Paper deep dive

Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling

Wengan He, Yongsheng Luo, Lihong Jiang, Wenhui Xu, Yu Li

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 83

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%

Last extracted: 8/23/2026, 1:24:07 AM

Summary

This review paper outlines the methodological evolution of protein structure prediction, categorizing it into four phases: explicit evolutionary modeling, end-to-end deep folding, unified heterogeneous modeling, and generative design. It highlights three key transitions: from explicit co-evolutionary features to learned sequence representations (e.g., AlphaFold2, ESMFold); from monomer folding to complex and heterogeneous system modeling (e.g., AlphaFold-Multimer, AlphaFold3); and from prediction to generative design (e.g., RFdiffusion). The paper emphasizes how deep learning has shifted the field from modular, MSA-dependent pipelines to integrated, end-to-end architectures capable of handling diverse biomolecular assemblies.

Entities (12)

Protein Structure Prediction · field · 99%AlphaFold2 · model · 98%Deep Learning · technology · 98%ESMFold · model · 98%RoseTTAFold · model · 98%AlphaFold3 · model · 97%AlphaFold-Multimer · model · 97%RFDiffusion · model · 97%RoseTTAFoldNA · model · 97%Protein Data Bank · database · 96%Multiple Sequence Alignment · method · 95%Diffusion Models · method · 94%

Relation Signals (8)

AlphaFold-Multimer → models → Protein Complexes

confidence 96% · integrated modeling of heterogeneous molecular systems in AlphaFold-Multimer

AlphaFold3 → models → Heterogeneous Molecular Systems

confidence 96% · integrated modeling of heterogeneous molecular systems in ... AlphaFold3

RFDiffusion → enables → Protein Design

confidence 95% · design-oriented generative modeling in RFdiffusion

Protein Structure Prediction → evolvedfrom → Explicit Evolutionary Modeling

confidence 95% · field is organized into four methodological phases ... from explicit evolutionary coupling features

Protein Structure Prediction → evolvedto → Generative Modeling

confidence 95% · from prediction-oriented structure inference to design-oriented generative modeling

ESMFold → uses → Learned Sequence Representations

confidence 95% · learned sequence representations in ... ESMFold

AlphaFold2 → uses → Learned Sequence Representations

confidence 95% · learned sequence representations in AlphaFold2

Multiple Sequence Alignment → isreplacedby → Learned Sequence Representations

confidence 90% · from explicit evolutionary coupling features ... to learned sequence representations

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Accurate protein structure prediction is fundamental to structural biology because protein structure underlies molecular function and provides a basis for mechanistic interpretation. Recent advances in deep learning have transformed the field from multiple sequence alignment (MSA)-driven monomer folding into broader frameworks capable of modeling protein complexes and increasingly heterogeneous molecular systems. Existing reviews have summarized this progress from the perspectives of representative models, application domains, and protein design. Building on these efforts, this review focuses on the methodological evolution of the field itself. It examines recent developments through three closely related dimensions: representations and data, architectures and learning strategies, and confidence and evaluation. Within this perspective, the field is organized into four methodological phases and three cross-cutting transitions: from explicit evolutionary coupling features and early contact prediction to learned sequence representations in AlphaFold2, RoseTTAFold, and ESMFold; from protein-only monomer folding to increasingly integrated modeling of heterogeneous molecular systems in AlphaFold-Multimer, RoseTTAFoldNA, and AlphaFold3; and, more recently, from prediction-oriented structure inference to design-oriented generative modeling in RFdiffusion and related frameworks. This framework provides a clearer understanding of how methodological shifts have shaped the capabilities, limitations, and practical roles of recent models.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

79,719 characters extracted from source content.

