Jun 2026· ACM International Conference on Bioinformatics, Computational Biology and Biomedicine· 0 citations· 30 references
Computer Science
Abstract
Accurate detection of ligand binding sites on proteins plays a crucial role in drug design, functional annotation, and biological classification. Deep learning (DL) models applied to computational binding site prediction have improved the characterization of binding site properties that are often difficult to obtain experimentally. However, most binding site detection DL algorithms search exhaustively around the whole protein, which is computationally expensive. Existing DL models are also sensitive to the choice of hyperparameters and architecture, which can lead to inaccurate predictions. To address these limitations, we present VoxelProt-ligand, a local voxel-based deep learning method for binding site detection in protein-ligand interactions. VoxelProt-ligand is built upon an architecture that encodes protein surfaces as sparse, octree-based 3D voxel grids. These sparse grids are used to train a 3D-CNN for classifying surface regions as binding or non-binding. Accurate position and volume of binding sites is achieved with a local search that leverages protein surface geometry and an energy score. VoxelProt-ligand is trained on the MaSIF-ligand dataset and evaluated on HOLO4K and COACH420. Our results shows close agreement with the ground truth in all evaluation datasets for the joint criteria of whether a protein contains binding sites and average number of detected binding sites per protein. Importantly, VoxelProt-ligand improves detection of binding site volume, achieving significantly higher binding pocket shape overlap than existing methods, while maintaining competitive binding site localization success rates. These results indicate that VoxelProt-ligand offers a viable path toward applications in computational drug discovery that depend on accurate determination of binding pocket geometry.
Despite challenges related to data sparsity and conformational variability, ViTs show strong performance and high robustness in structure-based affinity prediction tasks, underscore their effectiveness in learning spatial patterns and suggest broader applicability to related tasks, such as protein-protein or protein-nucleic acid interaction modeling.
Jakub Poziemski, Paweł Siedlecki· Scientific Reports· 0 citations
Deep learning models are data-hungry, and synthetic (artificial) data has been shown to be invaluable when data availability is low. While this has been demonstrated in certain technology areas, adopting such an approach is new in machine learning (ML) applications in chemistry, except for some pre-training tasks. In drug discovery, predicting binding energy between proteins and ligands is crucial. Many ML-based studies have been proposed to predict protein-ligand binding affinity using existing experimental data. However, these models suffer from inherent biases. Recent efforts have produced PLAS-20k, a synthetic dataset of multiple protein-ligand complex (PLC) conformations generated using molecular dynamics (MD) simulations as a viable option to complement existing experimental data and improve binding affinity prediction. For the binding affinity prediction task, we employ Pafnucy, a deep convolutional neural network, and propose using multiple structures for each PLC from PLAS-20k for training. We compare four different statistical and ML-based result-aggregation techniques. This work demonstrates the utility of dynamic datasets in enhancing binding affinity predictions, laying the foundation for future improvements in predicting similar protein properties using synthetic datasets and more sophisticated models and methods. We propose that physics-based synthetic datasets can significantly help develop more accurate data-driven methods.
Scientific contribution
This study shows that incorporating synthetic molecular dynamics data improves deep learning models for protein–ligand binding affinity prediction beyond static experimental structures. By systematically evaluating frame selection and prediction aggregation strategies, we demonstrate that training on diverse conformational snapshots significantly enhances generalization and accuracy. Our results highlight that dynamic synthetic datasets can enable deep learning models to outperform conventional methods such as MM-PBSA while remaining computationally efficient.
P. Agrawal, Prathit Chatterjee, U. Priyakumar· Journal of Cheminformatics· 0 citations
This work used its strategy, termed neural iterative selection–expansion (NISE), to design proteins that, using different folds, specifically bind to two chemically distinct small-molecule drugs, exatecan and apixaban, with success rates of 100%, respectively.
Benjamin Fry, Kaia Slaw, Nicholas F. Polizzi· Nature· 7 citations· ⚡1
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.
Zhenqiang Zhang, Zhihao Wang, Yang Liu et al.· Journal of Chemical Informat...· 0 citations
Protein-ion interactions are essential for many cellular processes, including enzymatic catalysis, signaling, and allosteric regulation. However, mapping ion-binding sites experimentally remains labor-intensive and expensive. Here, we present BiteNetI, a structure-based deep learning model that uses 3D convolutional neural networks to simultaneously localize ion-binding centers and predict binding residues for 14 biologically relevant ions. Trained on a carefully curated dataset of over 10,000 high-resolution protein–ion complexes, in which near-identical binding sites are consistently annotated by transferring ions between homologous structures, BiteNetI shows strong generalization ability across diverse ions within a unified multitask architecture. On two different test benchmarks, BiteNetI achieves state-of-the-art performance compared to existing ion-binding predictors as well as to a more general method, AlphaFold3, when used to predict the entire structure of protein bound to ions. Finally, for physiologically relevant ions such as Ca2+, Na+ and K+, BiteNetI achieves two- to three-fold improvement in accuracy. Structure-based deep learning enables accurate prediction of protein-ion binding sites across 14 biologically relevant ions, supporting comprehensive and large-scale annotation of protein-ion interactions.
Igor Kozlovskii, Petr Popov· Communications Biology· 0 citations
Abstract Motivation Identifying protein binding sites in protein–protein complexes is a central challenge in structural biology. Binding sites, consisting of groups of residues, govern how proteins recognize, and interact with protein partners. Thus, identifying them is essential for understanding biological function and guiding the design of effective biomolecules and even drug molecules. Despite major progress in computational approaches, their performance remains limited because most models underrepresent the combined influence of surface properties and residue-level information, leaving room for improvement. Recent advances in state-space models and vision-based deep learning offer an opportunity to address these limitations by efficiently modeling long-range spatial dependencies on protein surfaces. Here, we introduce BiMba (protein Binding site prediction using Vision Mamba), a state-space–driven deep learning framework that leverages the efficient long-range modeling capability of the Vision Mamba architecture to learn from three-dimensional (3D) protein surfaces represented as two-dimensional (2D) geometric or physicochemical grids. Results BiMba integrates complementary sources of information, capturing geometric and physicochemical determinants of molecular recognition as surface patches, encoded as 2D images, along with residue-level descriptors, yielding a unified representation that couples spatial topology with biochemical context. BiMba demonstrates competitive performance across diverse and specialized benchmark datasets, often outperforming existing state-of-the-art methods. In addition, BiMba incorporates perturbation-based and gradient-based interpretability analyses by extracting hidden attentions from Mamba layers, enabling visualization of feature relevance and biologically meaningful residue clusters. Overall, our findings establish state-space models as efficient, interpretable, and scalable architectures for molecular surface learning, advancing the application of deep learning in structural bioinformatics. Availability and implementation The BiMba source code, training, test, and benchmark datasets are available at https://github.com/Azam-Shi/BiMba.