Protein-protein interactions (PPIs) are fundamental to many biological processes, including cell signaling, gene expression regulation, immune responses, and protein complex formation. Small molecule inhibitors targeting specific PPIs are expected to treat diseases such as cancer and viral infections by modulating pathophysiological processes. Despite their clinical importance, the development of PPI inhibitors is challenging due to limited experimental validation data, which complicates the accurate prediction of novel inhibitors. Therefore, there is an urgent need for advanced computational methods that can effectively integrate multiple data types and improve prediction accuracy. In this study, we proposed a new framework, PTPPI, to efficiently predict protein-protein interaction inhibitors (PPIIs). PTPPI integrates multiple molecular features, including extended connectivity fingerprints (ECFPs) for structural representation and deep semantic embeddings of SMILES sequences generated by the ChemBERTa pre-trained model. These features are processed by independent encoders and fused using an interactive attention mechanism, which enhances the molecular representation. In addition, PTPPI adopts a multi-task learning approach, enabling the model to both reconstruct input features and accurately predict inhibition scores. Experimental results on eight PPI target families, focusing on inhibitor identification and potency prediction, demonstrate that PTPPI outperforms existing methods. It not only integrates multiple molecular features effectively but also achieves superior prediction performance. This makes PTPPI a valuable and reliable tool for discovering new PPI inhibitors, thus opening up new possibilities for drug discovery and disease treatment.

Accurate prediction of molecular properties is a fundamental component of Artificial Intelligence-driven Drug Design (AIDD). In molecular property prediction, even slight structural variations, such as atomic arrangement or bond length, could impact molecular properties. Existing methods primarily focus on coarse-grained molecular modeling on SMILES, 2D, or 3D graphs. However, this paradigm struggles to adequately capture complex internal molecular interactions and overlooks finer-grained structural details, such as atom-bond-spatial positions. In this paper, we propose a molecular property prediction model, TriPercept, based on unified fine-grained representation learning. Specifically, TriPercept integrates atomic, topological, and geometric features, fully exploiting their intrinsic complementarity. We introduce three dedicated encoders to separately learn features at the atomic, bond, and spatial distance levels, thus addressing the neglect of complex structural details. Additionally, we design a structure-aware graph neural network to integrate these features and enhance structural consistency modeling with a contrastive and self-supervised based pretraining strategy. To evaluate the performance of TriPercept, we conduct systematic experiments on six classification datasets and six regression datasets. The results demonstrate that TriPercept performs excellently on most datasets. Visualization experiments further verify that TriPercept not only confirms the effectiveness of multi-level structural integration but also clearly reveals its interpretability in focusing on chemically relevant features and capturing complex structural relationships. The data and code are available at https://github.com/TiAW-Go/TriPercept.

Drug-target interaction (DTI) prediction plays a crucial role in drug discovery and repositioning. However, existing computational methods often overlook the internal chemical structure of drugs, the structural context of proteins, and their dynamic conformational adaptability during binding. To address these limitations, we propose MoFA-DTI, a deep learning-based multimodal fusion model for accurate DTI prediction. Our approach employs a graph convolutional network (GCN) to extract molecular graph features and a bidirectional gated recurrent unit (BiGRU) to capture protein sequence dependencies. A modalityaware feature aggregation module integrates self-attention and cross-attention mechanisms to learn both intra-modal and intermodal representations. Positional encoding and convolutional layers further enhance protein feature refinement. Experimental results on multiple public datasets show that MoFA-DTI consistently outperforms existing methods.

Graph neural networks have demonstrated impressive performance in molecular representation learning. However, when dealing with out-of-distribution data, the generalization ability of existing models often drops significantly. To address this challenge, we have proposed an adversarial domain generalization framework aimed at achieving robust molecular characterization in heterogeneous environments. Specifically, we designed a multi-stage task that first discovers latent domain distributions through latent category-independent features, and then applies adversarial optimization to enhance class invariance and domain invariance respectively. In addition, we have introduced two types of adversarial balance strategies to reduce class domain correlation and stabilize adversarial training. Extensive experiments have shown that our model performs significantly better than the state-of-the-art baseline models. T-SNE visualization and paired H -divergence measurements further confirm that the learned feature space exhibits clear domain separability and domain distribution differences.

