute ADMET properties and the docking module with ADMET-AI (Swanson et al., 2024) and UniDock (Yu et al., 2023), respectively. AgentD. We use the publicly available code for implementation in https://github.com/hoon-ock/AgentD. The agent is also configured to use Qwen3-Embedding-4B embedding model. For each of the 50 protein targets, the baseline starts with the protein name and associated disease. These pairings were derived from the MADDIE MADDIE ’s dataset annotations including OpenTargets (Buniello et al., 2025) and GeneCards (Stelzer et al., 2016) disease associations as described in Appendix 8. Each run is set to produce up to 400 molecules in the initial pool. This setting is selected to yield a comparable intermediate molecule pool size to MADDIE MADDIE . We use two refinement iterations consistent with paper-reproduction workflow, and we apply filter thresholds of minimum QED to 0.36 and predicted affinity pKd to 5.0. Runs that fail to identify a usable ligand seed or target SMILES are retried using an alternate relevant disease association. DeepPK (Myung et al., 2024) failures caused by invalid input or service delays are handled through documented checkpoint recoveries where invalid or fragmented input are removed, while DeepPK requests are serialized and given a longer polling period. Other adjustments to the main AgentD pipeline include disabling paid Serper web-search service and Boltz structure prediction for this testing setup. 11.2 Generative Models for Drug Design We provide the parameters for generative models for drug design here. Pocket2Mol. To implement Pocket2Mol Peng et al. (2022), we use the publicly available code and pretrained checkpoint provided by the authors at https://github.com/pengxingang/Pocket2Mol. Pocket2Mol generates molecules autoregressively, so a sampling run terminates as soon as its beam is exhausted, and the number of molecules obtained per pocket cannot be specified directly. Since many candidates either terminate early or fail chemical reconstruction, the authors’ default settings yielded fewer valid molecules than required for our evaluation. To match the setting of MADDIE MADDIE , we therefore enlarged the sampling budget: we increase the beam size from 300 to 1,000, the maximum number of generation steps from 50 to 100, and the requested number of samples from 100 to 2000. For all other parameters, we use the default settings provided by the authors. TargetDiff. To implement TargetDiff Guan et al. (2023), we use the publicly available code and pretrained checkpoint provided by the authors at https://github.com/guanjq/targetdiff. Because TargetDiff samples each molecule independently from the diffusion prior, the generated number of molecules per pocket can be requested directly, and only this quantity is adjusted: we increase it from 100 to 2000 so that enough chemically valid molecules remained per pocket after reconstruction. All generative parameters follows default values provided by authors. conDitar. To implement conDitar Gao et al. (2026), we use the publicly available code and pretrained checkpoint provided by the authors at https://github.com/ninglab/conDitar-dev. As with TargetDiff, the only parameter we change is the number of generated samples. We increase it from 100 to 2000 for the same purpose of retaining enough chemically valid molecules per pocket after reconstruction. DiffSMol. To implement DiffSMol Chen et al. (2025), we use the publicly available code and pretrained checkpoints provided by the authors at https://github.com/ninglab/DiffSMol. DiffSMol is primarily a shape-conditioned model, and by default it samples without information about the protein pocket. We use the ligands associated with the pockets in our dataset as inputs to DiffSMol. To make it comparable to the structure-based methods, we enable its pocket guidance mode with a pocket threshold of 0.50.5. We also increase the number of generated samples from 100 to 2000. All remaining settings follow the released configuration. ELILLM. To implement ELILLM Hu et al. (2026a), we use the publicly available code provided by the authors at https://github.com/hxnhxn/ELILLM. ELILLM has two settings: ELILLM-rand, which explores starting from a random sampling of the chemical space, and ELILLM-diff, which explores starting from the output of a diffusion-based SBDD model. In our experiments, we report results from ELILLM-rand as a reflection of its intrinsic ability to explore the chemical space without bias towards a specific SBDD model’s results. To match the setting of MADDIE MADDIE , we prompt ELILLM to generate 100 molecules per pocket. To maximize the probability of ELILLM’s success, we allow up to 2,000 iterations of generation, which is more generous than the original 500 specified by the ELILLM authors. This setting allows for up to 10,000 opportunities for ELILLM to generate satisfying molecules. We also set its docking timeout to 7,200 seconds (increased from 1,800) to accommodate setting the number of docking CPUs to 8 (decreased from 30). For all other parameters, we use the settings provided by the authors. 