MSCA-PLA: multi-scale cross-attention with differentiable pooling for protein-ligand binding affinity prediction.

Structure-agnostic protein–ligand binding affinity prediction via hierarchical representation alignment

Abstract Motivation To enable real-world protein-ligand affinity prediction, not only out-of-distribution generalization but also robustness to variable structural availability and quality should be considered in model design. Results We present AlignNet, a hierarchical representation alignment framework that mitigates intra- and inter-molecular heterogeneity to learn robust protein-ligand embeddings for generalizable affinity prediction, even from sequence-level inputs. Its intra-molecular module projects unimodal and multimodal features into a unified space, aligning augmented multimodal views for feature fusion and unimodal with multimodal embeddings to distill multimodal priors for structure-agnostic inference. Its inter-molecular module aligns protein and ligand embeddings for cross-molecular integration. Extensive experiments show that AlignNet (i) achieves highly competitive performance, with up to a 20.4% gain in SCC on the challenging LBA 30% split under sequence-only settings, suggesting improved out-of-distribution generalization; and (ii) learns well-separated affinity-related clusters, supporting reliable structure-independent prediction. Availability and implementation AlignNet is available at https://github.com/altriavin/AlignNet.

MSIGR-PLA: Integrating Multi-Scale Interaction and Global Representations for Protein-Ligand Affinity Prediction.

The proposed MSIGR-PLA is an integrative framework that integrates local multi-scale interaction features with global protein-ligand representations to improve the accuracy of PLA prediction and consistently outperforms existing methods on four benchmark datasets.

GAMT-GINE: A Graph Isomorphism Network Integrating Continuous Spatial Awareness and Multi-Task Learning for Protein–Ligand Binding Affinity Prediction

Protein–ligand interactions (PLIs) play a crucial role in drug discovery, and accurately predicting protein–ligand binding affinity (PLA) remains a central challenge in computer-aided drug design. Although graph neural networks (GNNs) have demonstrated considerable potential in molecular modeling, existing methods still face several limitations, including excessive reliance on hand-crafted chemical features, loss of spatial information, and difficulties in integrating heterogeneous affinity labels, which restrict their generalization capability in PLA prediction. To address these challenges, we propose GAMT-GINE, a graph isomorphism network that integrates continuous spatial awareness with multi-task learning. The model employs minimalist atomic features and a batch-normalization-free mechanism, together with a multi-task branch that uses a large amount of half-maximal inhibitory concentration (IC50) data as an auxiliary prediction target. Experimental results show that GAMT-GINE achieves a Pearson’s correlation coefficient (Rp) of 0.791 and a root mean square error (RMSE) of 1.403 on the CASF-2013 benchmark dataset. In the generalization evaluation on CASF-2016, Rp further increases to 0.831, while RMSE decreases to 1.227, demonstrating performance comparable to that of current State-of-the-Art models. Furthermore, comprehensive evaluations, including ablation studies, feature importance analysis, analysis of the effects of data filtering on model performance and data composition, and analysis of the influence of training–test data similarity on prediction results, indicate that GAMT-GINE can effectively utilize continuous spatial information and heterogeneous affinity labels, achieving good predictive accuracy and cross-dataset generalization capability.

Dual Cross-Attention Network for Hierarchical Feature Fusion in Protein-Protein Interaction Prediction

Characterizing protein-protein interactions (PPIs) is essential for deciphering core biological processes, including signal transduction, metabolic pathway regulation, immune recognition, and cell cycle control. However, experimental PPI determination remains time-consuming and expensive, driving the adoption of deep learning as an efficient and accurate computational approach. Current deep-learning-based PPI prediction models typically process both intra- and inter-protein as isolated units in feature extraction, thereby ignoring mutual information transfer within a single protein and the interacting pair. To address this limitation, we propose DCAPPI (Dual Cross-Attention network for Protein-Protein Interaction prediction), a novel framework leveraging dual cross-attention modules for hierarchical feature fusion at both intra- and inter-protein levels. First, the Channel Cross-Attention module processes protein sequence and structure as distinct input channels. It generates deep intra-protein representations by performing cross-attention between sequence-derived and structure-derived tokens, achieving multimodal feature integration. Second, the Partner Cross-Attention module models the target protein and its interacting partner as a pair of correlative units. By performing cross-attention operations across these units, it enables collaborative feature fusion and constructs context-aware inter-protein interaction features. Evaluation results indicate that DCAPPI achieves superior performance over state-of-the-art methods on benchmark datasets.

Meta-learning GNN with MD-informed attention for cross-species prediction of phosphoinositide-dependent kinase-1 (PdK1) inhibitors in termite control

DeepGCL: Multi-View Graph Contrastive Learning for Enhanced Drug-Target Binding Affinity Prediction Through Protein Pocket-Drug Interaction Modeling.

Drug-target binding affinity (DTA) prediction plays a critical role in drug discovery. Although recent approaches have demonstrated the importance of incorporating protein pocket information, they still struggle to jointly model local atomic-level interactions and global sequence contexts, as well as capture their intricate cross-modal dependencies, within a unified framework for accurate prediction. To address this limitation, we present DeepGCL, a novel multi-modal framework that leverages multi-view graph contrastive learning to capture latent representations of pocket-drug interactions and their underlying molecular determinants. The DeepGCL integrates protein sequences, pocket-drug interaction graphs, and drug molecular graphs within a unified framework to learn complementary molecular representations. By combining sequence encoding with multi-view graph contrastive learning through augmented view generation, the DeepGCL effectively captures contextual protein information and structural interactions between pockets and drugs. This enables accurate characterization of spatial and chemical relationships between binding partners. Comprehensive evaluation on multiple benchmarks shows that DeepGCL consistently outperforms state-of-the-art methods across standard metrics. Ablation studies further confirm that the protein feature module and graph contrastive learning module both make significant and complementary contributions to performance. These findings underscore the effectiveness of multi-view learning paradigms in capturing the multifaceted nature of drug-target interactions and highlight the importance of integrating diverse molecular representations for accurate binding affinity prediction. Our code is available at https://github.com/domtryto/DeepGCL.