Yan Kang

Large Language Models (LLMs) are increasingly demonstrating considerable promise in predicting disease comorbidities. Although structural data from disease networks and retrieval-augmented generation (RAG) have been adopted to enhance disease comorbidity prediction, three underlying limitations remain. (i) Current RAGs implementations often compromise semantic coherence by relying on fixed chunking strategies. (ii) Structural information for disease networks is generally extracted using a single metric, regardless of potential disease associations driven by complex biological mechanisms. (iii) Conventional Chain-of-Thought (CoT) prompts are normally susceptible to error propagation due to their dependence on coarse-grained entity extraction. To overcome these limitations, this study introduced a novel multi-strategy RAG framework integrating a dual-track retrieval approach, a multi-stage CoT strategy, and a multi-metric heuristic to support high-quality, domain-specific reasoning. Built upon a dynamic vector database, the dual-track retrieval mechanism identified implicit relationships among diseases. We also developed a specialized dynamic ensemble heuristic strategy, which adapts supervised ensemble methods to improve disease association prediction by integrating multiple topological metrics. Furthermore, a multi-stage CoT strategy was further developed to enhance the transparency and robustness of the inference process. Finally, a dynamic selection mechanism was applied to adaptively integrate structured signals with semantic signals for optimized predictive performance. Extensive evaluations across multiple benchmarks indicated that our framework exceeded state-of-the-art baselines, registering AUC increases of 10.41% and 5.88%, Average Precision (AP) increases of 9.70% and 5.10%, and F1-score increases of 8.37% and 3.65%, respectively. Further experiments under zero-shot, few-shot, and ablation settings validated the model's effectiveness, generalizability, and interpretability.

Deep learning-based models have been widely used to predict electronic health record (EHR) events by exploiting diagnostic characteristics. Despite significant progress, three limitations remain: 1) effectively modeling dynamic relationships among diseases, 2) fully leveraging diagnosis code ontologies from multiple perspectives, and 3) incorporating unstructured discharge summaries. To address these challenges, we propose a coarse-to-fine disease learning framework with patient notes for EHR event prediction, tailored to capture both dynamic and static disease characteristics. First, we construct a fine-grained dynamic disease graph by removing disease weakly correlated disease pairs based on co-occurrence distributions. Second, disease embeddings are refined by integrating coarse and fine-grained information within the hierarchical structure of ICD-9-CM codes. In addition, discharge summaries are combined with auxiliary patient notes for collaborative disease learning. Finally, gated recurrent units, location-based attention, and soft attention mechanisms are utilized to further enhance embedding representations. Experiments on two real-world EHR datasets, MIMIC-III and MIMIC-IV, demonstrate that our model consistently outperforms nine baseline methods in EHR prediction. The source code can be found at https://github.com/YNU-L/CCDLD

Protein-protein interaction (PPI) prediction is essential for understanding cellular functions and enabling applications in drug development and disease research. PPI networks exhibit multi-scale structural and semantic heterogeneity and introduce representational bias under topological imbalances, leading to insufficient exploitation of subgraph-level semantic complexity. Moreover, many existing approaches rely heavily on external data during modeling, resulting in high computational costs for large-scale inferences. In this paper, we propose JCCLPPI, a joint curriculum- and contrastive-learning framework for lightweight PPI prediction. First, we implement a PPI network-structure encoding module designed to mitigate topological bias and learn biologically interpretable representations without relying on handcrafted features or external annotations. Next, we propose a motif-based curriculum learning module that incrementally introduces training samples according to their structural-semantic complexity, thereby enhancing model robustness to long-tail distributions and structural heterogeneity. Finally, our approach incorporates two graph neural network-based modules during PPI inference to perform local structural modeling and global context encoding, facilitating multi-scale feature extraction. Experiments conducted on two widely used human PPI benchmark datasets, SHS27k and SHS148k, demonstrate that JCCLPPI improves model generalization. It achieves an approximate 4% increase in micro-F1 score compared to state-of-the-art methods, while also improving computational efficiency by reducing memory consumption by 76% and inference time by 35%. Furthermore, JCCLPPI provides a scalable basis for therapeutic target prioritization and early-stage drug discovery.