Are generalist models better for biology than specialized models?

According to a new case study, a single generalist AI model now outperforms specialist models trained on specific cancer types by up to 8% on some indications.

Bioptimus just dropped results for M-Optimus-1, their multimodal foundation model that reads across three biological layers simultaneously: tissue images (H&E pathology), bulk gene expression, and spatial transcriptomics.Interestingly, when they trained one model on diverse tissue types instead of building separate models per cancer, it transferred knowledge between indications.

In ther words, patterns learned from lung tissue improved predictions in colon and head and neck cancers. Even on cancer types the model had never seen during training, like kidney, skin, bone marrow, it outperformed models specifically trained on those tissues.

This matters because it challenges a deep assumption in computational biology: that biology is too varied across organs for a generalist approach to work.

Turns out, immune cells, vasculature, and tumor microenvironment interactions follow shared rules across the body. A model that sees enough diversity learns the grammar, not just the vocabulary of individual diseases, so to speak.

The practical angle is that M-Optimus-1 can predict spatial gene expression (~20,000 genes) from a standard $10 H&E slide, reconstructing information that would cost ~$10,000 per specimen to measure directly with spatial transcriptomics.

If that holds up in validation, it means retrospective analysis of historical clinical trial cohorts becomes vastly cheaper.

Worth noting: performance scales with data and hasn't plateaued yet, which is the kind of curve you want to see if you're building a platform.

Still early. The real test will be whether these in silico spatial maps prove reliable enough for actual biomarker discovery and patient stratification decisions.

But the generalist-beats-specialist result is one of those findings that, if it replicates broadly, reshapes how the field thinks about building these models.

Image credit: Bioptimus