Pan-cancer benchmarking reveals complementary copy number signatures with distinct multi-omic predictability

Abstract

Copy number signatures provide compact representations of the processes that shape cancer genomes, but signatures derived with different feature encodings are often interpreted as if they were interchangeable. We established a matched-sample pan-cancer benchmark of three major copy number signature compendia, comparing their activity structure, cross-framework concordance, patient stratification, outcome associations, and predictability from non-copy-number molecular data. Signature- level concordance was sparse and concentrated in a limited set of biologically related patterns. Clustering of high-activity signatures produced distinct patient partitions with limited overlap between compendia, although one cluster in each framework showed a directionally favorable outcome association after accounting for cancer-type-specific baseline hazards. Prediction from gene expression, DNA methylation, somatic mutations, age, and tumor purity was strongly framework dependent: test-set F1 scores were 0.93 for Drews, 0.80 for Steele, and 0.24 for Tao. Gene expression provided the largest contribution and largely retained the performance of the full models. These results show that compendium choice is an analytical decision rather than an interchangeable preprocessing step. The benchmark provides a reproducible framework for selecting and interpreting copy number signature representations in pan-cancer studies.