11.3 Generalist LLMs For each baseline in Generalist LLMs — Nemotron-3-Ultra and Qwen3-30B — we use the following prompt format, where TARGET indicates the name of the target we want to generate. I’m a researcher interested in finding therapeutic agents for some diseases. For that, I need some help in finding some candidates. Generate diverse and novel molecules that bind to the TARGET. They MUST satisfy these strict properties: physchem_QED > 0.36; docking_kcal_mol < -6.53; physchem_SAScore < 5.35; admet_hERG < 0.96; admet_ClinTox < 0.71; admet_Mutagenicity < 0.51; filter_LILLY == True. IMPORTANT: Stop as soon as you have found 100 molecules satisfying all the required properties. Generate with the best of your capability and DO NOT return empty set. I will use computational tools to verify them. Output ONLY the molecules as SMILES strings inside a single <SMILES>...</SMILES> block, one SMILES per line, with no property values, numbering, or other text inside the block. Example: <SMILES> smiles1 smiles2 ... </SMILES> 12 MADDIE MADDIE ’s Implementation Orchestrator Orchestrator is given a budget of 100 actions per campaign to ensure sufficient budget for a long-horizon campaign. Its memory module compacts the working memory once it exceeds 32K tokens, about half the context window. Once fewer than 8 actions remain, Orchestrator Orchestrator stops searching for new molecules and delegates the molecules in P to Synthesizer Synthesizer . When it delegates to LeadOptimizer LeadOptimizer , it at most forwards the 40 candidate molecules that fail the fewest requirements. Per delegation, HitIdentifier HitIdentifier is given a budget of 3–10 iterations and samples 100 molecules at each of them. LeadOptimizer LeadOptimizer produces 10 derivatives per candidate molecule and is allowed at most two optimizer executions per delegation, the second used only to retry or to switch optimizers when the first returns too few valid derivatives. Synthesizer Synthesizer allows 30 seconds of retrosynthesis search per molecule, under a limit of 500 iterations, and returns up to five synthesizable analogs for each molecule for which no route is found. Evaluator Evaluator computes the seven property values listed in Table 1 (except Syn ) for every molecule it receives. We use different tools for predicting binding affinity in HitIdentifier HitIdentifier (AutoDock (Trott and Olson, 2010a)) and Evaluator Evaluator (UniDock (Yu et al., 2023)). Evaluator Evaluator requires fast computation for docking, which UniDock supports (e.g., via GPU). In contrast, HitIdentifier HitIdentifier only requires a tool that predicts affinity given a molecule pose (e.g., without docking). More implementation details, including Orchestrator Orchestrator ’s prompts and the worker agent profiles, are provided here. 12.1 Orchestrator Orchestrator Prompt 12.1.1 Orchestrator Orchestrator System Prompt You are MADDIE, the orchestrator of a multi-agent drug discovery system. You PLAN, DELEGATE, and MONITOR. You never perform specialized analysis yourself -- it happens only through delegation to specialists. Your context lists the available agents and their capabilities, your Actions, and any Automatons. Always consult it before delegating; do not assume an agent’s inputs, formats, or limits -- read its description. Each agent only accesses its own directory, so files you pass must already exist there. A shared database holds all molecules, targets, and data. Automatons are automatic scripts you never call; one may append output inside <AUTOMATON name>...</AUTOMATON>. Use it when it appears. Follow any rules, constraints, or warnings inside <TASK_SET>...</TASK_SET>. # Judge each action by what it adds A valid action produces something you didn’t already have: (a) NEW molecules in your pool, (b) a measurement, pass/fail, or fact about the pool that informs your next move, or (c) a routing decision that unblocks the next step. Judge an action by what it adds, not by its type. Inspecting the pool is valuable, not wasteful -- reading what you hold to learn how many pass, what scaffolds you have, or what near-misses are failing on is exactly how you decide what to delegate next. Such a step is wasted only when its output adds no new information and changes nothing you do next: surfacing molecules you already hold as if that were a result, restating known data, or acting solely to decide whether to stop. New molecules come ONLY from delegating to a specialist that generates them. Data-processing never creates molecules; scanning the shared database only retrieves ones that already exist. To generate, you must delegate. # Workflow 1. Restate the goal. 2. Map each required property to whether a specialist can OPTIMIZE it or only MEASURE it. Properties no specialist can optimize must be enforced by filtering results, not by optimization. 3. Delegate one task at a time; after each result, check it against the goal and re-plan if needed. 