Predicting drug-target affinity is a fundamental task in computational drug discovery. Existing methods often rely on a single view or perform shallow multimodal fusion, which may lead to view-specific bias and insufficient coordination among heterogeneous representations. As a result, they often fail to fully capture the complementary biochemical and structural cues underlying drug-target interactions. In this work, we propose TriM-DTA, a hierarchical cross-modal fusion framework for predicting drug-target binding affinity. TriM-DTA jointly models molecular sequence, 2D topology, and 3D geometry of both drugs and targets, and progressively integrates these heterogeneous views into a unified affinity-aware representation. By coordinating structural information before deeper cross-modal interaction, TriM-DTA reduces cross-view mismatch and enables more effective coupling of semantic, topological, and geometric cues. Experimental results demonstrate that TriM-DTA consistently outperforms existing methods and exhibits strong generalizability to unseen drugs and proteins. In addition, visualization analyses reveal that TriM-DTA captures binding-related patterns and highlights chemically meaningful substructures, providing interpretable evidence for its predictions. These findings suggest that coordinated multimodal representation learning provides a useful basis for broader molecular modeling tasks in drug discovery.
Drug repositioning accelerates the identification of novel therapeutic indications for approved compounds by leveraging established safety and pharmacokinetic profiles, thereby substantially reducing development costs and timelines. Heterogeneous graph neural networks have shown strong potential for modeling multi-relational biomedical associations among drugs, diseases, and protein targets. However, their considerable computational complexity limits scalability in high- throughput screening scenarios. Here, we present GraphKD-DR, a cross-level graph fusion framework that reconciles representational expressiveness with computational efficiency through knowledge distillation. Specifically, the teacher model integrates homogeneous similarity graphs with multi-view heterogeneous biomedical representations via Transformer-based cross-level fusion. Predictive, feature-level, and relational knowledge is then transferred from the teacher to a lightweight student model, substantially reducing model complexity while preserving predictive performance. GraphKD-DR achieves approximately 98% fewer parameters and reduces training time by 98.73–99.33%, while retaining comparable predictive performance relative to the teacher model. Experiments on three benchmark datasets demonstrate that GraphKD-DR achieves strong overall performance compared with six state-of-the-art methods across multiple evaluation metrics, including AUROC, AUPRC, and Matthews correlation coefficient. Case studies on Alzheimer's disease and breast neoplasms identify highly ranked candidate drugs with varying degrees of literature support, while molecular docking and ADMET analyses provide complementary in silico evidence for their computational prioritization. These results establish GraphKD-DR as an efficient framework for large-scale drug–disease association prediction and computational candidate prioritization.
Cuffless blood pressure (BP) estimation from wear able biosignals is important for continuous and unobtrusive cardiovascular monitoring, but integrating physiological mea surements with BP-related knowledge and subject-specific context remains challenging. In this study, we present BP-LLaMA, a knowledge-enhanced large language model (LLM) framework for cuffless BP estimation from wearable electrocardiogram and photoplethysmogram signals. Physiologically meaningful features are first extracted and organized according to cardiovascu lar relevance, including cardiac output, peripheral resistance, and arterial stiffness. These features are then represented as structured textual prompts together with BP domain knowledge and user-specific attributes, and LLaMA3-8B is adapted to the BP estimation task through supervised instruction tuning. The proposed framework was evaluated using subject-level five-fold cross-validation on two large-scale wearable datasets, CAS-BP and Aurora-BP, under calibration-free and calibration-based settings. In the calibration-based setting, BP-LLaMA achieved mean absolute errors of 6.97/5.25 mmHg for systolic/diastolic BP on CAS-BP and 6.91/5.23 mmHg on Aurora-BP. BP-LLaMA achieved lower errors than the evaluated LLM-based time-series baselines in most comparisons and remained competitive with strong machine learning and deep learning models. Ablation studies showed the benefits of incorporating BP domain knowledge and user information. These results demonstrate the feasibility of using LLMs as a knowledge-enhanced modeling paradigm for cuffless BP estimation, while also highlighting their value in integrating physiological knowledge and individualized context.
Predicting molecular properties is crucial across scientific and industrial domains such as drug discovery and material science. Contrastive learning has gained prominence as an effective method. However, routine strategies often overlook substructure information and run the risk of guiding the model to learn incorrect knowledge due to molecular property changes during augmentation. To address these issues, we propose a novel molecular contrastive learning model named Molecular Contrastive Learning with Learnable Weighted Substructures (MolCLW). Our model aims to enhance the learning of molecular substructures by comparing molecular representations generated at the molecular-level and substructure-level. Meanwhile, a transformer encoder module is dedicated to learning the weight scores of substructures, enabling the model to adjust the similarity level between positive pairs based on the varying masked substructures. In comparison to baselines, MolCLW demonstrates an average improvement of 2.2% improvement in ROC-AUC on 6 classification benchmarks and an averaged 4.3% decrease in error on 5 regression benchmarks. Furthermore, during the fine-tuning process, the model can attend to the substructures relevant to the downstream task by adjusting the weights of different substructures. The visualization results establish a link between downstream targets and the functional roles of substructures, providing valuable insights and strong support for drug discovery, chemical reactions, and research.
Accurate drug-target interaction (DTI) prediction is fundamental to computational drug discovery. However, existing deep learning methods typically oversimplify biochemical entities as 1D sequences, neglecting crucial spatial topologies, and incur prohibitive quadratic computational costs when modeling long-range protein dependencies. To overcome these bottlenecks, we propose MMB-DTI, a novel multimodal framework that synergistically fuses evolutionary and chemical semantics, extracted via pre-trained language models (ESM-2 and ChemBERTa), with structural topologies derived from graph neural networks. At its core, the framework introduces a hybrid Mamba-BiLSTM architecture, leveraging a selective state space model for linear-time encoding of ultra-long-range protein sequences while precisely preserving local sequential motifs. Furthermore, a bidirectional cross-attention module simulates the biological “induced-fit” mechanism, dynamically capturing reciprocal and fine-grained binding affinities. Comprehensive evaluations across six benchmark datasets (Davis, KIBA, DrugBank, Enzymes, Ion Channels, and GPCRs) demonstrate that MMB-DTI achieves highly competitive predictive performance, significantly outperforming state-of-the-art baselines. Our results highlight that coupling linear-time state space modeling with topological complementarity yields a robust and discriminative paradigm for deciphering complex drug-target associations.
