OBJECTIVE: Precision medicine applications depend on elucidating the underlying molecular mechanisms of disease. However, many disorders, such as endometriosis, remain poorly characterised genetically due to data scarcity, positive-unlabelled (PU) imbalance, and the heterogeneous structure of biomedical knowledge. This study aims to develop HetBio-CLiP, a novel heterogeneous graph-based contrastive learning (CL) methodology that improves candidate gene prioritisation across heterogeneous biomedical graphs by explicitly addressing positive-unlabelled learning constraints and structural heterogeneity. METHODS: HetBio-CLiP (Heterogeneous Biomedical graph Contrastive Learning with Interactions and Positive-Unlabelled learning) integrates multi-relational genomic, variant, and clinical data from real-world patient cohorts. To address the challenges of data scarcity and class imbalance, the graph neural network (GNN) methodology combines heterogeneous graph CL with PU learning. Furthermore, the model incorporates an interpretable GNNShap explainer to provide transparency at both the feature and edge levels, to assess the biological relevance of the predictions. RESULTS: HetBio-CLiP achieved superior performance across key evaluation metrics, attaining the highest Area Under the Curve (AUC) of 0.9489 ± 0.04 and Area Under the Precision-Recall Curve (AUPR) of 0.9401 ± 0.05. This outperformed state-of-the-art baselines such as M-SAGEGraph (AUC 0.9149 ± 0.04) and GAT (AUC 0.9138 ± 0.07). In terms of ranking candidate genes, the model achieved the highest Precision@42 (0.6933 ± 0.06) and TP@42 (29.1 ± 2.4) and second-highest Normalised Discounted Cumulative Gain (NDCG@42) (0.827 ± 0.06). Ablation studies confirmed that the integration of PU and contrastive learning was essential for these performance gains. Biologically, the model successfully prioritised known pathophysiology drivers, with top-ranked validated gene predictions including WNT4, GREB1, and ESR1. CONCLUSION: HetBio-CLiP presents an effective and interpretable methodology for uncovering candidate gene-disease relationships in data-scarce environments. By combining CL with PU-training, the proposed method addresses critical limitations of current gene prioritisation models and produces biologically meaningful predictions. The model demonstrates a strong balance between ranking accuracy and recall, identifying broader sets of plausible candidates than existing methods. This approach benefits researchers, clinicians, and downstream genomic studies by offering an improved foundation for identifying novel disease-associated genes and guiding future experimental investigation in endometriosis and other complex diseases.

Current genome-wide association studies provide valuable insights into the genetic basis of ischaemic stroke (IS) risk. However, polygenic risk scores, the most widely used method for genetic risk prediction, have notable limitations due to their linear nature and inability to capture complex, nonlinear interactions among genetic variants. While deep neural networks offer advantages in modeling these complex relationships, the multifactorial nature of IS and the influence of modifiable risk factors present additional challenges for genetic risk prediction. To address these challenges, we propose a Chromosome-wise Multi-task Genomic (MetaGeno) framework that utilizes genetic data from IS and five related diseases. The framework includes a chromosome-based embedding layer to model local and global interactions among adjacent variants, enabling a biologically informed approach. Incorporating multi-disease learning further enhances predictive accuracy by leveraging shared genetic information. Among various sequential models tested, the Transformer demonstrated superior performance, and outperformed other machine learning models and PRS baselines, achieving an AUROC of 0.809 on the UK Biobank dataset. Risk stratification identified a two-fold increased stroke risk (HR, 2.14; 95% CI: 1.81-2.46) in the top 1% risk group, with a nearly five-fold increase in those with modifiable risk factors such as atrial fibrillation and hypertension. Finally, the model was validated on the diverse All of Us dataset (AUROC = 0.764), highlighting ancestry and population differences while demonstrating effective generalization. This study introduces a predictive framework that identifies high-risk individuals and informs targeted prevention strategies, offering potential as a clinical decision-support tool.

Cancer risk prediction is a cornerstone of personalized medicine that offers opportunities for early detection and preventive interventions. However, the current models are designed to predict cancer risk face several challenges. First, most rely on traditional statistical methods, which struggle to capture the complexity of genetic, family medical history, and lifestyle factors. Hence, the accuracy of these models is limited. Additionally, the models neglect to integrate multidimensional data sources, particularly genetic information like single nucleotide polymorphisms (SNPs), which could enhance prediction accuracy. Third, while the system might effectively predict risk, it cannot translate those predictions into actionable healthcare recommendations to reduce cancer risk. In this study, we address all three of these limitations. With a focus on six prevalent cancers—we extracted SNP data from the UK Biobank and designed a novel risk prediction model for cancer and personalized healthcare recommendations based upon the mixture of experts (MoE) paradigm and large language models (LLMs), respectively. Named MoE-HRS, experts based two router networks for separate processing by the Transformer and the convolutional neural network (CNN). Experiments on UK Biobank data show that our model outperforms state-of-the-art cancer risk prediction models. To bridge the gap between risk prediction and practical healthcare applications, we devised a healthcare recommender system powered by LLMs. This approach holds promise for enhancing early detection rates and promoting preventive healthcare management (relevant coding and data are available at https://github.com/bjtu-lucas-nlp/MoE-HRS ).