Paper deep dive

BERT4DTI : BERT-based Model for Predicting Drug-Protein Interactions

Thanina Hamitouch, Khadidja Henni, Abdelkrim Arie, Amina Selma Haichour, Neila Mezghani, Lina Abou-Abbas

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 30

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Understanding how drugs interact with protein targets is fundamental to drug discovery, drug repurposing and the early identification of promising therapeutic candidates before costly experimental testing. Sequence-based DTI models face three practical limitations: labelled interactions are scarce and unevenly distributed, large pretrained chemical and protein encoders are expensive to fine-tune end-to-end, and independently encoded sequences do not capture pair-specific dependencies. We present BERT4DTI, which encodes SMILES strings with ChemBERTa and amino-acid sequences with ProtBERT, applies bidirectional mutual attention between token-level representations, and classifies the resulting interaction features using convolutional layers and a multilayer perceptron. To reduce trainable size, ProtBERT is truncated to 18 retained layers and only the last two layers of each encoder are fine-tuned. On BIOSNAP, DAVIS and BindingDB, BERT4DTI is competitive, achieving the best ROC-AUC and PR-AUC on BIOSNAP and the highest sensitivity on all three benchmarks. An ablation on DAVIS shows that mutual attention improves PR-AUC and specificity. With 125M trainable parameters compared with 353M for full BERT fine-tuning, BERT4DTI provides a favourable performance-parameter trade-off for sequence-based DTI screening, while leaving runtime profiling, calibration and leakage-audited validation for future work.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

26,942 characters extracted from source content.

Expand or collapse full text

BERT4DTI : BERT-based Model for Predicting Drug-Protein Interactions Conference: Proceedings of the 35th ACM International Conference on Information and Knowledge Management; November 07–11, 2026; Rome, ItalyProceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM ’26), November 07–11, 2026, Rome, ItalyDOI: 10.1145/3799682.3840046ISBN: 979-8-4007-2539-5/2026/11CCS: Computing methodologies Machine learning approachesCCS: Applied computing BioinformaticsCCS: Computing methodologies Information extraction Thanina Hamitouch Affiliation: Ecole Nationale Supérieure d’Informatique (ESI), Algiers, Algeria , Khadidja Henni Affiliation: Institut d’Intelligence Artificielle Appliquée, TELUQ University, Montreal, Canada email: khadidja.henni@teluq.ca , Abdelkrim Aries Affiliation: Ecole Nationale Supérieure d’Informatique (ESI), Algiers, Algeria email: ab_aries@esi.dz , Amina Selma Haichour Affiliation: Ecole Nationale Supérieure d’Informatique (ESI), Algiers, Algeria , Neila Mezghani Affiliation: Institut d’Intelligence Artificielle Appliquée, TELUQ University, Montreal, Canada and Lina Abou-Abbas Affiliation: Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon © c Abstract. Understanding how drugs interact with protein targets is fundamental to drug discovery, drug repurposing and the early identification of promising therapeutic candidates before costly experimental testing. Sequence-based DTI models face three practical limitations: labelled interactions are scarce and unevenly distributed, large pretrained chemical and protein encoders are expensive to fine-tune end-to-end, and independently encoded sequences do not capture pair-specific dependencies. We present BERT4DTI, which encodes SMILES strings with ChemBERTa and amino-acid sequences with ProtBERT, applies bidirectional mutual attention between token-level representations, and classifies the resulting interaction features using convolutional layers and a multilayer perceptron. To reduce trainable size, ProtBERT is truncated to 18 retained layers and only the last two layers of each encoder are fine-tuned. On BIOSNAP, DAVIS and