HyperNetWalk: A Unified Framework for Personalized and Cohort-Level Cancer Driver Gene Identification via Reverse Inference on Layered Signaling-Regulatory Network

Accurate identification of cancer driver genes from passenger mutations is essential for understanding tumorigenesis and clinical translation.

We present HyperNetWalk, an unsupervised framework that unifies personalized and cohort-level driver gene identification within a shared inference architecture.

HyperNetWalk builds a layered signaling-regulatory network by integrating protein-protein interactions, approximating upstream signaling, with a gene regulatory network for downstream transcriptional regulation, with transcription factors serving as interface nodes.

Driver identification is formulated as an inverse problem in which observed transcriptional dysregulation is traced back to candidate upstream drivers by reverse random walk.

The resulting sample-specific scores are used directly for personalized prediction and as node weights for cross-sample refinement through hypergraph random walk, enabling both local personalized and global cohort-level prediction.

Across 12 TCGA cancer types, HyperNetWalk outperformed representative existing methods at both prediction levels.

Ablation analyses supported the contributions of the reverse inference formulation and layered network architecture.

Further analyses showed that HyperNetWalk captured cancer-type-specific driver signals, prioritized both recurrent and low-frequency candidate drivers, and produced predictions supported by drug-gene interaction and clinical actionability annotations.