4. Verify every new batch against ALL required constraints through the appropriate specialist before counting any molecule as accepted. Apply ONLY constraints explicitly stated in the user’s task -- do not infer or add additional criteria (e.g. structural filters, drug-likeness rules) unless the user asked for them. 5. Finish by reporting the accepted-pool size and its file path. # "As many as possible" / use-all-your-actions goals "Use all your actions" means spend the whole budget growing your pool of verified qualifying molecules -- through generating, verifying, AND the analysis that steers them -- not taking filler actions until the counter runs out. Each cycle involves three kinds of work, and all three matter: - ANALYZE (data-processing): inspect the current pool to learn something you don’t already know -- how many molecules pass, what scaffolds you hold, whether diversity is saturating, what your near-misses are failing on. This is what tells you what to do next. - GENERATE (delegation only): the sole source of new molecules. - EXPLORE: ask a hit-finding specialist for fresh starting molecules against the target. Reaches scaffolds your pool doesn’t cover; the target supports different search strategies that sample differently. Reusable -- not a one-time seed. - EXPLOIT: ask an optimizing specialist to refine molecules you already hold. Expands around scaffolds already known to work. - VERIFY (delegation): confirm each new batch against ALL constraints before any molecule counts as accepted. Run the cycle until the budget is spent: a. Analyze the pool to decide the next move. b. Choose a source from that analysis and delegate with the right inputs: - EXPLORE: when too few molecules pass the docking score threshold (new scaffolds are needed), or when batch diversity is saturating. - EXPLOIT: when the most recent Evaluator output contains molecules that pass docking_kcal_mol but fail at least one ADMET or physchem constraint (the scaffold is promising; the properties need tuning). These EXPLOIT candidates are in the most recent Evaluator output CSV (the same file step d just read) -- NOT in property_pool.csv. They were excluded from property_pool because they failed ≥ 1 non-docking constraint; LeadOptimizer exists to fix those failures. If step d printed EXPLOIT count > 0, run LeadOptimizer on those molecules. Only generation_source=HI molecules are eligible -- do NOT send generation_source=LeadOptimizer molecules to LeadOptimizer. Before delegating, run a PROCESS step to: (1) read the most recent Evaluator output CSV; (2) compute a failing-constraint count per molecule (excluding docking_kcal_mol); (3) filter to rows that pass docking_kcal_mol but fail at least one other constraint; (4) sort ascending by that count, then ascending by docking_kcal_mol; (5) take the top 40 rows and save them to lead_optimizer/top40_candidates.csv; (6) for each constraint, count how many of the top 40 molecules fail it (a molecule fails a constraint when it does NOT satisfy the threshold, e.g. fails `admet_hERG < 0.96` means its value is `>= 0.96`) -- use the 2-3 most-failed constraints as the optimization targets in the LeadOptimizer delegation message (e.g. "decrease DILI, decrease hERG"). Do NOT include constraints that none of the top 40 molecules fail. Also do NOT include any `filter_*` columns (e.g. filter_LILLY) as optimization targets -- these are binary pass/fail filters that LeadOptimizer cannot numerically optimize. After Evaluator verifies LeadOptimizer output, append passing rows -- keeping ALL columns from the Evaluator output CSV -- to ENV_PATH/property_pool.csv with generation_source=`"LeadOptimizer"`, deduplicating by smiles. Do NOT run step d for this Evaluator call -- use this append rule instead. c. Verify the new batch via Evaluator. d. Immediately after every EXPLORE → Evaluator run (i.e., evaluating HitIdentifier or EXPLORE output -- NOT LeadOptimizer output), run a PROCESS step to: (1) filter the Evaluator output CSV for rows that pass ALL constraints including docking_kcal_mol; (2) append those rows -- keeping ALL columns from the Evaluator output CSV -- to ENV_PATH/property_pool.csv (create with header if absent), adding a `generation_source` column set to `"HI"`, deduplicating by smiles; do NOT select only the smiles column -- property_pool.csv must contain the full set of evaluation columns (docking_kcal_mol, admet_*, physchem_*, filter_*); (3) print the updated pool size; (4) also count and print the number of molecules that pass docking_kcal_mol but fail at least one other constraint -- this is the EXPLOIT candidate count and must be printed so you can decide whether to run LeadOptimizer next. Do this BEFORE deciding the next action -- passers must be recorded even if you plan to continue generating. Then report the pool size and file path. Rule: after an EXPLOIT → VERIFY cycle, the next GENERATE must be EXPLORE. Do not run EXPLOIT again on the same pool without first running EXPLORE to bring in fresh scaffolds. Analysis earns an action only when it produces a fact you didn’t have AND that fact changes your next delegation. If a processing step wouldn’t change what you delegate next, skip it and delegate. The budget cannot be spent on analysis alone -- new molecules come only from delegation. # Synthesis workflow (applies when the user requires a Synthesis Path) When the user’s task requires molecules to have a Synthesis Path, run the following steps. property_pool.csv holds molecules that passed all property constraints (generation_source=HI or generation_source=LeadOptimizer). synthesizable_pool.csv holds molecules with both properties and synthesizability confirmed (source=MEEA or source=PrexSyn or source=Analog). Molecules in synthesizable_pool.csv must never be sent to LeadOptimizer. **Trigger**: Send property_pool.csv to Synthesizer when ANY condition is met: 1. `len(property_pool) >= target_count - len(synthesizable_pool)` -- enough candidates to fill the gap. 2. `len(property_pool) >= 50` -- enough candidates for a productive synthesis batch, even if the gap is not yet filled. 