OmicsLM preprint reports gains in interpreting gene expression measurements

Researchers report that OmicsLM, a 9B multimodal language model, outperformed evaluated alternatives across all six variants of a benchmark testing interpretation and comparison of gene expression measurements. The preprint describes a model that combines compact numerical representations of biological samples with natural-language instructions, allowing multiple samples in the same context. Training included 20 billion tokens of continued pretraining and 5.21 million supervised conversations. The researchers introduced GEO-OmicsQA, a benchmark containing 3,000 questions drawn from real studies. Tests favored continuous numerical representations over text-only gene lists, although OmicsLM led on three of four held-out annotation tasks.