Prediction of protein–protein interactions (PPIs) is fundamental for identifying drug targets and understanding cellular processes. The rapid growth of PPI studies necessitates the development of efficient and accurate tools for automated prediction of PPIs. In recent years, several robust deep learning models have been developed for PPI prediction and have found widespread application in proteomics research. Despite these advancements, current computational tools still face limitations in modeling both the pairwise interactions and the hierarchical relationships between proteins. We present HI-PPI, a novel deep learning method that integrates hierarchical representation of PPI network and interaction-specific learning for protein–protein interaction prediction. HI-PPI extracts the hierarchical information by embedding structural and relational information into hyperbolic space. A gated interaction network is then employed to extract pairwise features for interaction prediction. Experiments on multiple benchmark datasets demonstrate that HI-PPI outperforms the state-of-the-art methods; HI-PPI improves Micro-F1 scores by 2.62%–7.09% over the second-best method. Moreover, HI-PPI offers explicit interpretability of the hierarchical organization within the PPI network. The distance between the origin and the hyperbolic embedding computed by HI-PPI naturally reflects the hierarchical level of proteins. Overall, the proposed HI-PPI effectively addresses the limitations of existing PPI prediction methods. By leveraging the hierarchical structure of PPI network, HI-PPI significantly enhances the accuracy and robustness of PPI predictions.
Background: Multimodal breast ultrasound, including B-mode imaging, color Doppler flow imaging, and elastography, provides complementary information for lesion characterization. However, effectively integrating heterogeneous modalities remains challenging due to inconsistent feature distributions, limited cross-modal interaction, computational cost in existing methods, and sensitivity to noise and missing data. Methods: We presented an efficient Cross-Modal Interaction and Dynamic Fusion Network (CIDFNet) for multimodal breast ultrasound analysis. The framework integrates a multi-scale feature enhancement module to improve modality-specific representations, a cross-modal interaction module to enable early-stage feature exchange across modalities, and a dynamic fusion strategy to adaptively combine modality information based on feature reliability estimation. In addition, an invertible neural network is incorporated to reconstruct missing modality features during training. Results: Experiments on an internal dataset of 248 patients with 1532 images show that CIDFNet obtains an AUC of 85.69%, accuracy of 75.51%, recall of 50.00%, F1-score of 62.50%, and precision of 83.33%, while requiring 49.51 M parameters and 79.79 G FLOPs, respectively. Under a simplified Gaussian noise perturbation setting, performance degradation is observed. Conclusions: CIDFNet presents a framework for multimodal breast ultrasound analysis that reflects a trade-off between performance and computational efficiency.
Background: Accurate breast lesion detection in ultrasound images remains challenging due to speckle noise, acoustic artifacts, low contrast, and blurred lesion boundaries. Although YOLO-based detectors are efficient, they may not fully capture long-range contextual dependencies and directional structural information that are important for reliable lesion localization. Methods: We proposed a lightweight context-structure synergistic framework based on YOLOv13. A Dual-Stream Mamba Aggregation (DSMA) module is introduced to enhance contextual feature aggregation with linear-complexity state-space modeling, while a Structure-aware Axial Attention (SAA) module is used to model horizontal and vertical structural dependencies. The two modules are integrated in a stage-specific manner to improve feature representation with limited computational overhead. Results: On the BUV and WH-BUS datasets, the proposed method achieved competitive detection performance while maintaining 2.50M parameters, 6.4 GFLOPs, and 161.29 FPS. Ablation, cross-dataset, robustness, and visualization analysis showed that DSMA and SAA provide complementary benefits for contextual representation and structure-aware localization. Conclusion: The proposed method provides a lightweight detection framework for breast ultrasound images by jointly modeling contextual and structural features.
Background: Prediction of protein–protein interactions (PPIs) is fundamental for identifying drug targets and understanding cellular processes. The rapid growth of PPI studies necessitates the development of efficient and accurate tools for automated prediction of PPIs. In recent years, several robust deep learning models have been developed for PPI prediction and have found widespread application in proteomics research. Despite these advancements, current computational tools still face limitations in modeling both the pairwise interactions and the hierarchical relationships between proteins. Results: We present HI-PPI, a novel deep learning method that integrates hierarchical representation of PPI network and interaction-specific learning for protein-protein interaction prediction. HI-PPI extracts the hierarchical information by embedding structural and relational information into hyperbolic space. A gated interaction network is then employed to extract pairwise features for interaction prediction. Experiments on multiple benchmark datasets demonstrate that HI-PPI outperforms the state-of-the-art methods, HI-PPI improves MicroF1 scores by 2.62%–7.09% over the second-best method. Moreover, HI-PPI offers explicit interpretability of the hierarchical organization within the PPI network. The distance between the origin and the hyperbolic embedding computed by HI-PPI naturally reflects the hierarchical level of proteins. Conclusions: Overall, the proposed HI-PPI effectively addresses the limitations of existing PPI prediction methods. By leveraging the hierarchical structure of PPI network, HI-PPI significantly enhances the accuracy and robustness of PPI predictions.
Introduction: Sparse event annotations limit drug-drug interaction (DDI) prediction, particularly for rare events and event categories absent from training data. Methods: We propose MSER-DDI, a framework with two independently trained pathways that represent target events using labeled drug pairs or textual descriptions. The sample-driven pathway, PharDDIE, combines molecular structure and biological context with reconstruction-regularized latent encoding to construct event prototypes from a few labeled pairs. The semantics-driven pathway, EviDDIE, maps event descriptions into the drug-pair representation space and uses evidential inference without labeled target-event examples. Results: In 1-shot rare-event prediction, PharDDIE achieved higher mean area under the receiver operating characteristic curve (AUROC), accuracy, and event-macro F1 than retrained RareDDIE across five matched training seeds. These means also exceeded the published baseline estimates. Under the respective evaluation conditions, 5-shot predictions had lower Brier score, negative log-likelihood, and calibration error than 1-shot predictions, together with lower high-confidence error rates and broader coverage across all three event-frequency groups. EviDDIE predicted interactions for event categories excluded from molecular training. Literature-based assessment identified drug- or class-level pharmacological consistency for a subset of selected positive benchmark examples. Discussion: MSER-DDI supports event-level prediction when target-event annotations are scarce or unavailable, while joint evaluation of discrimination and probability quality characterizes predictive performance and confidence. The resulting predictions provide interaction hypotheses for further pharmacological investigation.