Life
Pathway-Masked Kolmogorov-Arnold Networks Yield an Interpretable Transcriptomic Score Associated with Colorectal Cancer Progression Independently of Clinical Stage
Par Allemailem KS, Alharbi HO, Alrumaihi F et 1 autres.
Contenu en anglais
Résumé
Background
Prognostic gene expression signatures in colorectal cancer are usually reported by standalone discrimination, a standard that random gene sets can meet, and are rarely tested for what they add to routine clinical variables.
Methods
TCGA-COAD and TCGA-READ were pooled into a single colorectal cohort (596 patients, 153 progression-free events). PathKAN places learnable univariate spline functions on a sparse gene-to-pathway topology drawn from MSigDB and BioGRID, reducing 20,681,472 dense weights to 127,985 edges, and is fitted under a Cox partial-likelihood objective. Performance was assessed by nested, repeated stratified cross-validation with dependence-corrected inference against clinical, elastic-net Cox and random survival forest models, and the frozen network was applied to four independent cohorts (1012 patients, 306 events) on two array platforms.