MOTIVATION: Predicting protein-protein interaction (PPI) sites is essential for advancing our understanding of protein interactions, as accurate predictions can significantly reduce experimental costs and time. While considerable progress has been made in identifying binding sites at the level of individual amino acid residues, the prediction accuracy for residue subsequences at transitional boundaries-such as those represented by patterns like singular structures (mutation characteristics of contiguous interacting-residue segments) or edge structures (boundary transitions between interacting/non-interacting residue segments) still requires improvement. RESULTS: we propose a novel PPI site prediction method named MVSO-PPIS. This method integrates two complementary feature extraction modules, a subgraph-based module and an enhanced graph attention module. The extracted features are fused using an attention-based fusion mechanism, producing a composite representation that captures both local protein substructures and global contextual dependencies. MVSO-PPIS is trained to jointly optimize three objectives: overall PPI site prediction accuracy, edge structural consistency, and recognition of unique structural patterns in PPI site sequences. Experimental results on benchmark datasets demonstrate that MVSO-PPIS outperforms existing baseline models in both accuracy and structural interpretability. AVAILABILITY AND IMPLEMENTATION: The datasets, source codes, and models of MVSO-PPIS are all available at https://github.com/Edwardblue282/MVSO-PPIS.

Accurate prediction of molecular properties is a key component of Artificial Intelligence-driven Drug Design (AIDD). Despite significant progress in improving these predictive models, balancing accuracy with computational complexity remains a challenge. Molecular topological and geometric features provide rich spatial information, crucial for improving prediction accuracy, but their extraction typically increases model complexity. To address this, we propose TGF-M (Topology-augmented Geometric Features for Molecular Property Prediction), a novel predictive model that optimizes feature extraction to enhance information capture and improve model accuracy, and reduces model complexity to lower computational cost. This approach enhances the model's ability to leverage both topological and geometric features without unnecessary complexity. On the re-segmented PCQM4Mv2 dataset, TGF-M performs remarkably, achieving a low mean absolute error (MAE) of 0.0647 in the HOMO-LUMO gap prediction task with only 6.4M parameters. Compared to two recent state-of-the-art models evaluated within a unified validation framework, TGF-M demonstrates comparable performance with less than one-tenth of the parameters. We conducted an in-depth analysis of TGF-M's chemical interpretability. The results further validate the method's effectiveness in leveraging complex molecular topology and geometry during model learning, underscoring its potential and advantages. The trained models and source code of TGF-M are publicly available at https://github.com/TiAW-Go/TGF-M.

Biomedical knowledge graph reasoning is capable of discovering hidden new knowledge based on existing biomedical data, providing ideas and references for new drug discovery, disease research, and so on. The entire graph topology structure formed by triplets and the attribute descriptions for each entity are crucial information for discovering new knowledge. Some add a variety of additional information to aid reasoning, namely multimodal reasoning. However, current multimodal reasoning techniques often rely solely on vector space distance inferences based on triplets themselves, making it difficult to capture more complex relationships and dependencies between facts. This work integrated triplet entity relations, graph topology structures, and attribute descriptions for each entity to incorporate richer information, and utilized logical rules as external knowledge for relational reasoning in biomedical knowledge graphs. We have evaluated our approach on PharmKG.

Retrosynthesis constructs rational synthetic pathways by predicting reactants from the target product. Previous research utilizing Graph Neural Networks often relies on two-dimensional molecular representations. The two-dimensional representations of functional groups of the same type are identical, and ignoring differences in their three-dimensional(3D) structures can cause confusion in identifying reaction centers. Additionally, they disconnect the reaction centers, hindering the accurate identification of chemical transformation rules. To address these issues, we propose the Retro3D model based on a semi-template-based method, integrating 3D structural information to enhance molecular representation. Due to the discrepancies in bond lengths and bond angles of the same type of functional groups in different 3D spatial structures, our method improves the accuracy of identifying reaction centers in the target product. Furthermore, our method determines whether reaction centers are disconnected, differentiating between types of reaction centers to obtain two types of chemical transformation rules. Given the known reaction class, experiments demonstrate that our model achieves superior accuracy in predicting reaction centers and final reactants compared to previous benchmark models. In summary, our methodology aligns with established chemical principles, enhancing its interpretability and broad applicability across diverse synthetic challenges.

