Mapping regulatory constraints in colorectal cancer: Patient-specific pathway analysis using graph neural networks

Sammanfattning

Colorectal cancer (CRC) is one of the most common cancer types worldwide and is driven by the accumulation of somatic mutations that disrupt key driver gene pathways. Both coding and non-coding somatic mutations contribute to cancer development and progression. While driver mutations in protein-coding genes have been extensively studied, non-coding region mutations remain less explored despite their potential role as important functional and regulatory drivers of tumorigenesis. Evolutionarily constrained somatic mutations are often associated with functional and regulatory pathway disruptions. Focusing on these conserved mutations reduces data dimensionality considerably; however, a substantial amount of complex, non-linear structure remains that cannot be captured effectively by linear or traditional similarity-based clustering methods. Conventional machine learning approaches are further limited in their ability to capture patient-level heterogeneity due to the sparsity and variability of clinical and metadata features. In this work, evolutionarily constrained somatic mutations from an 820-patient cohort with CRC, together with associated metadata, were systematically filtered, explored, and pre-processed to generate a clean and robust dataset for downstream analysis. A heterogeneous graph neural network (HGNN) was implemented using patient, gene, and gene region as nodes, and patient–gene, gene–gene (protein–protein interaction), and gene–region relationships as edges. Patient metadata, including age and sex, were incorporated as node features, while mean phyloP conservation scores were used as edge weights. A graph model was trained on coding, non-coding, and combined somatic mutation data to predict patient tumor stage. The model performed strongly on both coding and non-coding clustering and showed a reasonable yet low performance on the combined dataset. However, all three models were unable to predict tumor stage with reasonable accuracy. Downstream pathway analysis identified key colorectal cancer pathways including p53 and WNT with stronger signals when non-coding data were combined with coding data. We conclude that HGNN with a combination of evolutionary constraint somatic mutations on coding and non-coding significantly improves patient-level regulatory and pathway mapping. Moreover, hyperparameter optimisation and strategies to address class imbalance are expected to further improve model performance.