BindingDB, BERT4DTI is competitive, achieving the best ROC-AUC and PR-AUC on BIOSNAP and the highest sensitivity on all three benchmarks. An ablation on DAVIS shows that mutual attention improves PR-AUC and specificity. With 125M trainable parameters compared with 353M for full BERT fine-tuning, BERT4DTI provides a favourable performance–parameter trade-off for sequence-based DTI screening, while leaving runtime profiling, calibration and leakage-audited validation for future work. Keywords: drug–target interaction, pretrained language models, ChemBERTa, ProtBERT, mutual attention, partial fine-tuning, virtual screening †c-license: by 1. Introduction The therapeutic action of a drug depends strongly on its interactions with biological targets, usually proteins. Identifying such drug–target interactions (DTIs) is therefore central to drug discovery, off-target analysis and drug repurposing (11, 15). Experimental screening remains costly and time-consuming, making in silico DTI prediction useful for prioritizing plausible compound–protein pairs before laboratory validation. DTI prediction remains challenging. Known interactions form a small and biased subset of the chemical–protein space, and many benchmarks construct negative pairs from unknown interactions, introducing label uncertainty (3, 4). Moreover, drugs and protein targets are heterogeneous modalities with distinct vocabularies, lengths and biological meanings. A model must therefore capture interaction-relevant information using sequence- or graph-based representations. Graph-based approaches model structural relationships (5), whereas others learn directly from drug and protein sequences (6). Sequence-based deep learning methods avoid manual descriptor engineering. DeepDTA (7) uses CNNs over SMILES and amino-acid sequences, MolTrans (10) models substructure interactions with Transformers, and attention-based models such as ICAN (13) and HyperAttentionDTI (18) highlight the value of cross-sequence alignment. In parallel, pretrained language models became strong representation learners: ChemBERTa (1) learns from large SMILES corpora, while ProtBERT from ProtTrans (2) learns from protein sequences. Kang et al. (12) combined ChemBERTa and ProtBERT for DTI prediction, and DLM-DTI (14) explored more economical adaptation through hint-based learning. Related affinity-oriented models such as OdinDTA (16) and ArkDTA (8) further show the interest of attention mechanisms, but they mainly target continuous affinity or interpretability objectives. BERT4DTI focuses instead on binary sequence-based screening under common DTI benchmarks. Pretrained encoders improve representation quality but create a practical trade-off. Full fine-tuning of large chemical and protein encoders requires many trainable parameters, whereas frozen encoders may not adapt sufficiently to DTI prediction. In addition, late concatenation of independently encoded drug and protein embeddings does not explicitly model which tokens become relevant in the paired sequence. This motivates a model that combines restricted fine-tuning with explicit cross-modal interaction learning. We propose BERT4DTI, a sequence-based DTI classifier built around partially fine-tuned ChemBERTa and ProtBERT encoders, bidirectional mutual attention, and a convolutional classification head. Our contributions are: (i) a DTI architecture that aligns drug and protein token representations through mutual attention; (i) an evaluation on BIOSNAP, DAVIS and BindingDB with a metric-level interpretation that separates sensitivity gains from PR-AUC and specificity trade-offs, together with a component ablation; and (i) a trainable-parameter analysis and a cautious statement of the remaining limits related to runtime, calibration, SMOTE/merged-data evidence and leakage-aware validation. 