HGS-BioAbductor: hypergraph-based explainable survival modelling and LLM-based biological hypothesis abduction
Explainable deep survival modeling remains challenging due to the performance-explainability trade-off in prognostic models and the gap between feature importance and biological abduction. To address these limitations, we present HGS-BioAbductor, a framework that first constructs a biologically embedded hypergraph neural network for accurate survival prediction. Then, from predictive biomarkers and pathways identified by eXplainable Artificial Intelligence (XAI), it leverages a multi-agent-based tournament process to derive novel molecular mechanisms, and the plausibility and novelty of all hypotheses are quantitatively assessed with automatically queried literature. Across 18 independent clinical proteomics/transcriptomics datasets, HGS achieved the highest C-index in 16/18 benchmarks compared to seven other baselines and outperformed interpretable counterparts by a large margin (+0.1-0.2 C-index). Furthermore, in a case study of hepatocellular carcinoma (HCC), the system demonstrates its ability to propose new hypotheses about previously underexplored proteins in HCC prognosis, e.g., PFKFB2, with strong plausibility in the literature (>0.8), and knockdown experiments in Huh7 and MHCC-LM3 cell lines confirmed that PFKFB2 promotes HCC cell proliferation and colony formation. This work bridges data-driven omics discovery with biological interpretability supported by literature, offering an automated and end-to-end paradigm for clinical omics analysis.
Authors
- Yupeng Chen (ORCID: https://orcid.org/0000-0003-1605-9480)
- Pei Jiang
- Linhai Xie
- Qingyuan Zhang
Institutions
- Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
- Southern University of Science and Technology (CN)
- Beijing Proteome Research Center (CN)
Publication Details
- Journal
- npj Precision Oncology
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41698-026-01764-9
- Primary Topic
- Bioinformatics and Genomic Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00