A new review maps the fast-moving landscape of AI-driven drug-target interaction (DTI) prediction, revealing how the field has progressed in areas such as hand-crafted molecular features, deep learning, graph-based models, pretraining, and multimodal fusion.
The authors organised the review around data foundations, model architectures, benchmark resources, and evaluation practices that determine if predictions are trustworthy. Additionally, they argue that future progress is dependent on improving data quality, cross-dataset robustness, mechanistic interpretability, and links to experimental testing.
DTI prediction can help researchers explain mechanisms of action, identify candidate targets, reposition existing medicines, and prioritise compounds before laboratory work. However, the field still faces uneven structural coverage, scarce high-quality negative examples, inconsistent activity labels, and strong bias towards well-studied drugs and targets.
Models that perform well on familiar benchmark datasets can lose accuracy when asked to evaluate a new compound, a new protein family, or data collected under different experimental conditions. Multimodal systems can add chemical, structural, functional and network context, as well as introducing alignment problems, higher computational demands, and new sources of noise. These challenges mean further investigation is required in standardised dataset construction, data quality control, realistic evaluation and experimentally testable prediction frameworks.
The review, published online on July 20, 2026, in the Medical Journal of Peking Union Medical College Hospital, was prepared by researchers from the School of Artificial Intelligence at Beijing University of Posts and Telecommunications and the Shandong Computer Science Centre, also known as the National Supercomputing Centre in Jinan.
The team assesses the data foundations, benchmark resources, modelling strategies, and translational barriers shaping AI-driven DTI prediction, and outlines a path from algorithm-centred comparison toward frameworks that support real-world drug discovery workflows, pharmacological research and preclinical validation.
The review starts with three core data types: drug representations, target representations, and drug-target association labels. Drugs can be encoded via the simplified molecular input line entry system (SMILES), molecular fingerprints or molecular graphs, and targets can be described through amino-acid sequences, three-dimensional structures, and functional annotations.
The authors then trace three stages of technical development. Traditional machine learning approaches include similarity-based methods, matrix factorisation, network-based models, engineered-feature models, and hybrid models combining multiple strategies; deep learning automatically extracts local sequence patterns, molecular topology, long-range dependencies, semantic representations learned via large-scale pretraining, and enhanced signals from external knowledge or higher-order structural priors; and multimodal models integrate multi-source information, including drug, protein, disease, side-effect, pertubation phenotype, and knowledge-network data.
Graph neural networks (GNNs), attention mechanisms, Transformer architectures and pretrained representations have broadened what models can capture, particularly when molecular structure and biological context must be considered together.
The review also compares benchmark resources and demonstrates that each serves a different task, including binary interaction classification, quantitative affinity prediction, and structure-based virtual screening. However, performance cannot be compared fairly without consistent datasets, split protocols, and metrics. Random splits can exaggerate success by placing highly similar examples in training and test sets. More realistic cross-distribution, target-family and cold-start evaluations are required to reveal whether a model can generalise beyond familiar chemical and biological space.
The authors stated that the next advance should not be measured only by a higher benchmark score. They believe useful DTI systems must show where their evidence comes from, remain reliable when data distributions change and generate hypotheses that researchers can test.
In their view, predictions are more valuable when they narrow candidate lists, suggest plausible binding mechanisms, and guide molecular docking, structural modelling, and laboratory experiments. The authors also stressed that adding more data types is not automatically beneficial: each modality must be relevant, well aligned, and strong enough to enhance biological reasoning rather than increase model complexity.
The framework also has the potential to support faster candidate screening, drug repurposing, lead optimisation, and potential risk assessment; however, the review treats AI as a decision aid rather than a replacement for experiments.
A credible workflow would connect target biology, computational ranking, structural modelling, and staged validation via binding assays, cellular studies, and animal research. The authors also predict specialised biomedical foundation models and intelligent agents capable of integrating literature, databases, candidate generation, and validation design in one traceable workflow.
For pharmaceutical research, the central implication is practical: AI-driven DTI prediction will have its greatest impact when it shortens the path from large search spaces to a smaller set of biologically justified, experimentally verifiable options.