Expand or collapse full text

Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling Wengan He a,b,∗,1 , Yongsheng Luo a,1 , Lihong Jiang b , Wenhui Xu b and Yu Li a,∗ a Zhuhai College of Science and Technology, Zhuhai, China b Zhuhai UNO Technology Co., Ltd., Zhuhai, China A R T I C L E I N F O Keywords: protein structure prediction structural bioinformatics deep learning diffusion models protein design A B S T R A C T Accurate protein structure prediction is fundamental to structural biology because protein structure underlies molecular function and provides a basis for mechanistic interpretation. Recent advances in deep learning have transformed the field from multiple sequence alignment (MSA)-driven monomer folding into broader frameworks capable of modeling protein complexes and increasingly heteroge- neous molecular systems. Existing reviews have summarized this progress from the perspectives of representative models, application domains, and protein design. Building on these efforts, this review focuses on the methodological evolution of the field itself. It examines recent developments through three closely related dimensions: representations and data, architectures and learning strategies, and confidence and evaluation. Within this perspective, the field is organized into four methodological phases and three cross-cutting transitions: from explicit evolutionary coupling features and early contact prediction to learned sequence representations in AlphaFold2, RoseTTAFold, and ESMFold; from protein-only monomer folding to increasingly integrated modeling of heterogeneous molecu- lar systems in AlphaFold-Multimer, RoseTTAFoldNA, and AlphaFold3; and, more recently, from prediction-oriented structure inference to design-oriented generative modeling in RFdiffusion and related frameworks. This framework provides a clearer understanding of how methodological shifts have shaped the capabilities, limitations, and practical roles of recent models. 1. Introduction Protein structure is central to structural biology because it constrains molecular function and enables mechanistic interpretation [6, 18]. The Protein Data Bank provides the principal global archive of experimentally determined macromolecular structures [17]. Traditional computational approaches, including homology modeling, fragment as- sembly, and physics-based refinement, achieved important progress but remained constrained by template availability, conformational sampling complexity, and the accuracy of energy functions [95, 70, 18]. The emergence of deep learning fundamentally rede- fined the methodological foundations of structure prediction [66, 59]. Within a short period, the field progressed from co-evolution-assisted contact inference [82] to end-to-end coordinate prediction [7, 47] with near-experimental accu- racy for many monomeric proteins [40]. Subsequent devel- opments expanded modeling capabilities to protein–protein complexes [23], protein–nucleic acid assemblies [8], and ligand-inclusive heterogeneous molecular systems [42, 2, 50, 75, 85]. More recently, diffusion- and flow-based gen- erative frameworks [83, 27] extended structural modeling toward controllable protein design. Existing reviews have summarized this progress from perspectives such as representative model families, appli- cation scope, and protein design [59, 63]. The present re- view instead emphasizes the methodological evolution of protein structure prediction itself. Specifically, it examines ∗ Corresponding author. wenganhe@icloud.com (W. He); jluzhliyu@zcst.edu.cn (Y. Li) ORCID(s): 1 These authors contributed equally to this work. recent developments through three closely related dimen- sions: (1) representations and data, (2) architectures and learning strategies, and (3) confidence and evaluation. As summarized in Figure 1, the field is organized here into four methodological phases spanning explicit evolutionary mod- eling, end-to-end deep folding, relatively unified complex structure modeling, and generative structure modeling and design. Figure 1 further highlights three associated transi- tions: from explicit evolutionary coupling features and early contact prediction [82] to learned sequence representations in models such as AlphaFold2 [40], RoseTTAFold [7], and ESMFold [47]; from monomer-focused modeling in the first two phases to increasingly integrated treatment of hetero- geneous molecular systems in AlphaFold-Multimer [23], RoseTTAFoldNA [8], RoseTTAFold All-Atom [42], and Al- phaFold3 [2]; and, more recently, from prediction-oriented structure inference to design-oriented generative modeling in RFdiffusion [83] and related frameworks [27]. Rather than replacing predictive approaches, generative protein design is treated here as a complementary paradigm that extends structural modeling toward controllable creation and explo- ration of novel sequence–structure solutions. This phase- and-transition perspective provides a clearer understanding of how methodological shifts have shaped the capabilities, limitations, and practical roles of recent models. Phases in this review are assigned primarily according to modeling objective and system scope rather than architec- ture alone. Thus, AlphaFold3 is placed in Phase I because its principal task is structure inference for specified hetero- geneous biomolecular assemblies, even though its structure module uses diffusion-based denoising [2]. He et al.: Preprint submitted to ElsevierPage 1 of 15 Protein Structure Prediction Figure 1: Four methodological phases in the evolution of deep learning for protein structure prediction. 2. Explicit Evolutionary Modeling The earliest methodological phase of protein structure prediction with machine learning is characterized by the ex- plicit modeling of evolutionary couplings derived from mul- tiple sequence alignments (MSAs). As illustrated in Figure 2, this paradigm follows a structured and modular pipeline consisting of (A) MSA construction, (B) co-evolutionary feature extraction, (C) contact or distance prediction, and (D) downstream structure reconstruction . The underlying data regime is dominated by curated structural datasets such as the Protein Data Bank (PDB) [17], together with homologous sequence collections obtained through large- scale alignment tools, including ClustalW [45], HMMER [26, 20], HHblits [62], and MMseqs2 [74]. At the core of this phase lies a key assumption: evolu- tionary constraints encoded in homologous sequences re- flect three-dimensional structural relationships. According to the thermodynamic hypothesis of protein folding [6] and subsequent theoretical developments [18, 12], amino acid sequences determine native structures, and residues that are spatially proximal tend to co-evolve to maintain structural stability and function. Statistical approaches such as direct coupling analysis (DCA) quantify these dependencies by estimating residue–residue coupling strengths from MSAs [53, 54, 31]. These coupling signals provide an indirect but powerful proxy for structural proximity and serve as the primary source of structural information in this paradigm. 2.1. From Statistical Couplings to Deep Learning-Based Contact Prediction Building on these statistical foundations, deep learning models introduced more expressive approaches to con- tact prediction. Convolutional neural networks and resid- ual architectures trained on MSA-derived features were used to predict residue–residue contact maps or distance distributions [82]. In these models, co-evolutionary fea- tures—such as covariance matrices, coupling scores, and position-specific profiles—are explicitly computed and used as input representations. Representative systems such as RaptorX and related frameworks substantially improved long-range contact pre- diction compared with purely statistical methods [82, 59]. However, despite these improvements, the overall method- ological logic remained largely unchanged. Evolutionary signals were still engineered first, predictive constraints were generated second, and full three-dimensional structures were reconstructed only in a downstream stage. In this sense, deep learning enhances predictive accuracy without fundamen- tally altering the pipeline structure. 2.2. Modular Pipeline and Structural Decoupling As shown in Figure 2, this phase exhibits a distinctly modular architecture. Contact or distance prediction is per- formed independently of coordinate reconstruction, and pre- dicted constraints are subsequently passed to external fold- ing or reconstruction procedures based on fragment assem- bly, distance geometry, or knowledge-based potentials [70, 72]. Although some implementations more tightly couple He et al.: Preprint submitted to ElsevierPage 2 of 15 Protein Structure Prediction Figure 2: Co-evolutionary modeling pipeline in early protein structure prediction. The schematic illustrates (A) MSA construction from sequence databases, (B) extraction of co-evolutionary features (e.g., covariance or coupling matrices), (C) deep learning–based contact or distance prediction, and (D) downstream structure reconstruction using external folding engines. constraint prediction and reconstruction, the dominant early pipeline remained fragmented: gradients did not propagate from final structural outputs through all preceding stages, and individual components were optimized separately. This design introduces several limitations. Errors in MSA construction, coupling estimation, or contact predic- tion can propagate into downstream folding without end- to-end correction. Moreover, global geometric consistency is not directly optimized at the contact-prediction stage, so locally plausible constraints may still be incompatible with a physically coherent three-dimensional structure. Conse- quently, gains in intermediate contact or distance accuracy do not necessarily yield proportional improvements in final structural quality [44, 5]. 2.3. Methodological Characteristics and System-Level Limitations These characteristics can be summarized from a system- level perspective, as shown in Table 1. The phase com- bines curated structural supervision with MSA-dependent homologous sequence information, relies on explicit co- evolutionary features as its principal representation, and formulates prediction in terms of contacts or distances rather than direct coordinate generation. From a representation standpoint, the strengths and weaknesses of this paradigm are tightly coupled. MSAs [45] encode evolutionary variation through covariance statistics and coupling scores [53, 54, 31], which can provide highly informative signals for structurally conserved protein fam- ilies. However, predictive performance depends strongly on the depth, diversity, and alignment quality of available homologs [59]; when coverage is limited, co-evolutionary signals become sparse or noisy [82]. This limitation is further reinforced by the data regime. Although sequence databases have expanded rapidly, usable target-specific information remains constrained by the abil- ity to identify and align informative homologs using tools such as HMMER [26, 20], HHblits [62], and MMseqs2 [74]. Proteins with shallow or biased evolutionary coverage there- fore remain challenging even when structural supervision is available from repositories such as the PDB [17]. The learning objective introduces an additional limita- tion. Contact maps and distance matrices are intermediate structural representations rather than final atomic coordi- nates, and the inverse mapping from a set of predicted con- straints to a unique, physically valid three-dimensional struc- ture is generally underdetermined. Reconstruction quality therefore depends on the downstream optimization proce- dure as well as on the accuracy and consistency of the predicted restraints [46, 78]. 