Where Black-box Drug-Target Interaction Prediction Models Look: Cross-Method Explainability

A new interpretability audit of the BridgeDPI model reveals how black-box drug-target interaction predictors use sequence, fingerprint, and graph features, exposing modality dominance and dataset-specific artifacts that could guide downstream validation in computational drug discovery. The study, submitted to arXiv in 2026, examines the BridgeDPI architecture across three datasets — Gao, Human, and C.elegans — using six gradient-based attribution methods combined with feature-wise occlusion ablation and a strict intersection consensus to reduce single-explainer bias [1][2]. The methods include integrated gradients, saliency, layer-wise relevance propagation, SmoothGrad, and SmoothGrad-IG [2]. Researchers summarized sensitivity and signed effects at raw inputs, at the bridge similarity scaffold, and through the graph convolution, including edge-level sensitivities and targeted edge removals [2]. The findings indicate that explainability is most informative when treated as model criticism, revealing padding and special-token artifacts, dataset-dependent cooperative versus suppressive effects across layers, and chemistry-consistent fragment and composition motifs where methods agree [2]. The authors caution that these analyses do not substitute for structural or experimental ground truth, but they can provide testable hypotheses for downstream validation in computational drug discovery pipelines [2]. Deep learning architectures, which underpin models like BridgeDPI, have been applied to bioinformatics and drug design, producing results that in some cases surpass human expert performance [4]. The arXiv preprint server, where the study appears, hosts over two million articles and receives approximately 24,000 submissions per month as of late 2024, serving as a primary dissemination channel for machine learning research before peer review [9]. The broader field of artificial intelligence has seen funding and interest increase substantially since 2012, when graphics processing units began accelerating neural networks, and growth accelerated further after 2017 with the transformer architecture [3]. Applying modern explainable AI to contemporary DTI and DTA models remains an early pass over the rich structure implicit in trained weights and data, yet even this first layer of scrutiny helps researchers relate predictions to drug- and target-side representations and prioritize external validation [2].