Source-linked AI summary
Interpretable bilinear attention network with domain adaptation improves drug-target prediction
Peizhen Bai, Filip Miljković, Bino John, Haiping Lu
TL;DR
Drug-target prediction needs models that capture local drug-protein interactions and generalize across data distributions. DrugBAN combines bilinear attention over molecular graphs and protein sequences with conditional domain adaptation, achieving improved performance in in-domain and cross-domain settings while providing interpretable interaction insights.
Problem
Drug-target prediction lacks explicit local-interaction modeling and robust generalization to novel pairs from different distributions, despite its importance for drug discovery.
Method
DrugBAN uses bilinear attention to learn pairwise local interactions from drug molecular graphs and protein sequences, with conditional adversarial domain adaptation for cross-domain prediction.
Results
DrugBAN consistently achieves improved DTI prediction performance in both in-domain and cross-domain settings and provides biological insights through attention-weight mappings.
Takeaways & Limitations
Bilinear attention weights can support interpretation by mapping predicted interaction contributions to protein subsequences and drug compound atoms.
Takeaways & Limitations
The study uses 1D protein sequences and 2D molecular graphs rather than modeling highly accurate 3D protein structures.
Abstract
from arXiv · showhide
Predicting drug-target interaction is key for drug discovery. Recent deep learning-based methods show promising performance but two challenges remain: (i) how to explicitly model and learn local interactions between drugs and targets for better prediction and interpretation; (ii) how to generalize prediction performance on novel drug-target pairs from different distribution. In this work, we propose DrugBAN, a deep bilinear attention network (BAN) framework with domain adaptation to explicitly learn pair-wise local interactions between drugs and targets, and adapt on out-of-distribution data. DrugBAN works on drug molecular graphs and target protein sequences to perform prediction, with conditional domain adversarial learning to align learned interaction representations across different distributions for better generalization on novel drug-target pairs. Experiments on three benchmark datasets under both in-domain and cross-domain settings show that DrugBAN achieves the best overall performance against five state-of-the-art baselines. Moreover, visualizing the learned bilinear attention map provides interpretable insights from prediction results.
Introduction
Drug-target interaction prediction supports drug discovery but remains limited by weak local-interaction modeling and poor generalization to out-of-distribution pairs. DrugBAN addresses both challenges with bilinear attention, domain adaptation, and interpretable attention maps.
- Motivation: In silico DTI prediction can reduce experimental search costs by narrowing compound candidates and identifying potential side-effect mechanisms.Traditional in vitro measurements are reliable but costly and time-consuming, limiting their scalability.
- Challenges: Existing methods often learn global drug and protein representations without explicitly modeling interactions between molecular substructures and protein binding sites.This limits direct learning of the local mutual effects that determine DTI.
- Challenges: Cross-domain DTI prediction is difficult because real-world drug-target pairs may be unseen, dissimilar to training pairs, and drawn from different distributions.The target domain may provide only unlabeled data, motivating transferable representation learning from source to target.
- DrugBAN: DrugBAN encodes molecular graphs and protein sequences with GCN and CNN modules, then uses bilinear attention to learn pairwise local interaction representations.A fully connected decoder uses the joint representation for DTI prediction.
- DrugBAN: For cross-domain prediction, DrugBAN applies conditional domain adaptation to align transferable representations across source and target distributions.The framework is designed to improve generalization on out-of-distribution drug-target pairs.
- Interpretability: DrugBAN provides interpretable predictions by visualizing bilinear attention weights that map contributions to drug substructures and protein subsequences.This distinguishes the framework from black-box predictions by exposing pairwise interaction patterns.
Results
DrugBAN evaluates drug-target interaction prediction using molecular graphs and protein sequences, with bilinear attention for local interactions and domain adaptation for cross-domain transfer. Across in-domain, cold-pair, and cross-domain evaluations, it generally outperforms the compared baselines, while its attention maps provide molecular-level interpretability with weaker protein-sequence interpretation.
- Evaluation setup: DrugBAN maps drug molecular graphs and protein sequences to interaction probabilities using GCN/CNN encoders, bilinear attention, and a fully connected classifier.The bilinear attention module produces joint local-interaction representations for prediction.
- Evaluation setup: The study evaluates three public datasets with random, cold-pair, and clustering-based pair splits, reporting AUROC, AUPRC, accuracy, sensitivity, and specificity across five runs.BindingDB and BioSNAP support in-domain and cross-domain evaluation, while Human is assessed with random and cold-pair splits.
- In-domain performance: DrugBAN consistently outperforms five baselines in AUROC, AUPRC, and accuracy on BindingDB and BioSNAP under random in-domain splits.Sensitivity and specificity are also reported as competitive.
- In-domain performance: All models lose performance on the Human cold-pair split, yet DrugBAN retains the best performance among the compared deep learning baselines.The cold-pair split removes associated drugs and proteins from training, reducing overlap with evaluation pairs.