2.4. Transition Toward End-to-End Modeling Despite these limitations, explicit evolutionary model- ing represents a critical foundation for subsequent develop- ments. It established the importance of pairwise residue rela- tionships, demonstrated that sequence variation can be sys- tematically exploited for structural inference, and provided the conceptual basis for learning structure from sequence- derived constraints. More importantly, the limitations of this phase—particularly its dependence on explicit feature engineering and its modu- lar separation between prediction and reconstruction—directly motivate the transition to end-to-end deep folding models. Later approaches seek to integrate representation learning, geometric reasoning, and coordinate prediction within a unified optimization framework, thereby addressing the He et al.: Preprint submitted to ElsevierPage 3 of 15 Protein Structure Prediction Table 1 Methodological characteristics and limitations of explicit evolutionary modeling in early protein structure prediction. AspectDescriptionLimitation Data regimeCurated structures + homologous sequences (MSA- dependent) Limited by MSA depth and coverage Representation Explicit co-evolutionary features (covariance, cou- pling) Requires high-quality alignments ObjectiveContact/distance predictionIndirect relation to final 3D structure PipelineModular (prediction + folding)Error propagation, no end-to-end optimization fragmentation and inefficiencies inherent in early pipelines [66, 40]. 3. From Explicit Evolutionary Features to Learned Sequence Representations A central methodological transition in protein structure prediction is the shift from explicit evolutionary feature en- gineering toward learned sequence representations (Figure 1). Rather than relying exclusively on query-specific co- evolutionary statistics derived from MSAs, some modern predictors learn latent representations from large-scale se- quence corpora through self-supervised pretraining [32, 22, 47] 3.1. MSA-Based Pipelines and Their Underlying Assumptions In earlier MSA-based pipelines, structural information is obtained through a multi-stage process (Figure 3). Starting from a query sequence, homologous sequences are retrieved from large databases to construct an MSA, which serves as the basis for extracting co-evolutionary features such as covariance matrices and coupling scores. These features are then used as inputs to neural networks for predicting residue–residue contacts or distance distributions, which are subsequently converted into three-dimensional structures through external folding engines or constraint-based recon- struction. This pipeline reflects the core assumption of explicit evolutionary modeling: structural constraints can be inferred from correlated mutations across homologous sequences. When sufficiently deep and diverse MSAs are available, such co-evolutionary signals provide strong indicators of spa- tial proximity and functional constraint [53, 54, 31]. Their reliability nevertheless varies with MSA depth, sequence diversity, and alignment quality. Query-specific MSA con- struction also introduces database-search and preprocessing costs, particularly at proteome scale [74, 56]. 3.2. Emergence of Learned Sequence Representations In explicitly MSA-free PLM-based predictors such as ESMFold and OmegaFold, the query sequence is encoded directly by a pretrained protein language model rather than being accompanied by a query-specific MSA. Self-supervised pretraining on large sequence corpora produces contextual embeddings that can encode information relevant to struc- ture and function [47, 87]. The precise structural content and transferability of these embeddings remain model- and task-dependent. These learned embeddings are then passed to down- stream folding or coordinate-prediction modules. In MSA- free implementations, removing query-specific homology search can simplify inference and improve throughput, al- though the magnitude of the speed and accuracy trade-off varies across models, target classes, and evaluation settings [47, 87]. This paradigm also shifts much of the computa- tional burden from online alignment construction to offline large-scale pretraining. 3.3. Pipeline-Level Comparison Between MSA and PLM Paradigms The contrast between these two paradigms is illustrated in Figure 3. In MSA-based pipelines (Figure 3), structural information is derived from explicitly computed evolution- ary features, and the workflow is multi-stage and alignment- dependent. In PLM-based pipelines (Figure 3), structural information is encoded in learned sequence representations, and prediction becomes more integrated and less dependent on external preprocessing. This transition reflects a broader methodological shift: from explicit feature engineering to representation learning. In the former, structural signals are manually derived from sequence variation; in the latter, they are implicitly captured within learned embedding spaces. As a result, the locus of structural information moves from external preprocessing steps to the internal representations of neural networks. 3.4. Methodological Differences and Scaling Behavior From a methodological perspective, the differences be- tween these paradigms can be summarized along several di- mensions, as shown in Table 2. MSA-based methods obtain target-specific evolutionary information from homologous sequences, whereas PLM-based methods encode statistical regularities acquired during pretraining. MSA-dependent performance is therefore strongly influenced by homolog availability, while the capabilities of PLM-based systems depend on the architecture, pretraining data, model scale, and downstream folding module. These are tendencies rather He et al.: Preprint submitted to ElsevierPage 4 of 15 Protein Structure Prediction Figure 3: Comparison between MSA-based and PLM-based structure prediction pipelines. The schematic contrast (A) traditional pipelines that rely on MSA construction and explicit co-evolutionary feature extraction with (B) PLM-based pipelines that directly encode sequence information into contextual embeddings. Key differences include information source, computational cost, and integration with downstream structure modules. Table 2 Comparison between explicit evolutionary modeling and learned sequence representations DimensionExplicit evolutionary modelingLearned sequence representations Data regimeMSA-dependent homologous sequencesLarge-scale unlabeled sequence corpora Representation Explicit co-evolutionary features (covariance, cou- pling) Implicit contextual embeddings (PLMs) InformationDirectly computed evolutionary couplingsLearned statistical patterns from sequences ScalabilityLimited by MSA depth and coverageCan improve with pretraining data and model scale, depending on architecture Generalization Strong when informative homologous coverage is available Some MSA-free PLM models can be advantageous for orphan or low-homology targets than universal properties of every model in either category [15, 24, 47]. These differences lead to distinct but overlapping gen- eralization profiles. MSA-based approaches remain highly effective when informative homologous coverage is avail- able. Some MSA-free PLM predictors, notably ESMFold and OmegaFold, can be useful for orphan or low-homology targets because they do not require a target-specific align- ment, but this advantage should not be generalized to all PLM architectures or all difficult proteins [47, 87]. The two information sources are therefore better viewed as comple- mentary than mutually exclusive. 3.5. Limitations and Hybrid Strategies PLM-based approaches introduce their own challenges. Learned embeddings are distributed representations whose biological and geometric content is often less directly in- terpretable than explicit coupling matrices. In addition, the apparent reduction in inference-time preprocessing can ob- scure the substantial computational and data requirements of PLM pretraining [21, 16]. Whether learned sequence representations consistently recover the fine-grained residue-pair constraints supplied by deep MSAs remains model- and target-dependent. Hybrid systems that combine PLM embeddings with MSA-derived or pairwise geometric information may therefore be advan- tageous in some settings [15]. He et al.: Preprint submitted to ElsevierPage 5 of 15 Protein Structure Prediction Figure 4: Evolution from monomer folding to relatively unified heterogeneous modeling. The schematic illustrates (A) monomer folding with residue-level representation, (B) multi-chain protein-complex modeling, (C) heterogeneous biomolecular systems including nucleic acids and ligands, and (D) mixed token- and atom-level modeling with iterative or diffusion-based refinement. 3.6. Transition Toward End-to-End and Unified Modeling Importantly, this transition should not be interpreted as a simple replacement of MSA-based methods. Rather, it represents a broader shift in how structural information is conceptualized and learned. Evolutionary signals are not dis- carded but transformed—from explicitly computed statistics to latent representations embedded within model parame- ters. More broadly, the move toward learned sequence repre- sentations lays the foundation for subsequent methodologi- cal developments. By reducing dependence on handcrafted features and enabling more integrated architectures, it facili- tates the emergence of end-to-end folding models and, later, unified frameworks for modeling heterogeneous molecular systems. In this sense, this transition serves as a critical bridge between early co-evolution-based approaches and modern deep learning architectures. 4. From Monomer Folding to Unified Modeling of Heterogeneous Molecular Assemblies A major development in contemporary protein structure prediction is the expansion of modeling scope from single- chain folding to the joint treatment of increasingly hetero- geneous molecular systems [23, 8]. This transition involves more than an increase in target complexity: it changes the entities represented by the model [42], the geometric and chemical relationships that must be learned, and the criteria by which predictions are evaluated [2]. As illustrated in Figure 4, the evolution of protein structure prediction can be understood as a progressive expansion of both system complexity and representational richness, moving from iso- lated protein folding toward relatively unified modeling of heterogeneous molecular interactions. 