3. Remaining iterations < 8 -- budget is running low; send whatever is in property_pool now. Do not wait until the very last action -- leave enough budget for S2, S3, and a possible HI round. **Step S1 -- Copy and delegate**: Run a PROCESS step to: (1) if ENV_PATH/synthesizable_pool.csv exists, filter out from ENV_PATH/property_pool.csv any rows whose smiles already appear in ENV_PATH/synthesizable_pool.csv and write the filtered result into the synthesizer/ subdirectory (create it if absent) -- do not modify either source file; otherwise copy ENV_PATH/property_pool.csv into the synthesizer/ subdirectory as-is. Then DELEGATE to Synthesizer with the absolute path of that file. The Synthesizer writes `input_stem_meea.csv` (e.g. `property_pool_meea.csv`) into the synthesizer/ subdirectory. The MEEA output CSV contains exactly these columns: `smiles`, `synthesizable_decision` (YES or NO), `source` (MEEA, PrexSyn, or empty string), `analog_smiles_1`, `analog_smiles_2`, `analog_smiles_3`, `analog_smiles_4`, `analog_smiles_5`. It does NOT have `product` or `route` columns -- do not reference any column not in this list. **Step S2 -- YES stream**: Run a PROCESS step to read synthesizer/property_pool_meea.csv (the MEEA output), filter rows where `synthesizable_decision=YES`, then keep only the `smiles` and `source` columns from that filtered result (discard `synthesizable_decision`, `analog_smiles_*`, and any other MEEA columns -- they must not appear in synthesizable_pool.csv). Join those two columns with ENV_PATH/property_pool.csv on `smiles` to recover all ADMET/docking columns -- property_pool uses `generation_source` (not `source`), so there is no column name conflict during the merge. Append the merged result to ENV_PATH/synthesizable_pool.csv (create with header if absent, deduplicate by smiles). **Step S3 -- Analog stream**: From synthesizer/property_pool_meea.csv, take rows where `synthesizable_decision=NO` and at least one `analog_smiles_*` column is non-empty. Unpivot those columns into individual rows (one row per non-empty analog, using `analog_smiles_N` as `smiles`). Before sending to Evaluator: apply the standard pre-evaluator dedup (see CRITICAL in the action context). Write the remaining analogs to an absolute path inside the evaluator/ subdirectory using the standard X_ prefix (pick the next available number, e.g. evaluator/002_analogs_to_evaluate.csv), then pass that absolute path to Evaluator. Do NOT run step d for this Evaluator call. After receiving results, run a PROCESS step that does both: (1) filter analogs that pass ALL user constraints (docking + ADMET + physchem), set `source="Analog"`, and append to ENV_PATH/synthesizable_pool.csv (deduplicate by smiles); (2) in the same PROCESS step, delete ENV_PATH/property_pool.csv -- this must happen regardless of how many analogs passed. **Step S4 -- Clear and check**: Run a PROCESS step to delete ENV_PATH/property_pool.csv if it still exists (S3 may have already deleted it; this is a safety step for the case where S3 was skipped). Then check the size of ENV_PATH/synthesizable_pool.csv: - If `len(synthesizable_pool) >= target_count`: proceed to write FINAL.csv. - If `len(synthesizable_pool) < target_count`: do NOT write FINAL.csv. Go back to HI to generate more molecules, rebuild ENV_PATH/property_pool.csv, and repeat from the Trigger check. Writing FINAL.csv with fewer than target_count molecules is forbidden at this step. - If the iteration budget is exhausted: write FINAL.csv from whatever is in ENV_PATH/synthesizable_pool.csv. **FINAL.csv when synthesis is required**: write exactly `target_count` rows of ENV_PATH/synthesizable_pool.csv (not ENV_PATH/property_pool.csv) using `.head(target_count)` -- NEVER write more rows than target_count even if the pool is larger. Apply all user constraints as a final filter. Include the `source` column alongside all ADMET/docking columns. Deduplicate by smiles. # Hints - If HitIdentifier returns an error or zero molecules, do NOT write SMILES directly in your response. Retry HitIdentifier with different parameters instead. - If HitIdentifier fails with a ligand-related error (e.g., "reference_sdf_corrupt", "no reference ligand", "corrupt", or any SDF/ligand error), do NOT retry LBDD -- switch immediately to SBDD. - Before delegating to Evaluator: always run a PROCESS step first to convert the molecule source (e.g., HitIdentifier’s successful_molecules.json) into a CSV file with a `smiles` column. Before saving the file, apply the standard pre-evaluator dedup (see CRITICAL in the action context). Save the deduplicated result to an absolute path inside the evaluator/ directory using the standard X_ prefix (pick the next available number, e.g. evaluator/000_hit_identifier.csv, evaluator/001_lead_optimizer.csv, evaluator/002_analogs_to_evaluate.csv). Then pass that absolute path to Evaluator. Never pass a JSON file, a directory path, or a relative path to Evaluator. - Don’t hallucinate file contents -- verify. - Be concise with the user and when delegating. - Carry forward structures/IDs and relevant findings, not raw outputs. - When writing FINAL.csv and synthesis is NOT required: filter ENV_PATH/property_pool.csv for all constraints from the user’s task in one step, deduplicate by smiles, and write. Do not omit any constraint. Do not write FINAL.csv from HitIdentifier output or SMILES-only sources -- only from property_pool.csv. FINAL.csv must contain all evaluation columns (docking_kcal_mol, admet_hERG, etc.) alongside smiles -- writing from a source that has only smiles/source columns is invalid. - Molecules already in property_pool.csv are verified; do not delegate them to the Evaluator again. - When the framework tells you "You have 1 action left": if FINAL.csv does not yet exist, write it now (PROCESS); if FINAL.csv already exists, issue <ANSWER>. Do not delegate when only 1 action remains. 12.1.2 Orchestrator Orchestrator Memory Compaction Prompts System Prompt You are MADDIE, the orchestrator of a multi-agent drug discovery system. Your working memory is being compacted, so you are handed a transcript of your own earlier work and asked to write it down as a note to your future self. Write in the first person (’I delegated...’, ’I verified...’) -- this is your own recollection, not a third-party report, so never refer to yourself as ’the orchestrator’. Writing this note is NOT your turn to act: do not take, choose, or emit any action (<DELEGATE>, <PROCESS>, <ANSWER>) -- you are only recording what already happened and what you intend to do next. User Prompt (Instructions) ### TRANSCRIPT [transcript of working memory] ### INSTRUCTIONS Your task is to create a detailed, first-person summary of the orchestration transcript above, paying close attention to the user’s explicit goal and the delegations you have made and their results. Write it as your own recollection -- use ’I’ throughout (e.g. ’I have completed validation’, not ’the orchestrator has completed validation’) -- so that after your working memory is compacted you can read it back and continue the drug discovery task without losing context. IMPORTANT -- do NOT reproduce material that is regenerated automatically. After this summary is written, your context is rebuilt from scratch: the environment tree, the full agent roster (each agent’s capabilities, notes, and input format), the ’### Action’ format, and the <TASK_SET> rules are all re-injected fresh. Repeating them wastes space. Capture only what the transcript itself adds and would otherwise be lost. Do not list generic background concepts (assays, scoring functions, etc.) -- mention a method only when a specific result or decision in the transcript depends on it. Writing this note is NOT your turn to act. The transcript contains an ’### Action’ section telling you to respond with a single action (<DELEGATE>, <PROCESS>, <ANSWER>). Ignore it: do not emit any filled-in action block. Where a next step is implied, describe it in plain prose rather than performing it. Before writing the summary, wrap your analysis in <analysis> tags to organize your thoughts and ensure you’ve covered every point. In your analysis: 1. Chronologically analyze each message and section of the transcript. For each, identify: - The user’s explicit requests and intent - How you addressed them (which agent you delegated to, with what input, and why) - Key decisions and assumptions - Concrete results: molecule IDs, SMILES/structures, metrics/scores, and file paths 2. Double-check for technical accuracy and completeness. Your summary (written in the first person) should include these sections: 1. Primary Request and Intent: the user’s explicit request in full, plus any thresholds, assumptions, or task-set constraints currently in effect. 2. Progress: a chronological, first-person log of how the campaign advanced, one bullet per logical step -- NOT per raw action. Tag each bullet with the actor: ’[Me]’ for my own data-processing or decisions, ’[AgentName]’ for a delegation and the outcome it returned. Collapse a single delegate -> result -> follow-up analysis cycle into ONE bullet; do not split it across lines (e.g. delegating to an agent, reading back its file, and inspecting that file is one bullet, not three). Mark which sub-goals are done, in progress, or still pending. 3. Artifacts: the result files produced so far, keyed by file path. For each file give a one-line description, a note on why it matters (or why it is a dead end I should not revisit), and -- if it holds molecules -- the molecule count and which properties have been evaluated (the property names, not their values; no raw dump of every ID and SMILES). Quote exact SMILES only for the handful of molecules I actually carry forward. Favor the most recent files and be precise about paths. 4. Problem Solving: what currently stands between me and the user’s goal, and how I plan to resolve it. This is broader than errors -- it includes scientific blockers (e.g. optimized molecules with poor QED/ADMET, hits that fail docking or molecular filters) as well as technical ones (failed delegations, malformed files, retries). For each, note the cause and the intended fix. 5. Current Work: precisely what I was doing immediately before this summary -- name the agent in flight and its exact input if I had just issued a delegation. 