# ONCOPLEX: an oncology-inspired hypergraph model integrating diverse biological knowledge for cancer driver gene prediction

## TL;DR

ONCOPLEX is a new AI model that uses complex biological networks to better identify genes that cause cancer.

## Contribution

ONCOPLEX introduces a novel hypergraph-based framework integrating diverse biological data for improved cancer driver gene prediction.

## Key findings

- ONCOPLEX outperforms existing methods in identifying cancer driver genes across various cancer types.
- The model accurately recovers known driver genes and identifies new candidates supported by biological evidence.

## Abstract

Cancer development is driven by a small subset of somatic mutations, known as driver mutations, that disrupt key regulatory processes in cells. These mutations occur in specific genes, called cancer driver genes, whose altered functions promote tumor initiation and progression. Accurately identifying driver genes remains a major challenge due to their rarity and the overwhelming presence of passenger mutations. Recent advances in graph-based deep learning have improved the modeling of gene interactions, but most approaches are limited to pairwise connections and fail to capture the higher-order complexity of biological systems. We introduce ONCOPLEX, a hypergraph-based neural network framework that models genes as nodes and curated cancer-related pathways as hyperedges, enabling the representation of multi-gene interactions. Unlike previous methods, ONCOPLEX integrates diverse molecular and phenotypic features, such as somatic mutations, gene expression, and DNA methylation, into a pathway-informed hypergraph structure to learn biologically meaningful gene representations. ONCOPLEX is trained in a supervised manner on labeled driver and non-driver genes, with unlabeled genes included as nodes during representation learning. Comprehensive evaluations across pan-cancer and cancer-type-specific settings show that ONCOPLEX consistently outperforms state-of-the-art methods in classification and ranking metrics. It accurately recovers known driver genes and highlights novel candidates supported by literature and enrichment analyses. These findings underscore the power of pathway-guided hypergraph modeling for advancing cancer driver gene discovery.