4.1. Monomer Folding as the Foundational Paradigm In the monomer-folding setting, the predictive task can be formulated as mapping a single amino acid sequence to a dominant three-dimensional structure. Such predic- tions are often assessed using global and local structural- similarity metrics, including TM-score, RMSD, and lDDT; these metrics are not exclusive to monomer prediction, al- though complex and ligand-containing systems require ad- ditional interface- and chemistry-aware measures [92, 52, 10]. This formulation underlies landmark end-to-end models such as AlphaFold2 [40]. Monomer folding nevertheless captures only a subset of biologically relevant structural phenomena. Protein func- tion is frequently determined by binding, multimerization, nucleic-acid recognition, ligand association, and higher- order assembly, motivating models that represent inter- molecular interfaces and chemically distinct entities [4, 64]. 4.2. From Multi-Chain Complexes to Heterogeneous Systems The first step beyond monomer folding is the exten- sion to multi-chain protein complexes, as exemplified by AlphaFold-Multimer [23]. In this setting, the model must He et al.: Preprint submitted to ElsevierPage 6 of 15 Protein Structure Prediction reason about intra-chain folding and inter-chain geome- try, including interface arrangement, chain identity, and stoichiometry. Accurate individual-chain structures do not guarantee a correct assembly, so complex prediction re- quires interface-aware representations, training objectives, and evaluation criteria [93, 13]. The subsequent expansion proceeds through distinct steps. RoseTTAFoldNA extends the three-track framework specifically to protein–DNA and protein–RNA complexes [8], whereas RoseTTAFold All-Atom broadens the mod- eling space to generalized biomolecular systems that can include ligands and other non-polymer entities [42]. Al- phaFold3 further supports joint structure inference for pro- teins, nucleic acids, ligands, ions, and modified residues within a common predictive system [2]. These models should therefore not be treated as having identical molecular scope. 4.3. Toward Relatively Unified Modeling Frameworks As these capabilities converge, the field moves toward what may be described as relatively unified modeling frame- works. The qualifier “relatively” is important: current sys- tems integrate several previously separate tasks within shared architectures [42, 2], but their accuracy, training coverage, and chemical treatment remain uneven across molecular classes and interaction types [36, 57]. This transition entails a corresponding change in rep- resentation. Earlier models emphasized residue-level se- quence, pair, and frame representations, whereas heteroge- neous systems require mixed token- and atom-level descrip- tions that encode chemical identity, connectivity, and cross- entity geometry. The predictive objective correspondingly broadens from isolated-chain folding to joint structure in- ference across interacting molecular components [42, 2]. 4.4. AlphaFold3 and the Redefinition of the Prediction Problem Within this phase, AlphaFold3 represents a key method- ological turning point [2]. Its significance lies not only in predictive performance, but also in reformulating structure inference around heterogeneous biomolecular interactions. AlphaFold3 expands the modeled entity space by sup- porting proteins, nucleic acids, ligands, ions, and modi- fied residues within a shared predictive framework. This is methodologically important because many biological in- teractions depend on precise geometry between chemically distinct components rather than on protein folds considered in isolation [2]. Its representation and output formulation also move to- ward joint atom-level modeling of heterogeneous systems, enabling the network to reason about local geometry, chem- ical identity, and interaction-specific contacts within a com- mon structure-prediction pipeline [2]. The term “all-atom” should not be interpreted as implying perfect stereochemical validity in every prediction. AlphaFold3 replaces the AlphaFold2 structure module with a diffusion module that denoises atom coordinates conditioned on processed input representations. Repeated random seeds can yield multiple candidate predictions, but the model remains prediction-oriented because the identi- ties of the molecular entities are specified and the primary objective is to infer their joint structure. The presence of a diffusion architecture therefore does not, by itself, place a model in the design-oriented generative phase [2]. AlphaFold3 also increases the importance of evaluating local chemical validity. The diffusion formulation accommo- dates diverse chemical components and reduces reliance on molecule-specific output modules, yet residual stereochemi- cal violations and clashes can still occur. Structural accuracy should therefore be assessed using both global or interface- level similarity and local chemistry-aware validation [2, 14]. 4.5. Methodological Implications and Remaining Challenges The transition from monomer folding to heterogeneous modeling introduces systematic changes across multiple as- pects of the problem. These changes extend beyond repre- sentation and encompass the definition of prediction targets, the nature of available data, and the criteria used for evalua- tion. To clarify these differences, a structured comparison between monomer folding and heterogeneous molecular sys- tem modeling is provided in Table 3. These changes introduce several challenges. High-quality training and benchmark data are unevenly distributed across complex types; representations must balance chemical detail with computational tractability; and evaluation must dis- tinguish global assembly accuracy from interface fidelity, ligand pose quality, and local stereochemical validity [51, 19, 89]. Furthermore, broader modeling scope does not eliminate the value of specialized inductive biases. Antibody model- ing, particularly in highly variable CDR loops, remains a prominent example in which domain-specific systems such as IgFold can offer practical advantages [1, 65]. 4.6. Transition Toward Generative Modeling The progression summarized in Figure 4 also highlights a broader methodological trajectory. As models become capable of representing heterogeneous molecular entities within shared all-atom spaces and performing iterative re- finement, the conceptual boundary between prediction and generation begins to diminish. In this sense, the expansion toward heterogeneous mod- eling serves as a bridge to subsequent developments in generative structural modeling. By redefining the prediction target from isolated protein folds to interaction-rich molec- ular systems, this phase lays the foundation for approaches that treat structure not as a single deterministic outcome, but as part of a broader distribution of plausible configurations. He et al.: Preprint submitted to ElsevierPage 7 of 15 Protein Structure Prediction Table 3 Comparison between monomer folding and heterogeneous molecular system modeling. The table contrasts the two phases in terms of system scope, representation, objective, data regime, and evaluation, highlighting the movement from single-chain residue-level prediction toward interaction-aware modeling of multi-molecular assemblies. DimensionMonomer folding (Phase I)Heterogeneous modeling (Phase I) System scope Single protein chainMulti-chain and multimolecular assemblies Representation Residue-level sequence–pair and frame representa- tions Mixed token and atom-level heterogeneous represen- tations ObjectiveStructure inference for a specified single-chain pro- tein Joint structure inference for specified multimolecular assemblies Data regimeRelatively richer monomer-focused structural dataMore heterogeneous and unevenly sampled complex data EvaluationGlobal and local structural-similarity metricsGlobal, interface-specific, and chemistry-aware met- rics 5. From Prediction-Oriented Structure Inference to Generative Protein Design As illustrated in the revised Figure 4, the key distinction is not whether a model produces one or multiple samples, nor whether it uses a diffusion module. It is whether the primary objective is inference for specified molecular entities or generation of novel design candidates under constraints. 5.1. Prediction-Oriented Structure Inference Prediction-oriented structure models estimate confor- mations for specified molecular inputs. AlphaFold2 largely behaves as a point-estimation system that returns a ranked structural hypothesis with confidence metrics [40]. Al- phaFold3, by contrast, uses diffusion-based coordinate de- noising and can generate different candidate predictions across seeds, while still remaining prediction-oriented be- cause the sequence and molecular composition are given [2]. Confidence measures such as pLDDT [77] and predicted aligned error [2] quantify expected reliability of predicted re- gions or relative placements, but they should not be equated with a complete conformational ensemble [3, 84]. Likewise, variability across random seeds may reflect model sampling behavior [2] without necessarily reproducing the biologi- cally populated distribution of states [84]. Prediction-oriented inference is highly effective when the scientific question concerns the likely structure of a specified target under conditions represented by the train- ing data. Its limitations become more apparent when the target occupies multiple biologically relevant states or when context-dependent dynamics are central to function. 5.2. Structural Heterogeneity and Distributional Prediction Many proteins and complexes occupy conformational ensembles shaped by thermodynamic fluctuations, molecu- lar partners, and environmental conditions. A single high- confidence structural hypothesis may therefore be insuf- ficient for intrinsically disordered regions, fold-switching proteins, flexible loops, or ligand-induced rearrangements [11, 58, 76]. This problem motivates a distinct line of work on dis- tributional structure prediction, in which a model samples alternative conformations for a specified sequence. Such approaches should be distinguished from de novo protein design: both can use generative machinery, but one aims to approximate a target’s conformational distribution, whereas the other aims to create new molecular solutions. EigenFold is one representative example of sequence-conditioned dis- tributional structure prediction [39]. Accordingly, neither point prediction nor repeated stochas- tic sampling should automatically be interpreted as a faithful physical ensemble. Validation against experimental ensem- ble data, molecular simulations, or state-specific bench- marks remains necessary [37, 68]. 