- Cross-domain performance: DrugBAN outperforms MolTrans by 2.9% and 7.4% in AUROC on BioSNAP and BindingDB, respectively, under clustering-based cross-domain evaluation.Overall performance drops because training and test data share less information, but vanilla DrugBAN remains strongest overall among the compared state-of-the-art models.
- Cross-domain performance: DrugBANCDAN improves over vanilla DrugBAN by 4.6% in AUROC and 16.9% in AUPRC on BioSNAP.The reported improvement follows introduction of the conditional domain adaptation module.
- Interpretability: Bilinear attention visualizations correctly highlight several ligand functional groups and some binding-site residues, while protein-sequence interpretability is weaker and sometimes misses specific interactions.The ligand examples include correctly identified interacting groups in 6QL2, 5W8L, and 4N6H; the paper also notes potential use for binding-property and off-target analysis.
Conclusion
DrugBAN is an end-to-end bilinear attention framework that integrates adversarial domain adaptation for drug-target interaction prediction. It improves prediction performance across in-domain and cross-domain settings, while its attention mappings provide biological interpretability; the work uses 1D protein sequences and 2D molecular graphs rather than 3D protein structures.
- DrugBAN integrates CDAN into an end-to-end bilinear attention framework to enhance cross-domain generalization for drug-target interaction prediction.
- DrugBAN consistently improves DTI prediction performance in both in-domain and cross-domain settings, while mapping attention weights to protein subsequences and drug atoms provides biological insights.
- The model uses 1D protein sequences and 2D molecular graphs, excluding highly accurate 3D protein structural information from its modeling scope.
Methods
DrugBAN combines local encoders and bilinear attention to model drug–protein substructure interactions, then uses conditional adversarial domain adaptation to improve cross-domain prediction. Its attention maps expose contributions of drug and protein substructure pairs.
- Feature encoding: DrugBAN uses GCN and CNN encoders to represent local structures in drug molecular graphs and protein sequences.The GCN aggregates neighboring atom features, while the CNN captures local residue patterns from overlapping protein 3-mers.
- Pairwise interaction learning: A bilinear attention module learns pairwise interactions between encoded drug atoms and protein subsequences.The interaction map has dimensions N×M, where N and M denote encoded drug atoms and protein substructures.
- Interpretability: The bilinear interaction map assigns intensities to drug–protein substructure pairs, mapping them to potential molecular substructures and binding sites.Representations are projected into a common feature space before element-wise interaction weighting.
- Pairwise interaction learning: Multi-head bilinear interaction sums multiple interaction maps and adds only one weight vector per additional head because the weight matrices are shared.The paper reports better performance for multi-head interaction than for a single head.
- Prediction: The joint representation is decoded by a fully connected classification layer followed by a sigmoid function for DTI prediction.This decoder operates on the joint representation produced by the interaction module.
- Cross-domain adaptation: DrugBAN embeds conditional adversarial domain adaptation to align source and target distributions when labeled source data and unlabeled target data are available.Adversarial training is described as reducing distribution shift and improving cross-domain generalization.
Data availability
The study uses public datasets and makes the experimental splits and source data available through cited repositories.
- Data availability: Experimental splits are available in the DrugBAN GitHub repository, and all datasets used are from public resources.The sources include BindingDB, BioSNAP, and Human dataset repositories.
Supplementary Material
The supplementary material details cross-domain dataset construction, evaluation settings, missing-data tests, and scalability analyses for DrugBAN. These experiments report strong performance under challenging splits and characterize computational behavior as dataset size and batch size increase.
- Cross-domain evaluation: Clustering-based pair splits construct cross-domain tasks by separating dissimilar drugs and proteins using a minimum distance threshold.Single-linkage clustering with γ=0.5 produced 2,780 drug and 1,693 protein clusters for BindingDB, and 2,387 drug and 1,978 protein clusters for BioSNAP.
- Unseen drugs and targets: DrugBAN achieves the best performance in five of six unseen drug/target settings and remains competitive on unseen targets in BioSNAP.These settings evaluate pairs involving randomly selected unseen drugs or target proteins, with separate validation and test subsets.
- Unseen drugs and targets: A naive unseen-drug split can underestimate cross-domain difficulty because BindingDB contains many highly similar molecules.The authors identify clustering-based splitting as a more challenging strategy that better separates dissimilar samples.
- Missing-data evaluation: DrugBAN achieves the best performance across all tested training-data fractions, with larger improvements on the bigger BindingDB dataset.Models were trained with only 5%, 10%, 20%, or 30% of each dataset and evaluated on the remaining data.
- Scalability: Optimization time and peak GPU memory increase almost linearly with dataset size and batch size, while CPU data loading accelerates with additional workers.For 49,199 BindingDB pairs, optimization takes about two hours; the default batch size of 64 uses 4.63 GB of GPU memory.