2. The BERT4DTI Method 2.1. Problem Definition and Inputs DTI prediction is a binary classification problem. Each example is a pair (d,p)(d,p) where d is a drug as a SMILES sequence and p a protein as an amino-acid sequence. The model learns F:×→[0,1]F:D×P→[0,1], where F⁡(d,p)F(d,p) estimates the probability of interaction; a pair is positive when this probability exceeds a validation-selected threshold. As shown in Fig. 1, the architecture has four stages: two pretrained encoders produce contextual token representations; both modalities are projected to a common dimension; a bidirectional mutual-attention block produces interaction-conditioned representations; and independent 1D CNN blocks followed by an MLP produce the prediction. This binary setting follows the standard benchmark protocol used by MolTrans and pretrained-BERT baselines, enabling direct comparison with prior sequence-based DTI studies. Figure 1. Overview of BERT4DTI. Contextual drug and protein token embeddings are mutually aligned before convolutional classification.Block diagram: SMILES tokens enter ChemBERTa and amino-acid tokens enter ProtBERT; both feed a mutual-attention block; the resulting drug and protein interaction embeddings feed two 1D CNN branches whose outputs are concatenated and classified by a feedforward network into a binary DTI class. 2.2. Chemical and Protein Encoders Drugs are represented by canonical SMILES and encoded with ChemBERTa, a RoBERTa-style model pretrained with masked language modelling on ≈ 10M SMILES strings (1):six Transformer layers, twelve heads, hidden size 768, ≈ 83.4M parameters. Proteins are encoded with ProtBERT from ProtTrans (2); the full model has 30 layers and ≈ 420M parameters. In our configuration ProtBERT is reduced to 18 retained layers (≈ 268M parameters before freezing) to control the large memory cost of long protein sequences. Partial fine-tuning. Let θd _d and θp _p denote ChemBERTa and the retained ProtBERT parameters, partitioned into frozen and trainable parts θd=(θdfr,θdtr) _d=( _d^fr, _d^tr), θp=(θpfr,θptr) _p=( _p^fr, _p^tr). Only the final two layers of each retained encoder are updated, i.e. ∇θdfrℒ=0 _ _d^frL=0 and ∇θpfrℒ=0 _ _p^frL=0. This preserves most pretrained parameters while allowing limited task adaptation; it is intended to reduce trainable size, not to claim runtime gains without measured times. Encoder outputs are linearly projected and normalized to a shared dimension ded_e: (1) D=LN⁡(EdWd+bd),P=LN⁡(EpWp+bp),D=LN(E_dW_d+b_d), P=LN(E_pW_p+b_p), with D∈ℝLd×deD ^L_d× d_e and P∈ℝLp×deP ^L_p× d_e. 2.3. Bidirectional Mutual Attention Late concatenation captures global information but does not ask which drug tokens matter given a protein, or vice versa. BERT4DTI uses bidirectional mutual attention. With h heads, head i uses projections QiD=DWiQ,D,KiD=DWiK,D,ViD=DWiV,DQ^D_i=DW^Q,D_i,\,K^D_i=DW^K,D_i,\,V^D_i=DW^V,D_i for the drug and QiP,KiP,ViPQ^P_i,K^P_i,V^P_i analogously for the protein. The drug-conditioned-by-protein and protein-conditioned-by-drug outputs are (2) HiD H^D_i =softmax⁡(QiD(KiP)⊤dh)ViP, =softmax\! ( Q^D_i(K^P_i) d_h )V^P_i, HiP H^P_i =softmax⁡(QiP(KiD)⊤dh)ViD, =softmax\! ( Q^P_i(K^D_i) d_h )V^D_i, with head dimension dhd_h. Heads are concatenated and projected to HD,HPH_D,H_P. We use de=384d_e=384, h=4h=4 and dh=128d_h=128 (≈ 1.6M attention parameters). Contextual and attention vectors are fused by an equal-weight residual combination: (3) Df=0.5HD+0.5D,Pf=0.5HP+0.5P.D_f=0.5\,H_D+0.5\,D, P_f=0.5\,H_P+0.5\,P. This fixed fusion keeps pretrained information while injecting cross-sequence context; alternative weights or a learnable gate are not evaluated here and are noted as a limitation. 2.4. Convolutional Classification and Objective Two independent 1D CNN branches process DfD_f and PfP_f. Each branch applies two 1D convolutions (32 channels, kernel size 5), each followed by ReLU, max pooling and dropout. The drug and protein feature vectors are concatenated into a 4960-dimensional vector and passed to a three-layer MLP (→14960\!→\!256\!→\!32\!