5.3. Generative Modeling for Protein Design Generative protein-design models are defined primarily by their ability to create novel molecular candidates rather than merely by their ability to sample multiple structures. Depending on the task, they may generate protein back- bones, amino acid sequences, side-chain arrangements, or joint sequence–structure representations under design con- straints. Diffusion models [30] have become an influential de- sign paradigm because conditional denoising can be guided by geometric motifs, target surfaces, symmetries, or other structural constraints [83]. Flow-matching [48] and autore- gressive approaches [81] provide alternative routes to the same broad objective: learning a generative process over molecular design space. RFdiffusion generates novel protein backbones condi- tioned on structural or functional constraints and has been applied to motif scaffolding, symmetric assemblies, and binder design [83]. Its primary contribution is therefore controllable de novo design rather than conformational sam- pling of a fixed natural sequence. La-Proteina extends this direction through partially la- tent flow matching for atomistic protein generation and joint He et al.: Preprint submitted to ElsevierPage 8 of 15 Protein Structure Prediction Table 4 Comparison between prediction-oriented structure inference and generative protein design. The table distinguishes the paradigms by conditioning, primary objective, output, and modeling mechanism; it deliberately avoids equating prediction with a single output or generation with multiple samples. DimensionPrediction-oriented structure inferenceGenerative protein design ConditioningSpecified sequence and/or molecular com- position Motifs, target geometry, interfaces, func- tion, and optional sequence constraints Primary objectiveInfer structures compatible with specified molecules Createnovelstructuresor sequence–structure solutions OutputOne or more candidate structures for the same specified target Diverse de novo backbones, sequences, or joint designs Inference/generation mechanism Deterministic regression or diffusion-based sampling Diffusion, flow matching, autoregressive, or other generative processes Representative modelsAlphaFold2, AlphaFold3RFdiffusion, La-Proteina treatment of sequence–structure information [27]. RFdiffu- sion and La-Proteina differ in representation and generation strategy, but both are better characterized as design-oriented generative systems than as generic uncertainty models. 5.4. Conceptual Differences Between Prediction and Generation The distinction between prediction-oriented inference and generative design therefore concerns conditioning and objective. Prediction starts from specified molecular identi- ties and asks what structure or interaction geometry is plau- sible. Generative design starts from a desired motif, target interface, geometry, or functional objective and searches for novel molecular realizations. Either paradigm may produce multiple candidates, and either may use diffusion. This distinction has direct implications for evaluation. Predictive models are assessed primarily by correspondence to experimentally determined structures and by confidence calibration, whereas design models must additionally be evaluated for novelty, validity, diversity, designability, and experimental function [43, 49, 90]. To clarify these differences, a structured comparison is provided in Table 4. 5.5. Complementarity and Emerging Unified Perspective Prediction-oriented inference and generative design are complementary. Predictive models can evaluate whether a designed sequence is compatible with a target fold [40, 83] or interaction geometry [2], while generative models can propose candidates that satisfy specified constraints [83]. In practical design loops, prediction, generation, ranking, and experimental validation may therefore be iterated rather than treated as isolated stages [61, 71]. This complementarity contributes to a broader conver- gence between structural inference and protein engineering. Nevertheless, prediction of a specified structure and genera- tion of a novel design remain different tasks, with different notions of success and different validation requirements [83, 27]. 5.6. Challenges and Future Directions Generative protein design introduces several challenges. Generated candidates must satisfy stereochemical, energetic, and folding constraints; computational metrics may not pre- dict expression, stability, binding, or biological function; and diversity must be distinguished from invalid or redundant sampling. Evaluation therefore requires a combination of structural validation, sequence analysis, physical modeling, and experimental testing. [90, 9] Computational cost and sampling efficiency also remain important, especially for iterative diffusion or flow-based systems [34, 91]. Despite these challenges, generative modeling repre- sents a fundamental expansion of computational structural biology. It extends the field from inference about specified molecules toward the controlled creation and exploration of new molecular solutions. This design-oriented phase com- plements, rather than supersedes, predictive structure mod- els. 6. Synthesis, Challenges, and Future Directions 6.1. Synthesis of methodological evolution Protein structure prediction has undergone a profound methodological transformation over the past decades, evolv- ing from physics-based and knowledge-based modeling to deep learning–driven inference and, more recently, to in- creasingly integrated and generative frameworks. As orga- nized in this review, this development is best understood not simply as a sequence of progressively stronger models, but as a broader reconfiguration of how structural information is represented, learned, and utilized. The four methodological phases and three cross-cutting transitions discussed through- out the review provide a structured way to interpret this evolution beyond a purely chronological narrative. At a fundamental level, the history of protein structure prediction reflects a continuing redefinition of what counts as the relevant source of structural information. Early ap- proaches relied on explicit physical principles, statistical potentials, and optimization over energy landscapes [18, 6, He et al.: Preprint submitted to ElsevierPage 9 of 15 Protein Structure Prediction 12, 72]. The subsequent rise of co-evolutionary analysis introduced a new form of sequence-derived structural sig- nal, enabling residue-level constraints to be inferred from variation across homologous sequences [53, 54, 31]. In this regime, structural reasoning was still mediated by explicitly constructed features and modular pipelines, but the field had already shifted toward data-informed inference. The integration of deep learning marked a decisive methodological turning point. Models such as AlphaFold2 and RoseTTAFold [40, 7] demonstrated that structure pre- diction could be reformulated as an end-to-end learning problem in which representation learning, geometric rea- soning, and coordinate prediction are tightly coupled. In parallel, protein language models introduced a complemen- tary route in which structural information was no longer extracted primarily through explicit evolutionary features, but learned implicitly from large-scale sequence corpora [47]. Together, these developments changed both the data regime and the representational logic of the field, shifting the emphasis from handcrafted evolutionary descriptors to learned internal representations. A further transformation occurred when the scope of pre- diction expanded beyond isolated monomer folding. Models such as AlphaFold-Multimer [23], RoseTTAFoldNA [8], RoseTTAFold All-Atom [42], and AlphaFold3 [2] moved the field toward increasingly integrated modeling of com- plexes and heterogeneous molecular systems. In this setting, the target of prediction is no longer just a single protein fold, but the structural organization of multiple interacting entities, often including nucleic acids, ligands, and ions. This shift is particularly important because it aligns struc- ture prediction more closely with biological reality, where function is frequently determined by interfaces, assemblies, and molecular context rather than by isolated folds alone. More recently, design-oriented generative modeling has introduced another conceptual change. RFdiffusion and La- Proteina use diffusion or flow-matching mechanisms to gen- erate novel backbones or sequence–structure solutions under design constraints [83, 27]. This direction extends structural modeling from inference toward generation, optimization, and molecular engineering. Taken together, these developments suggest that the recent history of protein structure prediction is not merely a story of accuracy improvement. It is also a story of method- ological expansion: from explicit evolutionary modeling to learned representations, from single-structure prediction to uncertainty-aware generation, and from isolated proteins to heterogeneous biomolecular systems. The phase-and- transition framework adopted in this review is intended to make these deeper shifts visible and to clarify how they shape the differing capabilities and practical roles of contemporary models. 6.2. Current challenges and unresolved issues Despite these advances, several important challenges remain. The first concerns the availability and distribution of data. Monomer-focused structure prediction benefits from relatively rich structural and sequence resources, whereas high-quality examples of heterogeneous complexes, alter- native conformations, and chemically diverse interactions are more unevenly distributed. This imbalance can bias training and evaluation toward well-represented subsets of biomolecular space [41, 67]. A second challenge concerns representation and inter- pretability. Explicit co-evolutionary features can be related to residue-pair constraints, whereas latent representations in PLMs, all-atom predictors, and generative models are harder to map onto biological mechanisms. Improved mechanistic interpretation would support model validation, failure anal- ysis, and appropriate scientific use [33, 69, 79]. A third unresolved issue concerns evaluation. Metrics such as TM-score, RMSD, and lDDT remain useful across monomer and complex tasks, but they are insufficient on their own as modeling scope broadens. Complex prediction requires interface-aware assessment [55]; ligand-containing systems require pose and chemistry validation [14]; dis- tributional prediction requires ensemble-aware benchmarks [38, 73]; and generative design requires measures of validity, novelty, diversity, designability, and experimental function [43, 49, 90, 9]. A fourth challenge concerns computational cost and ac- cessibility. MSA-free inference reduces query-time database search in some systems, but large PLMs incur substantial pretraining and memory costs. Heterogeneous all-atom pre- diction and iterative generative sampling can also increase inference expense. Broader deployment therefore requires transparent reporting of hardware, preprocessing, sampling count, and runtime [16, 94]. Finally, an important conceptual challenge concerns the relationship between generality and specialization. Broader architectures do not necessarily dominate every specialized setting. Antibody modeling, especially for variable CDR re- gions, illustrates how domain-focused training and inductive biases may remain valuable [60, 65, 86]. Determining when a general model is sufficient and when specialized modeling is justified will remain an important practical question. 