Representations of somatic mutations for cancer typing

Advisor: Dr. Bradley A. Maron, University of Maryland School of Medicine. With Aaron Ge and Xingyan Liu.

Question. Models for cancer typing and DNA-repair-deficiency prediction differ less in their algorithms than in how they represent mutations. Does a neural network that reads individual mutations beat a well-designed feature table under the same compute budget?

Method. Designed Bio MAF v4, 948 interpretable per-tumor summaries of mutation consequences, driver genes, hotspots, and allele fractions, and compared it with mutation counts, mutational signatures, and their combination using elastic net and XGBoost. Rebuilt the MuAt mutation transformer and trained it from scratch under the same compute budget as the comparison.

Findings. On 20 cancer types (n = 8,800), signatures plus Bio MAF v4 with XGBoost reached 0.738 balanced accuracy against 0.528 for the transformer, and the frozen model transferred to 913 PCAWG tumors. The gain depends on known cancer-gene annotations, which is itself informative about what the representation is encoding.

Paper: in review.