→\!1) with ReLU, dropout 0.10.1 and a final sigmoid returning y^=F⁡(d,p) y=F(d,p). The reported configuration uses SmoothL1Loss (9) for robustness to large residuals; because DTI is a classification task with class imbalance, results are interpreted with ROC-AUC, PR-AUC, sensitivity and specificity rather than the loss alone. 3. Experiments 3.1. Datasets, Protocol and Metrics We use BIOSNAP, DAVIS and BindingDB with the benchmark partitions reported for MolTrans and the pretrained BERT baseline (10, 12). BIOSNAP is built from known interactions balanced with sampled negatives; for DAVIS and BindingDB, pairs with Kd<30K_d<30 are treated as positive, leaving strongly imbalanced test sets. BIOSNAP has 4,510 drugs / 2,181 proteins, DAVIS 68 / 379, and BindingDB 10,665 / 1,413. The pair counts used for the main comparison are the original benchmark partitions: BIOSNAP 9619/9619 training, 1374/1374 validation and 2748/2748 test positives/negatives; DAVIS 1043/1043, 160/2846 and 303/5708; BindingDB 6334/6334, 927/5717 and 1905/11384. Drug sequences are truncated or padded to 100 tokens and proteins to 545. These thresholds were selected from the observed length distributions in the experimental report and cover approximately 95% of the benchmark sequences; larger protein cutoffs were tested during development but reduced feasible batch sizes because of GPU memory. Training uses Adam (learning rate 5×10−55× 10^-5), batch size 36, dropout 0.10.1 and up to 35 epochs, with PyTorch, PyTorch Lightning and Hugging Face Transformers on four NVIDIA P100 GPUs. BERT4DTI results are averaged over five runs; per-run model selection uses validation ROC-AUC. The hyperparameter search considered learning rates 10−6,10−5,5×10−5,10−4\10^-6,10^-5,5× 10^-5,10^-4\, dropout values 0.1,0.2,0.3\0.1,0.2,0.3\ and batch sizes 32,36\32,36\. We kept the fixed benchmark splits to remain comparable with published baselines. We report ROC-AUC, PR-AUC, sensitivity and specificity; PR-AUC and specificity are essential context for the imbalanced DAVIS/BindingDB settings, so a sensitivity gain is not read as uniform superiority. For additional operating-point context, precision, F1 and MCC are reconstructed for BERT4DTI from the reported test distributions and mean sensitivity/specificity: 0.823, 0.851 and 0.694 on BIOSNAP; 0.217, 0.349 and 0.389 on DAVIS; and 0.416, 0.572 and 0.527 on BindingDB. These reconstructed values should be treated as approximate because they are derived from aggregate rates rather than from per-seed confusion matrices. We report trainable-parameter counts. Table 1. Performance comparison on BIOSNAP, DAVIS and BindingDB. Best value in each dataset/metric group is in bold. Baseline values are taken from the corresponding benchmark reports. Dataset Method ROC-AUC PR-AUC Sensitivity Specificity BIOSNAP MolTrans 0.895±0.0020.895± 0.002 0.901±0.0040.901± 0.004 0.775±0.0320.775± 0.032 0.851±0.0140.851± 0.014 DeepDTA 0.876±0.0050.876± 0.005 0.883±0.0060.883± 0.006 0.781±0.0150.781± 0.015 0.824±0.0120.824± 0.012 Fine-tuned BERT 0.914±0.0060.914± 0.006 0.900±0.0070.900± 0.007 0.862±0.0250.862± 0.025 0.847±0.0070.847± 0.007 DLM-DTI 0.914±0.0030.914± 0.003 0.914±0.0060.914± 0.006 0.848±0.0160.848± 0.016 0.844±0.0240.844± 0.024 ICAN 0.8710.871 0.8860.886 0.7990.799 0.7860.786 BERT4DTI 0.920±0.0020.920± 0.002 0.924±0.0030.924± 0.003 0.882±0.0130.882± 0.013 0.810±0.0200.810± 0.020 DAVIS MolTrans 0.907±0.0020.907± 0.002 0.404±0.0160.404± 0.016 0.800±0.0220.800± 0.022 0.876±0.0130.876± 0.013 DeepDTA 0.880±0.0070.880± 0.007 0.302±0.0440.302± 0.044 0.764±0.0450.764± 0.045 0.865±0.0200.865± 0.020 Fine-tuned BERT 0.920±0.0020.920± 0.002 0.395±0.0070.395± 0.007 0.824±0.0260.824± 0.026 0.889±0.0150.889± 0.015 DLM-DTI 0.895±0.0030.895± 0.003 0.373±0.0170.373± 0.017 0.833±0.0440.833± 0.044 0.802±0.0700.802± 0.070 ICAN 0.9030.903 0.3720.372 0.8840.884 0.7660.766 BERT4DTI 0.920±0.0050.920± 0.005 0.370±0.0220.370± 0.022 0.891±0.0160.891± 0.016 0.829±0.0180.829± 0.018 BindingDB