6.3. Future directions and emerging opportunities Several directions are likely to shape the next phase of development. One is tighter integration of prediction and de- sign in iterative computational–experimental workflows, in which generative models propose candidates and predictive or experimental modules evaluate structural and functional constraints [28, 83]. Another important direction is the incorporation of dy- namics, ensembles, and context dependence. Most current predictors emphasize static structural hypotheses, whereas many proteins function through conformational transitions or partner-dependent states. Progress will require ensemble- aware training data, experimentally grounded benchmarks, and integration with statistical mechanics or molecular sim- ulation [38, 73]. A third opportunity lies in multimodal integration. Struc- tural predictions can potentially be constrained by cryo-EM He et al.: Preprint submitted to ElsevierPage 10 of 15 Protein Structure Prediction density, cross-linking, mutational scans, biochemical assays, ligand-binding measurements, and other experimental evi- dence [4, 64, 88, 25]. A fourth likely direction is the development of hybrid frameworks that combine prediction distributional infer- ence, and generative design without conflating their objec- tives. Predictive models may evaluate target compatibility [40, 2], distributional models may explore alternative states [39, 37], and design models may propose novel molecular solutions [83, 35]. Finally, future models will need greater emphasis on ef- ficiency, robustness, reproducibility, and accessibility. Use- ful reporting standards should include training-data prove- nance, inference settings, sampling counts, uncertainty cali- bration, and hardware requirements [29, 80]. 6.4. Closing perspective Protein structure prediction is no longer a narrowly de- fined problem of mapping one sequence to one static fold. It has become a broader modeling discipline that spans learned representation, interaction-aware structural reason- ing, and increasingly generative approaches to design and exploration. The methodological evolution reviewed here shows that recent progress is best understood not only in terms of benchmark performance, but also in terms of chang- ing assumptions about data, representation, objective, and scope. Seen from this perspective, the field is moving from a protein-centric prediction problem toward a more gen- eral framework for modeling biomolecular organization. At the same time, its future progress will depend on how effectively it addresses the challenges of data imbalance, interpretability, evaluation, and computational accessibility. The boundary between prediction and design is likely to become increasingly blurred, and the most influential future models may be those that integrate deterministic inference, heterogeneous system modeling, and generative sampling within coherent and practically usable frameworks. In that sense, the recent history of protein structure prediction is not only a record of rapid technical progress, but also a guide to the conceptual directions that are likely to shape its next phase. Acknowledgements The authors are grateful to the support of the Guangdong Key Disciplines Project (2024ZDJS137). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. CRediT authorship contribution statement Wengan He: Conceptualization, Methodology, Inves- tigation, Data curation, Visualization, Writing – original draft . Yongsheng Luo: Conceptualization, Methodology, Investigation, Writing – original draft . Lihong Jiang: Vali- dation, Writing – review & editing . Wenhui Xu: Validation, Writing – review & editing . Yu Li: Conceptualization, Supervision, Writing – review & editing . Data availability statement No new datasets were generated or analyzed in this review. All information discussed in the article was ob- tained from publicly available publications, databases, and resources cited in the manuscript. Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work, the authors used ChatGPT, provided by OpenAI, to assist with language refinement, structural organization, and the drafting of se- lected manuscript and submission-related text. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the article. References [1] Abanades, B., Wong, W.K., Boyles, F., Georges, G., Bujotzek, A., Deane, C.M., 2023. Immunebuilder: Deep-learning models for pre- dicting the structures of immune proteins. Communications Biology 6, 575. URL: https://doi.org/10.1038/s42003-023-04927-7, doi:10. 1038/s42003-023-04927-7. [2] Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A.J., Bambrick, J., Boden- stein, S.W., Evans, D.A., Hung, C.C., O’Neill, M., Reiman, D., Tun- yasuvunakool, K., Wu, Z., Žemgulyt ̇ e, A., Arvaniti, E., Beattie, C., Bertolli, O., Bridgland, A., Cherepanov, A., Congreve, M., Cowen- Rivers, A.I., Cowie, A., Figurnov, M., Fuchs, F.B., Gladman, H., Jain, R., Khan, Y.A., Low, C.M.R., Perlin, K., Potapenko, A., Savy, P., Singh, S., Stecula, A., Thillaisundaram, A., Tong, C., Yakneen, S., Zhong, E.D., Zielinski, M., Žídek, A., Bapst, V., Kohli, P., Jaderberg, M., Hassabis, D., Jumper, J.M., 2024. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature 630, 493– 500. URL: https://doi.org/10.1038/s41586-024-07487-w, doi:10. 1038/s41586-024-07487-w. [3] del Alamo, D., Sala, D., Mchaourab, H.S., Meiler, J., 2022. Sampling alternative conformational states of transporters and receptors with alphafold2. eLife 11, e75751. URL: https://doi.org/10.7554/eLife. 75751, doi:10.7554/eLife.75751. [4] Alber, F., Förster, F., Korkin, D., Topf, M., Sali, A., 2008. Integrating diverse data for structure determination of macro- molecular assemblies. Annual Review of Biochemistry 77, 443–477. URL: https://w.annualreviews.org/content/journals/ 10.1146/annurev.biochem.77.060407.135530, doi:https://doi.org/10. 1146/annurev.biochem.77.060407.135530. [5] AlQuraishi, M., 2019. End-to-end differentiable learning of protein structure. Cell Systems 8, 292–301.e3. URL: https: //w.sciencedirect.com/science/article/pii/S2405471219300766, doi:https://doi.org/10.1016/j.cels.2019.03.006. [6] Anfinsen, C.B., 1973.Principles that govern the folding of protein chains.Science 181, 223– 230.URL: https://w.science.org/doi/abs/10.1126/ science.181.4096.223,doi:10.1126/science.181.4096.223, arXiv:https://w.science.org/doi/pdf/10.1126/science.181.4096.223. He et al.: Preprint submitted to ElsevierPage 11 of 15 Protein Structure Prediction [7] Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G.R., Wang, J., Cong, Q., Kinch, L.N., Schaeffer, R.D., Millán, C., Park, H., Adams, C., Glassman, C.R., DeGiovanni, A., Pereira, J.H., Rodrigues, A.V., van Dijk, A.A., Ebrecht, A.C., Opper- man, D.J., Sagmeister, T., Buhlheller, C., Pavkov-Keller, T., Rathi- naswamy, M.K., Dalwadi, U., Yip, C.K., Burke, J.E., Garcia, K.C., Grishin, N.V., Adams, P.D., Read, R.J., Baker, D., 2021. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876. URL: https://w.science. org/doi/abs/10.1126/science.abj8754, doi:10.1126/science.abj8754, arXiv:https://w.science.org/doi/pdf/10.1126/science.abj8754. [8] Baek, M., McHugh, R., Anishchenko, I., Jiang, H., Baker, D., DiMaio, F., 2024. Accurate prediction of protein–nucleic acid complexes using rosettafoldna. Nature Methods 21, 117–121. URL: https://doi.org/ 10.1038/s41592-023-02086-5, doi:10.1038/s41592-023-02086-5. [9] Barnett, A.J., KC, R., Pandey, P., Somasiri, P., Fairfax, K.A., Hung, S., Hewitt, A.W., 2026. Benchmarking generative ai protein models re- veals differences between structural and sequence-based approaches. Genomics, Proteomics & Bioinformatics , qzag014URL: https:// doi.org/10.1093/gpbjnl/qzag014, doi:10.1093/gpbjnl/qzag014. [10] Basu, S., Wallner, B., 2016. Dockq: A quality measure for protein- protein docking models. PLOS ONE 11, 1–9. URL: https://doi.org/ 10.1371/journal.pone.0161879, doi:10.1371/journal.pone.0161879. [11] Boehr, D.D., Nussinov, R., Wright, P.E., 2009. The role of dy- namic conformational ensembles in biomolecular recognition. Na- ture Chemical Biology 5, 789–796. URL: https://doi.org/10.1038/ nchembio.232, doi:10.1038/nchembio.232. [12] Brini, E., Simmerling, C., Dill, K., 2020. Protein storytelling through physics. Science 370, eaaz3041. URL: https://w.science. org/doi/abs/10.1126/science.aaz3041, doi:10.1126/science.aaz3041, arXiv:https://w.science.org/doi/pdf/10.1126/science.aaz3041. [13] Bryant, P., Pozzati, G., Elofsson, A., 2022. Improved prediction of protein-protein interactions using alphafold2. Nature Communica- tions 13, 1265. URL: https://doi.org/10.1038/s41467-022-28865-w, doi:10.1038/s41467-022-28865-w. [14] Buttenschoen, M., Morris, G.M., Deane, C.M., 2024. Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences†electronic supplementary infor- mation (esi) available. see doi: https://doi.org/10.1039/d3sc04185a. Chemical Science 15, 3130–3139. URL: https://w.sciencedirect. com/science/article/pii/S2041652024002402, doi:https://doi.org/10. 1039/d3sc04185a. [15] Cao, H., Zhou, X., Gao, Z., Wang, C., Gao, X., Zhang, Z., de la Fuente-Nunez, C., Gu, C., Liu, G., Heng, P.A., 2025. Lightweight msa design advances protein folding from evolutionary embeddings. URL: https://arxiv.org/abs/2507.07032, arXiv:2507.07032. [16] Cheng, X., Chen, B., Li, P., Gong, J., Tang, J., Song, L., 2024. Training compute-optimal protein language models, in: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc.. p. 69386–69418. URL: https://proceedings.neurips.c/paper_files/paper/2024/ file/8066ae1446b2bbccb5159587c3b3bcc-Paper-Conference.pdf, doi:10.52202/079017-2216. [17] wwPDB consortium, 2019. Protein data bank: the single global archive for 3d macromolecular structure data. Nucleic Acids Research 47, D520–D528. URL: https://doi.org/10.1093/nar/gky949, doi:10. 1093/nar/gky949. [18] Dill, K.A., MacCallum, J.L., 2012. The protein-folding problem, 50 years on. Science 338, 1042–1046. URL: https://w.science. org/doi/abs/10.1126/science.1219021, doi:10.1126/science.1219021, arXiv:https://w.science.org/doi/pdf/10.1126/science.1219021. [19] Durairaj, J., Adeshina, Y., Cao, Z., Zhang, X., Oleinikovas, V., Duignan, T., McClure, Z., Robin, X., Kovtun, D., Rossi, E., Zhou, G., Veccham, S., Isert, C., Peng, Y., Sundareson, P., Akdel, M., Corso, G., Stärk, H., Carpenter, Z., Bronstein, M., Kucukbenli, E., Schwede, T., Naef, L., 2024. Plinder: The protein-ligand interactions dataset and evaluation resource. bioRxiv URL: https://w.biorxiv. org/content/early/2024/07/17/2024.07.17.603955, doi:10.1101/2024. 