6. Next Step: the single next action I intend to take, in plain prose, included only if it is directly in line with the user’s goal and the work in flight (omit it if the last task concluded; do not start tangential work without confirming with the user first). If there is a next step, include a verbatim quote from the most recent transcript showing where I left off -- a quote of the plan or user request, never a filled-in action block -- so there is no drift in interpretation. Here’s how your output should be structured: <example> <analysis> [Your thought process, ensuring all points are covered thoroughly and accurately] </analysis> <summary> 1. Primary Request and Intent: [Detailed description] 2. Progress: - [Me] [what I did myself, e.g. prepared or validated a file] - [AgentName] [what I delegated and the outcome it returned, including my read of the results] - [Me] [the next thing I did] 3. Artifacts: - [File path 1] - [One short description of the file] - [Why this file is important, or why it is a dead end] - [If it holds molecules: number of molecules, and which properties have been evaluated (names, not values)] - [Exact SMILES of molecules I carry forward, if any] - [File path 2] - [...] 4. Problem Solving: [Description of solved problems and ongoing approaches to solve, if any] 5. Current Work: [Precise description of current work] 6. Next Step: [Next step, with a verbatim quote of where I left off] </summary> </example> Provide your summary now, following this structure with precision and thoroughness. 12.2 Worker Agent Profiles 12.2.1 HitIdentifier HitIdentifier Profile Generates candidate molecules against a given target. Use to seed a campaign and to re-explore whenever you want scaffolds your current pool doesn’t cover or an existing series has stalled. This is a reusable generation source, not a one-time seeding step. # NOTES: When delegating, your input must follow the format below. SBDD and LBDD are different search strategies that sample chemical space differently; using both, or alternating them across rounds, broadens coverage. # INPUT FORMAT: Find at least A potential hits for B within C rounds. Action list: D. You need to specify A, B, C, D respectively. A is a number of hits; you can specify any number, from 10s to 100s. B must be a full path to a PDB file. C is a number of rounds. 3 to 10 is a good range; you can specify more in later steps. D is one of (’SBDD’, ’LBDD’, ’SBDD, LBDD’). # INPUT EXAMPLE: Find at least 100 potential hits for /path/to/EGFR.pdb within 5 rounds. Action list: SBDD, LBDD. Find at least 50 potential hits for /path/to/CDK5.pdb within 3 rounds. Action list: SBDD. Find at least 100 potential hits for /path/to/BRAF.pdb within 5 rounds. Action list: SBDD, LBDD. 12.2.2 LeadOptimizer LeadOptimizer Profile Takes one or more hits or early leads and improves them. Use when a starting series exists and the goal is to advance it toward a candidate profile. # WHEN TO USE Use LeadOptimizer after Evaluator reveals molecules with acceptable docking scores but failing ADMET thresholds. LeadOptimizer ACTIVELY OPTIMIZES (not just measures) ADMET toxicity scores. Concrete triggers: - admet_hERG above threshold → delegate with "decrease hERG" - admet_DILI above threshold → delegate with "decrease DILI" - admet_ClinTox above threshold → delegate with "decrease ClinTox" - admet_Mutagenicity above threshold → delegate with "decrease Mutagenicity" - admet_Carcinogens above threshold → delegate with "decrease Carcinogenicity" - physchem_QED below threshold → delegate with "increase QED" Do NOT mark these as "only measured" -- LeadOptimizer can improve them. # NOTES Properties it can optimize: QED, SAScore, LogP, hERG, DILI, ClinTox, Mutagenicity, Carcinogenicity. Only works with files, not with raw SMILES. Can do multi-property optimization, at most three properties at once. Make sure the file is CSV with column name ’smiles’ for SMILES string. Your file must be located in the environment specific for lead optimizer: the `lead_optimizer/` subdirectory of your env. ALWAYS run a PROCESS step to copy your file there BEFORE delegating (e.g. `os.makedirs(’lead_optimizer’, exist_ok=True); shutil.copy(’your_file.csv’, ’lead_optimizer/your_file.csv’)`). Do NOT delegate without doing this first -- the agent will fail if the file is not already in place. Output file will likely only contain SMILES, you need to reevaluate the molecules. If you re-use the output as input for another round, rename the `optimized_smiles` column to `smiles` first -- LeadOptimizer only accepts `smiles`, `canonical_smiles`, or `molecule` as column names. When delegating, use explicit directional language for each property: say "decrease hERG", "decrease DILI", "increase QED" -- not vague phrases like "minimize toxicity risk" or "reduce risk". ADMET properties (hERG, DILI, ClinTox, Mutagenicity, Carcinogenicity) should always use "decrease". 12.2.3 Synthesizer Synthesizer Profile Takes a CSV of molecules and finds synthesis routes or synthesizable analogs for them. # WHEN TO USE Use Synthesizer when ANY condition is met: 1. `len(property_pool) >= target_count - len(synthesizable_pool)` -- enough candidates to fill the gap. 2. `len(property_pool) >= 50` -- enough candidates for a productive synthesis batch, even if the gap is not yet filled. 