Interpretable cancer driver gene prediction from DNA sequences using a genomic language model

ICLR 2025 Workshop LMRL Submission72 Authors

Track: Tiny Paper Track

Keywords: genomic language model, cancer driver gene, post-hoc explainability

TL;DR: A new interpretable method to predict cancer driver genes from raw DNA sequences based on a genomic language model

Abstract:

Cancer driver genes are usually detected as positively selected genes with high fitness. We hypothesize
that subsequences within these genes carry signals of positive selection that can be learned by
genomic language models (gLMs). In this work, we fine-tuned Caduceus, a high-performing long-range
gLM, to predict cancer driver genes from DNA sequences. Post-hoc interpretations of our
fine-tuned model helped to explain important sequence features associated with gene fitness such
as known somatic mutations in driver genes. Our approach generates meaningful representations of
DNA sequences related to cancer driver genes and provides a framework toward interpretable cancer
driver gene prediction.

Submission Number: 72