Topological deep learning for drug–target interaction, virtual screening, and docking scoring: a practical, benchmark-driven review

Abstract Artificial intelligence is now central to computational drug discovery, yet performance in core tasks—drug–target interaction (DTI) prediction, virtual screening (VS), and docking scoring—is still limited by the multiscale geometric nature of molecular recognition and by evaluation pitfalls such as dataset bias and leakage. Topological deep learning (TDL) offers a complementary route to encode global and multiscale structure from ligands, binding pockets, surfaces, and protein–ligand complexes via persistent homology and related constructions. This review provides a practical, task-driven synthesis of TDL methods for DTI/VS/docking scoring, with an emphasis on design choices that determine real-world utility: (i) data modality (ligand, pocket, or complex/pose) under controllable uncertainty, (ii) topological objects and filtration families (distance/alpha versus physicochemical or interaction-field filtrations), and (iii) vectorizations and integration patterns (persistent homology-as-features, hybrid geometric deep learning, and emerging end-to-end approaches). Distinct from prior surveys, we present a decision-oriented taxonomy and a benchmark-driven evaluation playbook that specifies minimum standards for splits (scaffold, temporal, and target-wise/cluster), metrics (including early-recognition metrics for VS), baselines, and ablations to isolate the topological contribution. To support reproducibility, we provide a reporting checklist and curated summary tables (methods matrix and benchmark recommendations) that map tasks to recommended protocols and common failure modes.