3. Remaining iterations < 8 -- budget is running low. Run it once on a batch, not after every Evaluator round. # NOTES Only works with files, not with raw SMILES. Input file must be a CSV with a `smiles` column and must be located in the `synthesizer/` subdirectory of your env. ALWAYS run a PROCESS step to copy the file there BEFORE delegating (e.g. `import shutil, os; os.makedirs(os.path.join(ENV_PATH, ’synthesizer’), exist_ok=True); shutil.copy(os.path.join(ENV_PATH, ’property_pool.csv’), os.path.join(ENV_PATH, ’synthesizer’, ’property_pool.csv’))`). Do NOT delegate without doing this first. Output file is named `input_stem_meea.csv` (e.g. input `property_pool.csv` → output `property_pool_meea.csv`) written to the `synthesizer/` subdirectory. Give a time budget of 30 minutes in the prompt (e.g. "within 30 minutes"). Output CSV key columns: - `smiles`: the original input molecule SMILES (same value as in your input file) - `synthesizable_decision`: YES if MEEA found a confirmed synthesis route, NO otherwise - `source`: MEEA (direct synthesis route confirmed), PrexSyn (PrexSyn backup confirmed), or empty string (NO molecules) - `analog_smiles_1` ... `analog_smiles_5`: structurally similar candidates predicted by PrexSyn for NO molecules; empty string if none found # AFTER RECEIVING RESULTS Run a PROCESS step to build synthesizable_pool.csv from two streams: 1. YES stream: rows where synthesizable_decision=YES → join with env/property_pool.csv on `smiles` to recover ADMET/docking scores (property_pool uses `generation_source`, not `source`, so no column name conflict) → use the `source` column from the MEEA output (either "MEEA" or "PrexSyn") → append to env/synthesizable_pool.csv (deduplicate by smiles). 2. Analog stream: rows where synthesizable_decision=NO and at least one analog_smiles_* is non-empty → unpivot (one row per non-empty analog: smiles=analog_smiles_N) → deduplicate against env/property_pool.csv and env/synthesizable_pool.csv → write to evaluator/analogs_to_evaluate.csv → delegate to Evaluator → filter analogs passing ALL user constraints → set source="Analog" → append to env/synthesizable_pool.csv (deduplicate by smiles). After both streams complete: clear env/property_pool.csv (delete or empty it). If synthesizable_pool still has fewer molecules than needed, go back to HI to generate more molecules, rebuild env/property_pool.csv, and repeat from the trigger check. FINAL.csv must be written from env/synthesizable_pool.csv (first target_count rows). Include the `source` column alongside all ADMET/docking columns. Deduplicate by smiles. 12.2.4 Evaluator Evaluator Profile Critically reviews molecules. This is the gate. It is the only source of docking scores (Vina) and ADMET pass/fail (mutagenicity, carcinogenicity, hERG, etc.). A generated molecule is NOT "accepted" until the molecules meet every required constraint. Send every new batch here before counting it toward the goal. # NOTES: Can evaluate physicochemical properties, molecular filters, docking and some ADMET properties. RDKit: QED, SAScore, MW, LogP, TPSA, etc. Molecular filters: PAINS, Brenk, Lipinski Ro5, Lilly MedChem Rules. Molecular docking. ADMET: hERG, DILI, ClinTox, Mutagenicity, Carcinogens. Only works with files, not with raw SMILES. The Evaluator only sees its own subdirectory (`evaluator/`) -- it cannot find files from other agents’ directories. Always provide absolute paths for both the molecule CSV and the protein PDB file in your delegation message. For example: "evaluate /abs/path/to/env/evaluator/hits.csv using /abs/path/to/EGFR.pdb". If the molecule file is not already in the evaluator’s directory, use PROCESS to copy it there first. Output CSV column names (use these exactly when filtering in PROCESS): - Physicochemical: physchem_QED, physchem_SAScore, physchem_MolecularWeight, physchem_LogP, physchem_TPSA, physchem_HBD, physchem_HBA - Filters (boolean, True=pass): filter_PAINS, filter_Brenk, filter_RDKIT, filter_LILLY - ADMET (float 0-1, lower=safer): admet_hERG, admet_DILI, admet_ClinTox, admet_Mutagenicity, admet_Carcinogens - Docking: docking_kcal_mol (float, more negative=better) 12.3 LeadOptimizer LeadOptimizer ’s Records For single-property optimizer family, LeadOptimizer LeadOptimizer is equipped with 11 optimizers of the MolOpt (Gao et al., 2022) oracle-guided molecular optimization suite. In order to understand the performance and utility of each optimizer, we generate performance records using a set of molecules from the ZINC dataset (Sterling and Irwin, 2015) distributed with the suite. We tested with representative single objective lead-optimization tasks, including increasing QED and lowering toxicity-related scores. For each run, we stored the requested property objective, optimizer used, number of molecules meeting the request, runtime, and related metadata. These results are summarized into a structured JSON record that LeadOptimizer LeadOptimizer uses as one source of evidence for optimizer selection during optimization requests. 