AI-Based Drug-Target Interaction Prediction

Abstract

Drug discovery is a decade-long, USD 2.6 billion endeavour with a 90% clinical failure rate, driven largely by inadequate understanding of drug-target interactions (DTIs). Identifying which proteins a candidate drug binds to -- and with what affinity, selectivity, and mechanism -- is the foundation of both efficacy and safety prediction. Experimental DTI screening (high-throughput biochemical assays, surface plasmon resonance, isothermal titration calorimetry) is expensive (USD 5-50 per compound-target pair) and slow (months for a comprehensive target panel). AI-based DTI prediction models learn from known interactions to predict novel drug-target pairs, potentially reducing the experimental search space by 100-1,000 fold. Yet the DTI prediction landscape is fragmented: molecular fingerprint methods, sequence-based deep learning, structure-based docking surrogates, graph neural networks on molecular graphs, and knowledge graph embedding models each capture different aspects of the drug-target interaction. We present the Drug-Target Interaction Prediction Assessment Framework (DTIPAF), evaluating five AI architectures -- chemical fingerprint classifiers, sequence-based convolutional models, 3D structure-aware graph networks, heterogeneous knowledge graph embeddings, and multi-modal transformer fusion models -- across four DTI prediction tasks (binary interaction classification, binding affinity regression, selectivity profiling, and off-target prediction). Our DTI Prediction Performance Score (DPPS) measures prediction accuracy, generalisation to novel chemotypes, computational throughput, mechanistic interpretability, and prospective validation rate. Multi-modal transformer fusion achieves the highest DPPS (0.926) by integrating molecular graphs, protein sequences, and 3D binding pocket features, while knowledge graph embeddings achieve the highest off-target prediction accuracy (0.940) by leveraging the polypharmacology network structure