Gene expression is regulated by multiple interacting mechanisms, including DNA methylation, genetic variation, and alternative splicing, whose effects are often distributed across many weak and correlated elements. These regulatory influences are highly context-dependent, yet most studies rely on pairwise association testing to identify cis-regulatory elements, limiting the characterization of gene-specific regulatory architecture across cancer states.

We introduce GeneCL (Gene-anchored Contrastive Learning), a representation learning framework that models genes and their associated regulatory elements as nodes in a heterogeneous graph constructed from statistically supported associations. GeneCL learns context-specific embeddings using a contrastive objective, where proximity reflects shared regulatory context rather than marginal associations. Applied to primary tumors from The Cancer Genome Atlas (TCGA) and metastatic tumors from the Hartwig Medical Foundation, using matched RNA-seq and WES/WGS datasets, GeneCL captures gene–anchored relationships and enables systematic characterization of gene-level regulatory architecture across cancer contexts.

GeneCL provides a scalable and extensible framework for learning context-specific regulatory architecture beyond cis-regulatory for integrating and comparing multiple molecular modalities.

Co-authors: Maria Anisimova