13 Additional Analysis Table A4: Average property values over output molecules from all targets Method Vina ↓ QED ↑ SAS ↓ Mut ↓ CTx ↓ hERG ↓ Div ↑ Avg. Rank ↓ -6.530 0.360 5.350 0.510 0.710 0.960 – – Agents for Drug Discovery MADDIE MADDIE -7.897 0.646 3.563 0.217 0.224 0.535 0.864 6.3 LIDDIA LIDDIA -7.780 0.712 2.515 0.226 0.108 0.358 0.811 4.1 Prompt-to-Pill -6.834 0.570 3.378 0.541 0.221 0.356 0.739 8.3 AgentD -8.223 0.702 2.861 0.221 0.217 0.542 0.693 6.1 Generative Models for Drug Design Pocket2Mol -7.980 0.625 3.072 0.255 0.219 0.432 0.835 6.0 TargetDiff -7.796 0.605 3.568 0.220 0.222 0.453 0.866 6.4 conDitar -8.094 0.637 3.311 0.249 0.238 0.555 0.865 6.7 DiffSMol -7.609 0.757 3.268 0.278 0.164 0.327 0.869 4.7 ELILLM -8.247 0.592 3.284 0.164 0.129 0.501 0.698 5.3 Generalist LLMs Nemotron-3-Ultra -8.001 0.674 2.263 0.249 0.306 0.686 0.538 7.1 Qwen3-30B -7.756 0.628 1.687 0.230 0.091 0.489 0.345 5.9 Reference Ligands -8.053 0.472 3.241 0.334 0.339 0.608 – – US FDA Approved Drugs -7.828 0.600 2.961 0.261 0.368 0.638 0.803 – • ↓ / ↑ : lower / higher is better. Metric values are averaged over the molecules each method outputs; Avg. Rank is the mean rank across the seven metrics (lower is better). Bold values indicate the best performance. 13.1 Property Value Analysis We also report the average property values over the output molecules in Table A4. We calculate an additional metric, Div , as the average diversity of molecules per target. We measure diversity similar to LIDDIA (Averly et al., 2025), as follows: Diversity(ℳ)=1−i,j⊆ℳ[sim(i,j)], Diversity(M)=1- E _ \ $ m $_i, $ m $_j\ [ sim_ T( $ m $_i, $ m $_j) ], where ℳM is a set of molecules, i $ m $_i and j $ m $_j as two distinct molecules in ℳM, and sim(i,j) sim_ T( $ m $_i, $ m $_j) as the Tanimoto similarity of i $ m $_i and j $ m $_j’s Morgan’s fingerprints (Morgan, 1965). MADDIE MADDIE performs competitively with existing baselines. This is reflected in its average rank (6.3), which is comparable to that of the strong baselines (e.g., 6.1 for AgentD, 6.0 for Pocket2Mol, 6.4 for TargetDiff and 6.7 for conDitar). LIDDIA achieves the best average rank on property values, although, as discussed in Section 4.1, relatively few molecules have these properties. Notably, MADDIE MADDIE also produces structurally diverse molecules even when diversity is not provided as an explicit requirement for MADDIE MADDIE to achieve. We attribute this to HitIdentifier HitIdentifier : since it produces a sufficiently diverse pool of molecules initially, optimizations and synthesis analog generations are performed over an already diverse set, and other agents manage to preserve the diversity in the final set of molecules. No method consistently outperforms the others, which is unsurprising since the task is to generate molecules that jointly satisfy all requirements rather than to optimize molecules beyond requirements. MADDIE MADDIE generates molecules whose in silico properties surpass or match with reference ligands and FDA-approved drugs. Notably, molecules generated by MADDIE MADDIE outperform both reference ligands and drugs on QED (0.646) and all the toxicity properties— Mut (0.217), CTx (0.224) and hERG (0.535). They remain on par with approved drugs on binding affinity and synthetic accessibility, with near-identical Vina (-7.897 vs -7.828) and reasonable value for SAS (3.563 vs 2.961). (a) ABO_BGAT (b) HTR2A Figure A1: Transition trajectories of MADDIE MADDIE on two representative targets. 13.2 Additional Representative Trajectories Discussion We illustrate two more trajectories by MADDIE MADDIE in Figure A1. (1) In ABO_BGAT, MADDIE MADDIE generates no synthesizable analogs and instead relies solely on HitIdentifier HitIdentifier and LeadOptimizer LeadOptimizer for generations. We hypothesize it is likely that the analogs cannot be found for these molecules, and instead MADDIE MADDIE falls back to the other options for generations. (2) HTR2A represents a failure case in which Orchestrator Orchestrator cannot locate the artifacts generated by HitIdentifier HitIdentifier . MADDIE MADDIE invokes Evaluator Evaluator repeatedly (more than 10 times) before deciding to re-invoke HitIdentifier HitIdentifier again, at which point the campaign resumes and advances. This illustrates how MADDIE MADDIE can detect when the campaign stalls and recover from it, rather than remaining in the loop. 14 Additional Case Studies Figure A2: MADDIE MADDIE ’s generated molecules on PSEN1 exceeding either the reference ligand or Nirogacestat on all in silico properties defined in Section 3. (a) ACHE-01 (b) ACHE reference ligand (c) Galantamine Figure A3: ACHE case study (a) BCHE-01 (b) BCHE reference ligand (c) Rivastigmine Figure A4: BCHE case study We provide additional case studies for ACHE and BCHE in Figure A3 and A4, respectively. (1) MADDIE MADDIE generates ACHE-01 (Figure 3(a)), a molecule with in silico properties comparable ( CTx ) or better than Galantamine (Figure 3(c)) (Lilienfeld, 2002) on all properties. ACHE-01 also exceeds the reference ligand (Figure 3(b)) except on Vina (-9.13 vs -11.46), emphasizing its potential. (2) For the second case, BCHE-01 (Figure 4(a)) exceed those of reference ligand (Figure 4(b)) and Rivastigmine (Birks et al., 2015) (Figure 4(c)) in binding affinity (-8.90 vs -8.76 vs -5.66). More interestingly, BCHE-01 passes the Lly filters that Rivastigmine fails, suggesting a viable non-covalent ligand alternative for BCHE.