07.17.603955. [20] Eddy, S.R., 2011.Accelerated profile hmm searches. PLOS Computational Biology 7, e1002195.URL: https: //app.dimensions.ai/details/publication/pub.1001034143, doi:10.1371/journal.pcbi.1002195. [21] Elnaggar, A., Heinzinger, M., Dallago, C., Rehawi, G., Wang, Y., Jones, L., Gibbs, T., Feher, T., Angerer, C., Steinegger, M., Bhowmik, D., Rost, B., 2022. Prottrans: Toward understanding the language of life through self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 7112–7127. doi:10.1109/ TPAMI.2021.3095381. [22] Elofsson, A., 2023. Progress at protein structure prediction, as seen in casp15. Current Opinion in Structural Biology 80, 102594. URL: https://w.sciencedirect.com/science/article/pii/ S0959440X23000684, doi:https://doi.org/10.1016/j.sbi.2023.102594. [23] Evans, R., O’Neill, M., Pritzel, A., Antropova, N., Senior, A., Green, T., Žídek, A., Bates, R., Blackwell, S., Yim, J., Ronneberger, O., Bodenstein, S., Zielinski, M., Bridgland, A., Potapenko, A., Cowie, A., Tunyasuvunakool, K., Jain, R., Clancy, E., Kohli, P., Jumper, J., Hassabis, D., 2021. Protein complex prediction with alphafold- multimer. bioRxiv URL: https://w.biorxiv.org/content/early/ 2021/10/04/2021.10.04.463034, doi:10.1101/2021.10.04.463034. [24] Fang, X., Wang, F., Liu, L., He, J., Lin, D., Xiang, Y., Zhu, K., Zhang, X., Wu, H., Li, H., Song, L., 2023. A method for multiple- sequence-alignment-free protein structure prediction using a pro- tein language model. Nature Machine Intelligence 5, 1087–1096. URL: http://dx.doi.org/10.1038/s42256-023-00721-6, doi:10.1038/ s42256-023-00721-6. [25] Feng, S., Chen, Z., Zhang, C., Xie, Y., Ovchinnikov, S., Gao, Y.Q., Liu, S., 2024. Integrated structure prediction of protein– protein docking with experimental restraints using colabdock. Nature Machine Intelligence 6, 924–935. URL: https://doi.org/10.1038/ s42256-024-00873-z, doi:10.1038/s42256-024-00873-z. [26] Finn, R.D., Clements, J., Eddy, S.R., 2011. Hmmer web server: interactive sequence similarity searching. Nucleic Acids Research 39, W29–W37. URL: https://doi.org/10.1093/nar/gkr367, doi:10.1093/ nar/gkr367. [27] Geffner, T., Didi, K., Cao, Z., Reidenbach, D., Zhang, Z., Dallago, C., Kucukbenli, E., Kreis, K., Vahdat, A., 2026. La-proteina: Atomistic protein generation via partially latent flow matching. URL: https: //arxiv.org/abs/2507.09466, arXiv:2507.09466. [28] He, B., Qin, C., Zhao, Y., Huang, L.K., Wu, Z., Wang, F., Wu, F., Yang, F., Yao, J., 2026. Functional protein design and enhancement with ontology reinforcement iteration. Nature Communications 17, 4158. URL: https://doi.org/10.1038/s41467-026-69855-6, doi:10. 1038/s41467-026-69855-6. [29] Heil, B.J., Hoffman, M.M., Markowetz, F., Lee, S.I., Greene, C.S., Hicks, S.C., 2021. Reproducibility standards for machine learning in the life sciences. Nature Methods 18, 1132–1135. URL: https://doi. org/10.1038/s41592-021-01256-7, doi:10.1038/s41592-021-01256-7. [30] Ho, J., Jain, A., Abbeel, P., 2020. Denoising diffusion probabilistic models, in: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc.. p. 6840–6851.URL: https://proceedings.neurips.c/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf. [31] Hopf, T.A., Schärfe, C.P.I., Rodrigues, J.P.G.L.M., Green, A.G., Kohlbacher, O., Sander, C., Bonvin, A.M.J.J., Marks, D.S., 2014. Sequence co-evolution gives 3d contacts and structures of protein complexes. eLife 3, e03430. URL: https://doi.org/10.7554/eLife. 03430, doi:10.7554/eLife.03430. [32] Hu, B., Xia, J., Zheng, J., Tan, C., Huang, Y., Xu, Y., Li, S.Z., 2022. Protein language models and structure prediction: Con- nection and progression. URL: https://arxiv.org/abs/2211.16742, arXiv:2211.16742. He et al.: Preprint submitted to ElsevierPage 12 of 15 Protein Structure Prediction [33] Hunklinger, A., Ferruz, N., 2026. Towards the explainability of protein language models. Nature Machine Intelligence 8, 649– 662. URL: https://doi.org/10.1038/s42256-026-01232-w, doi:10. 1038/s42256-026-01232-w. [34] Jendrusch, M., Korbel, J.O., 2025. Efficient protein structure gen- eration with sparse denoising models. Nature Machine Intelligence 7, 1429–1445. URL: https://doi.org/10.1038/s42256-025-01100-z, doi:10.1038/s42256-025-01100-z. [35] Jendrusch, M.A., Yang, A.L.J., Cacace, E., Bobonis, J., Voogdt, C.G.P., Kaspar, S., Schweimer, K., Perez-Borrajero, C., Lapouge, K., Scheurich, J., Remans, K., Hennig, J., Typas, A., Korbel, J.O., Sadiq, S.K., 2025. Alphadesign: a de novo protein design frame- work based on alphafold. Molecular Systems Biology 21, 1166– 1189. URL: https://doi.org/10.1038/s44320-025-00119-z, doi:10. 1038/s44320-025-00119-z. [36] Jiang, Y., Li, X., Zhang, Y., Han, J., Xu, Y., Pandit, A., ZHANG, Z., Wang, M., Wang, M., Liu, C., Yang, G., Choi, Y., Lu, Y., Li, W.J., Fu, T., Wu, F., Liu, J., 2026. Posex: AI defeats physics- based methods on protein ligand cross-docking, in: The Fourteenth International Conference on Learning Representations. URL: https: //openreview.net/forum?id=qqzxKudD4T. [37] Jing, B., Berger, B., Jaakkola, T., 2024. Alphafold meets flow matching for generating protein ensembles, in: Proceedings of the 41st International Conference on Machine Learning, JMLR.org. [38] Jing, B., Berger, B., Jaakkola, T., 2026. Ai-based methods for simu- lating, sampling, and predicting protein ensembles. Current Opinion in Structural Biology 98, 103251. URL: https://w.sciencedirect. com/science/article/pii/S0959440X26000333, doi:https://doi.org/10. 1016/j.sbi.2026.103251. [39] Jing, B., Erives, E., Pao-Huang, P., Corso, G., Berger, B., Jaakkola, T., 2023. Eigenfold: Generative protein structure prediction with diffusion models. URL: https://arxiv.org/abs/2304.02198, arXiv:2304.02198. [40] Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ron- neberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S.A.A., Ballard, A.J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back, T., Petersen, S., Reiman, D., Clancy, E., Zielinski, M., Steinegger, M., Pacholska, M., Berghammer, T., Bodenstein, S., Silver, D., Vinyals, O., Senior, A.W., Kavukcuoglu, K., Kohli, P., Hassabis, D., 2021. Highly accurate protein structure prediction with alphafold. Nature 596, 583–589. URL: https://doi.org/10.1038/s41586-021-03819-2, doi:10.1038/s41586-021-03819-2. [41] Kosoglu, K., Aydin, Z., Tuncbag, N., Gursoy, A., Keskin, O., 2024. Structural coverage of the human interactome. Briefings in Bioin- formatics 25, bbad496. URL: https://doi.org/10.1093/bib/bbad496, doi:10.1093/bib/bbad496. [42] Krishna, R., Wang, J., Ahern, W., Sturmfels, P., Venkatesh, P., Kalvet, I., Lee, G.R., Morey-Burrows, F.S., Anishchenko, I., Humphreys, I.R., McHugh, R., Vafeados, D., Li, X., Sutherland, G.A., Hitchcock, A., Hunter, C.N., Kang, A., Brackenbrough, E., Bera, A.K., Baek, M., DiMaio, F., Baker, D., 2024. Gen- eralized biomolecular modeling and design with rosettafold all- atom. Science 384, eadl2528. URL: https://w.science. org/doi/abs/10.1126/science.adl2528, doi:10.1126/science.adl2528, arXiv:https://w.science.org/doi/pdf/10.1126/science.adl2528. [43] Kuang, J., Liu, N., Wang, J., Sun, C., Ji, T., Wu, Y., 2026. Pdfbench: A benchmark for de novo protein design from function. URL: https: //arxiv.org/abs/2505.20346, arXiv:2505.20346. [44] Laine, E., Eismann, S., Elofsson, A., Grudinin, S., 2021. Protein sequence-to-structure learning: Is this the end(-to-end revolution)? URL: https://arxiv.org/abs/2105.07407, arXiv:2105.07407. [45] Larkin, M., Blackshields, G., Brown, N., Chenna, R., McGettigan, P., McWilliam, H., Valentin, F., Wallace, I., Wilm, A., Lopez, R., Thompson, J., Gibson, T., Higgins, D., 2007. Clustal w and clustal x version 2.0. Bioinformatics 23, 2947–2948. URL: https://doi.org/ 10.1093/bioinformatics/btm404, doi:10.1093/bioinformatics/btm404. [46] Liberti, L., Lavor, C., Maculan, N., Mucherino, A., 2014. Eu- clidean distance geometry and applications. SIAM Review 56, 3– 69. URL: https://doi.org/10.1137/120875909, doi:10.1137/120875909, arXiv:https://doi.org/10.1137/120875909. [47] Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., dos Santos Costa, A., Fazel- Zarandi, M., Sercu, T., Candido, S., Rives, A., 2023. Evolutionary- scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130. URL: https://w.science. org/doi/abs/10.1126/science.ade2574, doi:10.1126/science.ade2574, arXiv:https://w.science.org/doi/pdf/10.1126/science.ade2574. [48] Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M., 2023. Flow matching for generative modeling. URL: https://arxiv.org/ abs/2210.02747, arXiv:2210.02747. [49] Liu, C., Ren, M., Guan, J., Gong, C., Sun, J., Chen, X., Xiao, W., 2026. Protdbench: A unified benchmark of protein binder design and evalu- ation. URL: https://arxiv.org/abs/2605.04118, arXiv:2605.04118. [50] Liu, L., Zhang, S., Xue, Y., Ye, X., Zhu, K., Li, Y., Liu, Y., Gao, J., Zhao, W., Yu, H., Wu, Z., Zhang, X., Fang, X., 2024. Technical report of helixfold3 for biomolecular structure prediction. URL: https://arxiv.org/abs/2408.16975, arXiv:2408.16975. [51] Liu, Z., Li, Y., Han, L., Li, J., Liu, J., Zhao, Z., Nie, W., Liu, Y., Wang, R., 2015. Pdb-wide collection of binding data: current status of the pdbbind database. Bioinformatics 31, 405–412. URL: https:// doi.org/10.1093/bioinformatics/btu626, doi:10.1093/bioinformatics/ btu626. [52] Mariani, V., Biasini, M., Barbato, A., Schwede, T., 2013. lddt: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics 29, 2722– 2728. URL: https://doi.org/10.1093/bioinformatics/btt473, doi:10. 1093/bioinformatics/btt473. [53] Marks, D.S., Colwell, L.J., Sheridan, R., Hopf, T.A., Pagnani, A., Zecchina, R., Sander, C., 2011.Protein 3d structure computed from evolutionary sequence variation. PloS one 6, e28766. URL: https://europepmc.org/articles/PMC3233603, doi:10. 1371/journal.pone.0028766. [54] Marks, D.S., Hopf, T.A., Sander, C., 2012. Protein structure predic- tion from sequence variation. Nature Biotechnology 30, 1072–1080. URL: https://doi.org/10.1038/nbt.2419, doi:10.1038/nbt.2419. [55] Mirabello, C., Wallner, B., 2024. Dockq v2: improved automatic quality measure for protein multimers, nucleic acids, and small molecules. Bioinformatics 40, btae586. URL: https://doi.org/10. 1093/bioinformatics/btae586, doi:10.1093/bioinformatics/btae586. [56] Mirdita, M., Schütze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., Steinegger, M., 2022. Colabfold: making protein folding accessible to all. Nature Methods 19, 679–682. URL: https://doi.org/10.1038/ s41592-022-01488-1, doi:10.1038/s41592-022-01488-1. [57] Morehead, A., Giri, N., Liu, J., Neupane, P., Cheng, J., 2026. As- sessing the potential of deep learning for protein–ligand docking. Nature Machine Intelligence 8, 32–41. URL: https://doi.org/10. 1038/s42256-025-01160-1, doi:10.1038/s42256-025-01160-1. [58] Motlagh, H.N., Wrabl, J.O., Li, J., Hilser, V.J., 2014. The ensemble nature of allostery. Nature 508, 331–339. URL: https://doi.org/10. 1038/nature13001, doi:10.1038/nature13001. [59] Pearce, R., Zhang, Y., 2021. Deep learning techniques have significantly impacted protein structure prediction and protein design.Current Opinion in Structural Biology 68, 194– 207. URL: https://w.sciencedirect.com/science/article/pii/ S0959440X21000142, doi:https://doi.org/10.1016/j.sbi.2021.01.007. protein-Carbohydrate Complexes and Glycosylation Sequences and Topology. [60] Polonsky, K., Pupko, T., Freund, N.T., 2023. Evaluation of the ability of alphafold to predict the three-dimensional structures of antibodies and epitopes. The Journal of Immunology 211, 1578– 1588. URL: https://doi.org/10.4049/jimmunol.2300150, doi:10.4049/ jimmunol.2300150. [61] Rapp, J.T., Bremer, B.J., Romero, P.A., 2024. Self-driving laborato- ries to autonomously navigate the protein fitness landscape. Nature He et al.: Preprint submitted to ElsevierPage 13 of 15 Protein Structure Prediction Chemical Engineering 1, 97–107. URL: https://doi.org/10.1038/ s44286-023-00002-4, doi:10.1038/s44286-023-00002-4. [62] Remmert, M., Biegert, A., Hauser, A., Söding, J., 2012. Hhblits: lightning-fast iterative protein sequence searching by hmm-hmm alignment. Nature Methods 9, 173–175. URL: https://doi.org/10. 