生信工程，Molecular subtype classification requires integrating complementary omics while preserving bi-ological organisation and accommodating patient-level variation in assay availability. Existing approaches typically address patient-specific assay contribution, pathway-guided representa-tion learning and higher-order patient relations separately, leaving their joint information flow difficult to trace.
数据处理，知识引导，超图神经网络，亚型分类，Multi-Omics Residual Encoding (MORE) produces view-specific representa-tions and a local state; Cross-Omics Pathway Activity (COPA) uses the encoded views to regulate a shared pathway anchor; Pathway-Anchored Hypergraph Learning (PHGL) propa-gates the conditioned state through patient-centred hyperedges; and Contribution-Balanced Gated Fusion (CBGF) coordinates local and propagated evidence before the classifier.
本项目面向高维异构多组学数据的癌症精准分型，构建了知识锚定的端到端超图学习框架 KAHM-Net，通过多组学残差编码、跨组学通路条件化、患者中心超图传播与贡献平衡门控融合，协同建模患者的分子特征、生物通路背景及群体高阶关联。核心难点在于协调不同组学的信息差异、实现生物知识与数据驱动表示的有效耦合，并控制小样本高维建模中的过拟合与信息泄漏；主要亮点是将患者自适应的组学整合与通路引导的关系学习纳入统一预测路径，并在三个 TCGA 癌症队列中开展性能比较、模块消融及部分组学缺失验证。
评论