Michele Waters

MetaHarmonizer improves metadata harmonization and addresses LLM contamination in biomedical data, outperforming existing methods on benchmark tasks.

This paper evaluates generalist large language models (LLMs) in tumor genomic interpretation tasks, finding they can complement or outperform task-specific models in some cases.

The paper develops Mosaic, a machine learning system that classifies tumors into detailed subtypes and identifies molecular phenotypes using histopathologic and genomic data from a large patient cohort.

This paper develops a multimodal transformer model using H&E images and text reports to predict recurrence scores in hormone receptor-positive breast cancer, achieving high accuracy and generalizability.

The paper uses machine learning to identify cancer driver mutations, finding that methods incorporating protein structure or functional genomic data outperform those using only evolutionary data, and validates their predictions using real-world patient cohorts.