Vijay Vankadaru

Vijay Vankadaru

MS Data Science, UC Berkeley
Mechanistic interpretability of LLMs and VLMs
Machine Learning Engineer, DASION

Benchmarking Somatic Mutation Representations for Cancer Typing and HRD Prediction

In reviewUniversity of Maryland Institute for Health Computing

A. Ge, V. Vankadaru, X. Liu, J. Almeida, J. B. Balasubramanian, B. A. Maron

With Aaron Ge, Xingyan Liu, Jonas Almeida, Jeya Balaji Balasubramanian, and Dr. Bradley A. Maron at the University of Maryland School of Medicine.

Machine learning on somatic mutations differs less in the algorithm than in how the mutations are described: a count, a pattern of sequence contexts, gene-level annotations, or individual mutations read by a neural network. We compared four representations with elastic net and XGBoost against a MuAt-compatible transformer trained from scratch under a fixed compute budget. On 20 cancer types, mutational patterns 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.

Cite as: Ge, A., Vankadaru, V., Liu, X., Almeida, J., Balasubramanian, J. B., & Maron, B. A. Benchmarking Somatic Mutation Representations for Cancer Typing and HRD Prediction. In review.