Predicting protein–protein interaction (PPI) patterns from primary sequences constitutes a complex pattern recognition task that demands advanced and explainable deep learning frameworks, as PPIs underpin critical cellular and pathological mechanisms.We propose PEPpip, a deep learning framework for binary sequence-based PPI prediction that also derives interaction maps via explainable-AI (XAI) methods. PEPpip encodes sequence pairs into 3D image-like feature maps using three pre-trained protein language models(PLMs). To learn PPI patterns, it applies two vision-based classifiers—ResNet and Vision Transformer( ViT)—whose predictions are fused through a novel Post-hoc Inference Combiner(PIC) leveraging their complementary strengths. In addition, PEPpip uniquely generates PPI maps by combining XAI-based interaction maps from integrated gradients and ViT attention, further refined through a modular noise-cleaning pipeline. We introduce a novel residue-residue interaction-aware, class-discriminative attention map(RiaCdAm) mechanism that embeds biological priors into ViT attention, improving interpretability and contact relevance. PEPpip achieves state-of-the-art AUPR (Area Under Precision-Recall curve) scores across species—91.1%(Mouse), 89.1%(Fly), 87.7%(Worm), 68.1%(Yeast), and 71.9%(E.coli)—outperforming existing methods by 2–3%. It also performs strongly on Human-PPI tasks. While its sequence-driven, XAI-based PPI map approach is not directly comparable to structure-based learners, PEPpip surpasses competing sequence-based classifiers in interaction map estimation, exceeding the next-best method by over 5% in top-L/10 and top-L/5 categories (L: shorter protein-length in the pair). PEPpip provides a unified and interpretable framework for binary PPI prediction with coarse-grained interaction map estimation, offering a practical first-pass screening tool for large-scale interactome analysis and paving the way for future XAI-integrated models.
Ghosh et al. (Thu,) studied this question.