1038/nmeth.1818, doi:10.1038/nmeth.1818. [63] Rennie, M.L., Oliver, M.R., 2025. Emerging frontiers in protein structure prediction following the alphafold revolution. Journal of The Royal Society Interface 22, 20240886. URL: https://doi.org/ 10.1098/rsif.2024.0886, doi:10.1098/rsif.2024.0886. [64] Rout, M.P., Sali, A., 2019. Principles for integrative structural biology studies. Cell 177, 1384–1403. URL: https://w.sciencedirect. com/science/article/pii/S0092867419305148, doi:https://doi.org/10. 1016/j.cell.2019.05.016. [65] Ruffolo, J.A., Chu, L.S., Mahajan, S.P., Gray, J.J., 2023. Fast, accurate antibody structure prediction from deep learning on mas- sive set of natural antibodies. Nature Communications 14, 2389. URL: https://doi.org/10.1038/s41467-023-38063-x, doi:10. 1038/s41467-023-38063-x. [66] Senior, A.W., Evans, R., Jumper, J., Kirkpatrick, J., Sifre, L., Green, T., Qin, C., Žídek, A., Nelson, A.W.R., Bridgland, A., Penedones, H., Petersen, S., Simonyan, K., Crossan, S., Kohli, P., Jones, D.T., Silver, D., Kavukcuoglu, K., Hassabis, D., 2020. Improved protein structure prediction using potentials from deep learning. Nature 577, 706–710. URL: https://doi.org/10.1038/s41586-019-1923-7, doi:10.1038/s41586-019-1923-7. [67] Shinada, N.K., Schmidtke, P., de Brevern, A.G., 2020. Accurate repre- sentation of protein-ligand structural diversity in the protein data bank (pdb). International Journal of Molecular Sciences 21. URL: https: //w.mdpi.com/1422-0067/21/6/2243, doi:10.3390/ijms21062243. [68] Monteiro da Silva, G., Cui, J.Y., Dalgarno, D.C., Lisi, G.P., Ruben- stein, B.M., 2024. High-throughput prediction of protein conforma- tional distributions with subsampled alphafold2. Nature Communica- tions 15, 2464. URL: https://doi.org/10.1038/s41467-024-46715-9, doi:10.1038/s41467-024-46715-9. [69] Simon, E., Zou, J., 2025. Interplm: discovering interpretable features in protein language models via sparse autoencoders. Nature Methods 22, 2107–2117. URL: https://doi.org/10.1038/s41592-025-02836-7, doi:10.1038/s41592-025-02836-7. [70] Simons, K.T., Kooperberg, C., Huang, E., Baker, D., 1997. As- sembly of protein tertiary structures from fragments with sim- ilar local sequences using simulated annealing and bayesian scoring functions. Journal of Molecular Biology 268, 209– 225. URL: https://w.sciencedirect.com/science/article/pii/ S0022283697909591, doi:https://doi.org/10.1006/jmbi.1997.0959. [71] Singh, N., Lane, S., Yu, T., Lu, J., Ramos, A., Cui, H., Zhao, H., 2025. A generalized platform for artificial intelligence-powered autonomous enzyme engineering. Nature Communications 16, 5648. URL: https://doi.org/10.1038/s41467-025-61209-y, doi:10. 1038/s41467-025-61209-y. [72] Sippl, M.J., 1990. Calculation of conformational ensembles from potentials of mean force: An approach to the knowledge-based prediction of local structures in globular proteins. Journal of Molecular Biology 213, 859–883. URL: https://w.sciencedirect. com/science/article/pii/S0022283605802694, doi:https://doi.org/10. 1016/S0022-2836(05)80269-4. [73] Sledzieski, S., Hanson, S.M., 2026. The landscape of machine learning approaches for modeling protein conformational ensembles. Current Opinion in Structural Biology 98, 103253. URL: https: //w.sciencedirect.com/science/article/pii/S0959440X26000357, doi:https://doi.org/10.1016/j.sbi.2026.103253. [74] Steinegger, M., Söding, J., 2017. Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature Biotechnology 35, 1026–1028. URL: https://doi.org/10.1038/nbt. 3988, doi:10.1038/nbt.3988. [75] team, C.D., Boitreaud, J., Dent, J., McPartlon, M., Meier, J., Reis, V., Rogozhonikov, A., Wu, K., 2024. Chai-1: Decoding the molecular interactions of life. bioRxiv URL: https://w.biorxiv.org/content/ early/2024/10/11/2024.10.10.615955, doi:10.1101/2024.10.10.615955. [76] Thomasen, F.E., Lindorff-Larsen, K., 2022. Conformational ensembles of intrinsically disordered proteins and flexible mul- tidomain proteins. Biochemical Society Transactions 50, 541– 554. URL: https://w.sciencedirect.com/science/article/pii/ S1470875222001143, doi:https://doi.org/10.1042/BST20210499. [77] Tunyasuvunakool, K., Adler, J., Wu, Z., Green, T., Zielinski, M., Žídek, A., Bridgland, A., Cowie, A., Meyer, C., Laydon, A., Velankar, S., Kleywegt, G.J., Bateman, A., Evans, R., Pritzel, A., Figurnov, M., Ronneberger, O., Bates, R., Kohl, S.A.A., Potapenko, A., Bal- lard, A.J., Romera-Paredes, B., Nikolov, S., Jain, R., Clancy, E., Reiman, D., Petersen, S., Senior, A.W., Kavukcuoglu, K., Birney, E., Kohli, P., Jumper, J., Hassabis, D., 2021. Highly accurate protein structure prediction for the human proteome. Nature 596, 590– 596. URL: https://doi.org/10.1038/s41586-021-03828-1, doi:10. 1038/s41586-021-03828-1. [78] Vendruscolo, M., Kussell, E., Domany, E., 1997. Recovery of protein structure from contact maps. Folding and Design 2, 295–306. URL: https://w.sciencedirect.com/science/article/ pii/S1359027897000412, doi:https://doi.org/10.1016/S1359-0278(97) 00041-2. [79] Vig, J., Madani, A., Varshney, L.R., Xiong, C., Socher, R., Ra- jani, N., 2021. BERTology Meets Biology: Interpreting Atten- tion in Protein Language Models, in: International Conference on Learning Representations. URL: https://mlanthology.org/iclr/ 2021/vig2021iclr-bertology/. [80] Walsh, I., Fishman, D., Garcia-Gasulla, D., Titma, T., Pollastri, G., Capriotti, E., Casadio, R., Capella-Gutierrez, S., Cirillo, D., Del Conte, A., Dimopoulos, A.C., Del Angel, V.D., Dopazo, J., Fariselli, P., Fernández, J.M., Huber, F., Kreshuk, A., Lenaerts, T., Martelli, P.L., Navarro, A., Broin, P.Ó., Piñero, J., Piovesan, D., Reczko, M., Ronzano, F., Satagopam, V., Savojardo, C., Spiwok, V., Tangaro, M.A., Tartari, G., Salgado, D., Valencia, A., Zam- belli, F., Harrow, J., Psomopoulos, F.E., Tosatto, S.C.E., Group, E.M.L.F., 2021. Dome: recommendations for supervised ma- chine learning validation in biology. Nature Methods 18, 1122– 1127. URL: https://doi.org/10.1038/s41592-021-01205-4, doi:10. 1038/s41592-021-01205-4. [81] Wang, C., Alamdari, S., Domingo-Enrich, C., Amini, A.P., Yang, K.K., 2025. Toward deep learning sequence–structure co-generation for protein design. Current Opinion in Structural Biology 91, 103018. URL: https://w.sciencedirect.com/science/article/pii/ S0959440X25000363, doi:https://doi.org/10.1016/j.sbi.2025.103018. [82] Wang, S., Sun, S., Li, Z., Zhang, R., Xu, J., 2017. Accurate de novo prediction of protein contact map by ultra-deep learning model. PLOS Computational Biology 13, e1005324. URL: https:// app.dimensions.ai/details/publication/pub.1010491660, doi:10.1371/ journal.pcbi.1005324. https://doi.org/10.1371/journal.pcbi.1005324. [83] Watson, J.L., Juergens, D., Bennett, N.R., Trippe, B.L., Yim, J., Eisenach, H.E., Ahern, W., Borst, A.J., Ragotte, R.J., Milles, L.F., Wicky, B.I.M., Hanikel, N., Pellock, S.J., Courbet, A., Sheffler, W., Wang, J., Venkatesh, P., Sappington, I., Torres, S.V., Lauko, A., De Bortoli, V., Mathieu, E., Ovchinnikov, S., Barzilay, R., Jaakkola, T.S., DiMaio, F., Baek, M., Baker, D., 2023. De novo design of protein structure and function with rfdiffusion. Nature 620, 1089– 1100. URL: https://doi.org/10.1038/s41586-023-06415-8, doi:10. 1038/s41586-023-06415-8. [84] Wayment-Steele, H.K., Ojoawo, A., Otten, R., Apitz, J.M., Pitsawong, W., Hömberger, M., Ovchinnikov, S., Colwell, L., Kern, D., 2024. Predicting multiple conformations via sequence clustering and al- phafold2. Nature 625, 832–839. URL: https://doi.org/10.1038/ s41586-023-06832-9, doi:10.1038/s41586-023-06832-9. [85] Wohlwend, J., Corso, G., Passaro, S., Reveiz, M., Leidal, K., Swider- ski, W., Portnoi, T., Chinn, I., Silterra, J., Jaakkola, T., Barzilay, R., 2024. Boltz-1 democratizing biomolecular interaction model- ing. bioRxiv URL: https://w.biorxiv.org/content/early/2024/11/ 20/2024.11.19.624167, doi:10.1101/2024.11.19.624167. He et al.: Preprint submitted to ElsevierPage 14 of 15 Protein Structure Prediction [86] Woo, H., Kim, Y., Seok, C., 2024. Protein loop structure prediction by community-based deep learning and its application to antibody cdr h3 loop modeling. PLOS Computational Biology 20, 1–17. URL: https://doi.org/10.1371/journal.pcbi.1012239, doi:10.1371/journal. pcbi.1012239. [87] Wu, R., Ding, F., Wang, R., Shen, R., Zhang, X., Luo, S., Su, C., Wu, Z., Xie, Q., Berger, B., Ma, J., Peng, J., 2022. High- resolution de novo structure prediction from primary sequence. bioRxiv URL: https://w.biorxiv.org/content/early/2022/07/22/ 2022.07.21.500999, doi:10.1101/2022.07.21.500999. [88] Xie, Y., Zhang, C., Li, S., Du, X., Lu, Y., Wang, M., Hu, Y., Chen, Z., Liu, S., Gao, Y.Q., 2025. Integrating diverse experimental information to assist protein complex structure prediction by grasp. Nature Methods 22, 2362–2374. URL: https://doi.org/10.1038/ s41592-025-02820-1, doi:10.1038/s41592-025-02820-1. [89] Xu, S., Feng, Q., Qiao, L., Wu, H., Shen, T., Cheng, Y., Zheng, S., Sun, S., 2025. Benchmarking all-atom biomolecular structure prediction with foldbench. Nature Communications 17, 442. URL: https://doi. org/10.1038/s41467-025-67127-3, doi:10.1038/s41467-025-67127-3. [90] Ye, F., Zheng, Z., Xue, D., Shen, Y., Wang, L., Ma, Y., Wang, Y., Wang, X., Zhou, X., Gu, Q., 2024. Proteinbench: A holistic evaluation of protein foundation models. URL: https://arxiv.org/abs/2409. 06744, arXiv:2409.06744. [91] Yue, A., Wang, Z., Xu, H., 2025. Reqflow: Rectified quaternion flow for efficient and high-quality protein backbone generation. URL: https://arxiv.org/abs/2502.14637, arXiv:2502.14637. [92] Zhang, Y., Skolnick, J., 2004.Scoring function for automated assessment of protein structure template quality. Proteins: Structure, Function, and Bioinformatics 57, 702– 710.URL: https://onlinelibrary.wiley.com/doi/abs/10. 1002/prot.20264,doi:https://doi.org/10.1002/prot.20264, arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/prot.20264. [93] Zhu, W., Shenoy, A., Kundrotas, P., Elofsson, A., 2023. Evalua- tion of alphafold-multimer prediction on multi-chain protein com- plexes. Bioinformatics 39, btad424. URL: https://doi.org/10.1093/ bioinformatics/btad424, doi:10.1093/bioinformatics/btad424. [94] Çelik, M.H., Xie, X., 2025.Efficient inference, training, and fine-tuning of protein language models.iScience 28, 113495. URL: https://w.sciencedirect.com/science/article/pii/ S2589004225017560, doi:https://doi.org/10.1016/j.isci.2025.113495. [95] Šali, A., Blundell, T.L., 1993. Comparative protein modelling by satisfaction of spatial restraints. Journal of Molecular Biology 234, 779–815. URL: https://w.sciencedirect.com/science/article/ pii/S0022283683716268, doi:https://doi.org/10.1006/jmbi.1993.1626. He et al.: Preprint submitted to ElsevierPage 15 of 15

Instant research discovery