MolTrans 0.914±0.0010.914± 0.001 0.622±0.0070.622± 0.007 0.797±0.0050.797± 0.005 0.896±0.0070.896± 0.007 DeepDTA 0.913±0.0030.913± 0.003 0.622±0.0120.622± 0.012 0.780±0.0350.780± 0.035 0.915±0.0160.915± 0.016 Fine-tuned BERT 0.922±0.0010.922± 0.001 0.623±0.0100.623± 0.010 0.814±0.0250.814± 0.025 0.916±0.0160.916± 0.016 DLM-DTI 0.912±0.0040.912± 0.004 0.643±0.0060.643± 0.006 0.888±0.0140.888± 0.014 0.793±0.0150.793± 0.015 ICAN 0.9000.900 0.6040.604 0.8460.846 0.8150.815 BERT4DTI 0.912±0.0030.912± 0.003 0.630±0.0040.630± 0.004 0.912±0.0050.912± 0.005 0.786±0.0150.786± 0.015 3.2. Main Benchmark Results Table 1 reports the comparison. On BIOSNAP, BERT4DTI obtains the best ROC-AUC (0.9200.920) and PR-AUC (0.9240.924) and the highest sensitivity, while MolTrans and the BERT baselines reach higher specificity: the model retrieves positive pairs well on a balanced test set but admits more false positives. On DAVIS, BERT4DTI matches the best ROC-AUC (0.9200.920) and gives the highest sensitivity (0.8910.891), whereas MolTrans has the best PR-AUC and fine-tuned BERT the best specificity; the strongly imbalanced test distribution makes PR-AUC and specificity essential context. On BindingDB, BERT4DTI again gives the highest sensitivity (0.9120.912), with full BERT fine-tuning best on ROC-AUC/specificity and DLM-DTI best on PR-AUC. BERT4DTI is therefore best viewed as a competitive model with a pronounced positive-retrieval profile rather than a uniformly superior method. 3.3. Ablation and Parameter Efficiency We ablate the CNN and mutual-attention modules individually and jointly on DAVIS (Table 2). Removing mutual attention lowers ROC-AUC from 0.920.92 to 0.890.89, PR-AUC from 0.370.37 to 0.290.29 and specificity from 0.830.83 to 0.750.75, indicating that attention contributes notably to PR-AUC under this setting. Removing both modules recovers PR-AUC close to the full model but yields a larger trainable size (183M vs. 125M). Table 3 compares trainable parameters: BERT4DTI uses 125M, far fewer than 353M for full BERT fine-tuning, though more than MolTrans (62.8M) and DLM-DTI (86.7M); it occupies an intermediate position, retaining more adaptation capacity than lighter alternatives while avoiding the full size of the paired encoders. A parameter count alone cannot establish wall-clock efficiency, which depends on sequence length, frozen-layer computation, attention memory, batch size and hardware. Accordingly, we report BERT4DTI as parameter-efficient relative to full BERT fine-tuning. Additional analyses in the experimental study (reported descriptively, not as controlled comparisons) include a cold-split evaluation holding out 20% of drugs or proteins (best held-out ROC-AUC 0.9010.901 for unseen drugs), a limited-data setting where BERT4DTI keeps the highest ROC-AUC down to 0.7700.770 with 95% of training data removed, a BIOSNAP protein-family stratification (strongest on GPCRs and ion channels, PR-AUC 0.990.99), and an exploratory positive-pair study on 13 CYP interactions (average score 0.7270.727 vs. 0.4790.479 for MolTrans). These analyses contain no negative pairs or similarity-clustered splits and do not support claims of general external generalization. Table 2. Ablation on DAVIS. Configuration ROC-AUC PR-AUC Sens. Spec. BERT4DTI 0.920.92 0.370.37 0.890.89 0.830.83 w/o CNN 0.910.91 0.360.36 0.910.91 0.760.76 w/o mutual attention 0.890.89 0.290.29 0.900.90 0.750.75 w/o CNN & attention 0.910.91 0.370.37 0.900.90 0.800.80 Table 3. Reported trainable parameter counts. Model Trainable parameters (M) MolTrans 62.8 DLM-DTI 86.7 BERT4DTI 125.0 BERT4DTI w/o CNN & attention 183.0 Full BERT fine-tuning 353.0 3.4. Discussion Table 1 reports the benchmark comparison on BIOSNAP, DAVIS and BindingDB. On BIOSNAP, BERT4DTI achieves the best ROC-AUC (0.9200.920) and PR-AUC (0.9240.924), together with the highest sensitivity. This indicates that the combination of pretrained sequence encoders and mutual attention is effective