Depression is a prevalent global mental health issue that significantly impacts both individuals and societies. Current methods often overlook the heterogeneity of depressive symptoms and fail to fully leverage available data, underscoring the need for more advanced computational methods DTI prediction. This paper introduces KGM4DTI, a multi-semantic, multi-view, and multi-level meta-path Transformer framework that integrates Transformer and GNN. KGM4DTI effectively captures both global semantic and local structural information between interaction pairs within a depression knowledge graph (DepKG), which comprises 2,258 drugs, 3,360 proteins, 6 subtypes of depression, 196 phenotypes, and 238 pathways. Comparative experiments have demonstrated that KGM4DTI significantly outperforms existing state-of-the-art methods, achieving an AUROC of 99.72 and an AUPR of 99.76. The robustness of framework is further confirmed through extensive ablation studies, which highlight the crucial role of integrating both global and local information. These results underscore the potential of KGM4DTI in accurately predicting DTI. This approach has promising implications for drug discovery, particularly in developing effective treatments for depression, and could extend its applicability to broader mental health disorders. The data and code are available at https://github.com/Tianxyuu/KGM4DTI/.

MOTIVATION: Protein-protein interaction sites (PPIS) are crucial for deciphering protein action mechanisms and related medical research, which is the key issue in protein action research. Recent studies have shown that graph neural networks have achieved outstanding performance in predicting PPIS. However, these studies often neglect the modeling of information at different scales in the graph and the symmetry of protein molecules within three-dimensional space. RESULTS: In response to this gap, this article proposes the MEG-PPIS approach, a PPIS prediction method based on multi-scale graph information and E(n) equivariant graph neural network (EGNN). There are two channels in MEG-PPIS: the original graph and the subgraph obtained by graph pooling. The model can iteratively update the features of the original graph and subgraph through the weight-sharing EGNN. Subsequently, the max-pooling operation aggregates the updated features of the original graph and subgraph. Ultimately, the model feeds node features into the prediction layer to obtain prediction results. Comparative assessments against other methods on benchmark datasets reveal that MEG-PPIS achieves optimal performance across all evaluation metrics and gets the fastest runtime. Furthermore, specific case studies demonstrate that our method can predict more true positive and true negative sites than the current best method, proving that our model achieves better performance in the PPIS prediction task. AVAILABILITY AND IMPLEMENTATION: The data and code are available at https://github.com/dhz234/MEG-PPIS.git.

Protein-protein interactions (PPIs) are a major component of the cellular biochemical reaction network. Rich sequence information and machine learning techniques reduce the dependence of exploring PPIs on wet experiments, which are costly and time-consuming. This paper proposes a PPI prediction model, multi-scale architecture residual network for PPIs (MARPPI), based on dual-channel and multi-feature. Multi-feature leverages Res2vec to obtain the association information between residues, and utilizes pseudo amino acid composition, autocorrelation descriptors and multivariate mutual information to achieve the amino acid composition and order information, physicochemical properties and information entropy, respectively. Dual channel utilizes multi-scale architecture improved ResNet network which extracts protein sequence features to reduce protein feature loss. Compared with other advanced methods, MARPPI achieves 96.03%, 99.01% and 91.80% accuracy in the intraspecific datasets of Saccharomyces cerevisiae, Human and Helicobacter pylori, respectively. The accuracy on the two interspecific datasets of Human-Bacillus anthracis and Human-Yersinia pestis is 97.29%, and 95.30%, respectively. In addition, results on specific datasets of disease (neurodegenerative and metabolic disorders) demonstrate the ability to detect hidden interactions. To better illustrate the performance of MARPPI, evaluations on independent datasets and PPIs network suggest that MARPPI can be used to predict cross-species interactions. The above shows that MARPPI can be regarded as a concise, efficient and accurate tool for PPI datasets.