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

Facilitating structure-based drug discovery with an artificial intelligence-driven virtual screening platform.

Geometric Deep Learning-Based Drug Design Models for Small-Molecule Drug Discovery.

Deep neural network (DNN)-based in silico models show great promise in predicting the properties and bioactivities of novel compounds, including small molecules. Among traditional approaches, structure-based drug design (SBDD) remains a fundamental approach for drug discovery using molecular docking, scoring functions, and molecular dynamics simulations. However, these approaches are often constrained by limited flexibility, resolution, and generalizability. Geometric deep learning (GDL) offers a transformative alternative by enabling models to learn directly from non-Euclidean molecular representations, such as graphs, point clouds, and meshes, capturing critical 3D spatial relationships inherent to protein-ligand interactions. This review highlights the theoretical underpinnings and practical applications of GDL in small-molecule drug discovery, focusing on tasks including binding affinity prediction, virtual screening, de novo molecule generation, pose prediction, ADMET profiling, and protein flexibility modeling. We explore key GDL architectures, graph neural networks, SE(3)-equivariant networks, 3D convolutional neural networks, point cloud models, and geometric transformers, and assess their performance across various drug discovery benchmarks. The integration of geometry-aware AI models with experimental and computational workflows was also highlighted for its potential to streamline hit-to-lead optimization and advance rational drug design. Despite remarkable progress, the field faces challenges including limited high-quality 3D structural datasets, protein flexibility representation, and the interpretability of deep models. Addressing these issues through hybrid modeling approaches, multi-resolution learning, and self-supervised training could further elevate GDL's impact. Ultimately, GDL stands at the frontier of AI-enhanced pharmaceutical innovation, offering unprecedented precision, efficiency, and insight in the pursuit of next-generation therapeutics.

Overcoming the accuracy–generalization tradeoff in docking and scoring for prospective virtual screening

Native Contact Ratio as a Topological Metric for Machine Learning Based Molecular Docking

Accurate identification of near-native ligand binding poses is a central challenge in structure-based drug design. From a physical point of view, the successful construction of a protein–ligand complex structure is dependent on whether protein and ligand can form enough atomically pairwise interactions that result in a global energy minimum. In this work, we report a machine learning scoring strategy for protein–ligand screening which explicitly considers the Native Contact Ratio (NCR), a topology inspired metric that quantifies the preservation of protein–ligand interfacial contacts as well as interaction energy. This physics-awared supervision strategy provides a simple but efficient gradient field that faithfully reflects the complicated protein energy landscape than conventional 3D coordinate-based objectives. Building on this principle, we present DeepNCR, an energy-informed Transformer framework that encodes approximate Coulombic and dispersive interaction potentials across the protein–ligand binding interface. Furthermore, we introduce a feature pruning step that compresses the interaction tensor from 1470 to 868 dimensions, further improving signal-to-noise ratio and directing model attention toward the interaction motifs critical for binding specificity. The model optimizes topological objectives and at inference drives pose refinement through a differentiable hybrid gradient field integrating predicted NCR and AutoDock Vina energetics. Extensive evaluation on the CASF-2016 benchmark and the 3D-DISCO cross-docking data set demonstrates consistently high performance: a Top-1 docking success rate of 94.7%, a 1% Enrichment Factor of 21.21 in virtual screening, and a Top-1 cross-docking success rate of 34.8%. Mechanistic analysis reveals that NCR-guided optimization enables decoy escaping from local energy minima and drives the recovery of disrupted native interactions, confirming that NCR captures the physical determinants of binding rather than mere geometric proximity.

SurroDock: A Deep Learning Surrogate for Accelerated Pre-Docking Ligand Prioritization in Structure-Based Virtual Screening

The rapid expansion of make-on-demand and public chemical libraries has made exhaustive docking-based structure-based virtual screening increasingly difficult. This study introduces SurroDock, a lightweight deep-learning surrogate designed to approximate AutoDock Vina docking scores from low-cost two-dimensional molecular features, serving as a practical pre-filter for docking. SurroDock was evaluated for estrogen receptor alpha using two distinct conformations: an agonist-bound (PDB ID: 1GWR) and an antagonist/SERM-bound (PDB ID: 3ERT). The dataset comprised approximately 334,000 unique compounds curated from the NCI Open Database, PubChem, and BindingDB, all docked using a standardized AutoDock Vina workflow. The model was trained on concatenated 2D molecular representations comprising Morgan fingerprints, MACCS keys, RDKit physicochemical descriptors, Vina-inspired ligand descriptors, atom-pair fingerprints, and 2D pharmacophore fingerprints. The docking-score distributions differed substantially between receptor states, with 3ERT exhibiting more favorable scores than 1GWR and weak inter-state score correlation supporting state-specific modeling. Using the integrated Unified-200k training set (200,000 compounds randomly sampled per receptor from the three docked sources), SurroDock achieved strong held-out validation performance, with R2 values of approximately 0.88 for 1GWR and 0.93 for 3ERT. In retrospective screening-style evaluation, SurroDock recovered substantial fractions of Vina’s top-ranked compounds at the top-1% recall (Recall@1%) of approximately 0.57 and 0.61 for 1GWR and 3ERT, respectively, yielding corresponding enrichment factors (EF@1%) of approximately 57-fold and 61-fold relative to random selection. Overall, the results indicate that 2D-based docking-score surrogate modeling can provide a reproducible and retrainable strategy for large-scale structure-based virtual screening by concentrating docking resources on a smaller, enriched subset of compounds. Because SurroDock emulates a docking scoring function rather than experimental binding affinity, its predictions should be used as prioritization aids and complemented by confirmatory docking, pose inspection, and experimental validation.

Development and validation of an attention-based cGAS-specific deep learning scoring function for structure-based virtual screening

Structure-based virtual screening (SBVS) is a cornerstone of modern drug discovery pipelines, yet conventional scoring functions often lack the resolution to model complex protein–ligand interactions accurately. To overcome these limitations, we developed DeepCGASPred, a cyclic GMP-AMP synthase (cGAS)-specific deep learning scoring function that integrates three-dimensional (3D) convolutional neural networks (CNNs) with multi-head attention mechanisms and composite structural descriptors, including Structural Protein–Ligand Interaction Fingerprints (SPLIF), hydrogen bond features, and extended connectivity fingerprints (ECFP). This integrative approach enables the model to capture spatial, physicochemical, and topological interaction patterns while prioritizing informative regions for accurate classification of active and inactive compounds. DeepCGASPred was trained using a chemically diverse dataset and rigorously validated through systematic hyperparameter optimization and multiple independent runs. Our results demonstrate that the combined feature set SPLIF+Hbonds+ECFP consistently outperforms all tested feature configurations, achieving a precision-recall area under the curve (PR-AUC) of 0.94–0.97, a median precision of approximately 0.99, a median recall of approximately 0.88, and a median F1 score above 0.92 across ten independent runs. Comparative evaluation reveals that DeepCGASPred surpasses established scoring functions such as SMINA, RF-Score, SCORCH, and CNN-Score on this cGAS-specific dataset, particularly in identifying actives under challenging test conditions. A perfect normalized enrichment factor at 1% (NEF1% = 1.00) confirms strong early enrichment performance. Optimal performance was achieved when the attention mechanism was inserted after the first CNN layer, reinforcing its role in enhancing generalization. DeepCGASPred offers a robust, interpretable framework for target-specific SBVS of cGAS inhibitors; the approach may also inform the development of similar tools for other biological targets.