on a balanced benchmark where positive and negative pairs are equally represented. However, its specificity remains lower than that of MolTrans and fine-tuned BERT, showing that BERT4DTI favours positive-pair retrieval over conservative rejection of negatives. This positive-retrieval profile is also observed on DAVIS and BindingDB. On DAVIS, BERT4DTI matches the best ROC-AUC and obtains the highest sensitivity, whereas MolTrans obtains the best PR-AUC and fine-tuned BERT the best specificity. On BindingDB, BERT4DTI again achieves the highest sensitivity, while full BERT fine-tuning gives the best ROC-AUC and specificity and DLM-DTI gives the best PR-AUC. These results suggest that BERT4DTI is well suited to early screening scenarios where recovering potential interactions is important, but they also show that it should not be interpreted as uniformly superior across all operating criteria. In imbalanced screening, PR-AUC and specificity remain important because false positives increase the experimental follow-up burden. Table 2 supports the contribution of the mutual-attention module. Removing mutual attention decreases ROC-AUC from 0.920.92 to 0.890.89, PR-AUC from 0.370.37 to 0.290.29, and specificity from 0.830.83 to 0.750.75 on DAVIS. This suggests that cross-sequence token alignment provides information beyond independent pretrained encodings followed by late fusion. Removing the CNN branch has a smaller effect on ROC-AUC and PR-AUC but still reduces specificity, indicating that local convolutional aggregation remains useful after mutual attention. The parameter analysis in Table 3 positions BERT4DTI between lighter sequence models and full BERT fine-tuning. With 125M trainable parameters, BERT4DTI is substantially smaller than full BERT fine-tuning (353M), while preserving more adaptation capacity than MolTrans and DLM-DTI. Therefore, the efficiency claim concerns trainable model size, not measured runtime. Runtime, peak memory and inference latency depend on sequence length, frozen-layer computation, batch size and hardware, and should be evaluated in a dedicated deployment study. Additional analyses, including held-out drug/protein settings, scarce-data experiments, protein-family stratification and the CYP positive-pair case study, provide complementary evidence that BERT4DTI preserves useful ranking behaviour beyond the main random-pair benchmark. These analyses remain descriptive because they do not replace similarity-clustered splits, duplicate-pair auditing, probability calibration or an external benchmark with both positive and negative evidence. Therfore, this comparison relies on the original benchmark partitions, and merged-dataset or SMOTE-based variants are not used as central evidence. 4. Conclusion BERT4DTI combines ChemBERTa and ProtBERT with partial fine-tuning, bidirectional mutual attention and convolutional classification for sequence-based DTI prediction. The model obtains strong results across three benchmarks, especially in terms of positive-pair retrieval, and the ablation study supports the value of mutual attention for cross-sequence interaction modelling. At the same time, the results reveal a clear trade-off: on imbalanced datasets, competing methods may obtain better PR-AUC or specificity. BERT4DTI should therefore be viewed as a competitive screening model with reduced trainable size relative to full BERT fine-tuning, rather than as a universally dominant predictor. Future work will extend BERT4DTI toward knowledge-grounded DTI prediction by integrating biomedical knowledge graphs and retrieval-augmented language models to support explainable, evidence-based interaction prioritization. Acknowledgment This research was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) through a Discovery Grant (RGPIN-2025-06792) and internal funding from TÉLUQ University (FAR3 program). References (1) Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. 