AI-Based Drug–Target Interaction Prediction: A Systematic Review of Machine Learning and Deep Learning Approaches

AI-Based Drug–Target Interaction Prediction

Abstract

Predicting which small molecules bind which proteins is among the slowest and most expensive steps in early drug discovery, and it is where computational methods have made their most visible progress. The field has moved from similarity kernels and matrix factorisation towards models that learn directly from molecular graphs, SMILES strings, protein sequences and, increasingly, predicted three-dimensional structures. This review surveys that shift. Following PRISMA 2020 guidance we screened 4,812 records published between 2008 and 2025 and retained 68 studies. The methods identified are grouped into five families: similarity and kernel models, matrix factorisation and feature-based classifiers, sequence-based deep networks, graph neural networks together with attention and transformer architectures, and heterogeneous-network or structure-aware models. For each family we compare representations, benchmark datasets, evaluation protocols and reported performance, and we state what each comparison contributes. Reported error on the standard affinity benchmarks has fallen by roughly half since the earliest kernel baselines, but almost all of that improvement has been measured under random splits that allow near-duplicate drugs and proteins to appear in both training and test data, and performance under cold-start and cross-domain conditions remains substantially weaker. Our central observation is therefore that architectural progress has outpaced evaluation practice. We identify ten open problems, among them leakage under random splits, weak cold-start generalisation, the absence of reliable negative samples, narrow coverage of protein families, limited interpretability and the under-use of predicted structures, and set out the directions most likely to address them: protein and chemical language models, structure-aware architectures built on proteome-wide predicted structures, uncertainty-aware prediction coupled to active experimental design, and standardised benchmarks with hidden test sets and prospective validation.