2020. ChemBERTa: large-scale self-supervised pretraining for molecular property prediction. arXiv preprint arXiv:2010.09885. (2) Ahmed Elnaggar, et al. 2022. ProtTrans: toward understanding the language of life through self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 10, 7112–7127. (3) Ali Ezzat, Min Wu, Xiao-Li Li, and Chee-Keong Kwoh. 2019. Computational prediction of drug–target interactions using chemogenomic approaches: an empirical survey. Briefings in Bioinformatics 20, 4, 1337–1357. (4) Vitor Magalhaes Silva, Khadidja Henni, Neila Mezghani, and Youcef Abdelliche. 2026. A matrix factorization and generative modeling framework for drug–target interaction prediction. In Proceedings of the 19th International Joint Conference on Biomedical Engineering Systems and Technologies – Volume 2: BIOINFORMATICS. SciTePress, 688–695. https://doi.org/10.5220/0014484300004070. (5) Khadidja Henni, Leila Hamdad, Neila Mezghani, NourElhouda Benazzoug, and Lina Abou-Abbas. 2025. Integrating structural graphs and cross-attention for drug–target interaction prediction. In 2025 IEEE/ACS 22nd International Conference on Computer Systems and Applications (AICCSA). IEEE, 1–7. https://doi.org/10.1109/AICCSA66935.2025.11315467. (6) Lina Abou-Abbas and Khadidja Henni. 2026. Protein and ligand novelty in drug–target interaction prediction: a dual-encoder fusion strategy for more interpretable and generalizable modeling. BMC Bioinformatics 27, 1, Article 121. https://doi.org/10.1186/s12859-026-06457-5. (7) Hakime Öztürk, Arzucan Özgür, and Elif Ozkirimli. 2018. DeepDTA: deep drug–target binding affinity prediction. Bioinformatics 34, 17, i821–i829. (8) Mogan Gim, et al. 2023. ArkDTA: attention regularization guided by non-covalent interactions for explainable drug–target binding affinity prediction. Bioinformatics 39, Supplement_1, i448–i457. (9) Ross Girshick. 2015. Fast R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, 1440–1448. (10) Kexin Huang, Cao Xiao, Lucas M. Glass, and Jimeng Sun. 2021. MolTrans: molecular interaction transformer for drug–target interaction prediction. Bioinformatics 37, 6, 830–836. (11) James P. Hughes, Stephen Rees, S. Barrett Kalindjian, and Karen L. Philpott. 2011. Principles of early drug discovery. British Journal of Pharmacology 162, 6, 1239–1249. (12) Hyeunseok Kang, et al. 2022. Fine-tuning of BERT model to accurately predict drug–target interactions. Pharmaceutics 14, 8, 1710. (13) Hiroyuki Kurata and Sho Tsukiyama. 2022. ICAN: interpretable cross-attention network for identifying drug and target protein interactions. PLOS ONE 17, 10, e0276609. (14) Jonghyun Lee, Dae Won Jun, Ildae Song, and Yun Kim. 2024. DLM-DTI: a dual language model for the prediction of drug–target interaction with hint-based learning. Journal of Cheminformatics 16, 14. (15) Kanica Sachdev and Manoj Kumar Gupta. 2019. A comprehensive review of feature based methods for drug target interaction prediction. Journal of Biomedical Informatics 93, 103159. (16) Shilong Xu and Ruili Wang. 2023. OdinDTA: combining mutual attention and pre-training for drug-target affinity prediction. In 2023 IEEE 35th International Conference on Tools with Artificial Intelligence (ICTAI), 680–687. (17) Zhihui Yang, Juan Liu, Feng Yang, Xiaolei Zhang, Qiang Zhang, Xuekai Zhu, and Peng Jiang. 2024. Advancing drug-target interaction prediction with BERT and subsequence embedding. Computational Biology and Chemistry, 108058. (18) Qichang Zhao, Haochen Zhao, Kai Zheng, and Jianxin Wang. 2022. HyperAttentionDTI: improving drug–protein interaction prediction by sequence-based deep learning with attention mechanism. Bioinformatics 38, 3, 655–662. GenAI Usage Disclosure Generative AI tools were used only for language polishing and formatting assistance. All scientific contributions, experimental results, interpretations and conclusions are the responsibility of the authors and were checked against the original experimental report.

Instant research discovery