Biological foundation models, from DNA to whole cells

"Foundation model" has arrived in biology, and it now spans three scales: the sequence of a single protein, the genome of an organism, and the state of a whole cell.

Proteins: the ESM family

EvolutionaryScale's ESM models treat protein sequences as a language. The latest open-weight ESM C models (300M, 600M and 6B parameters) are drop-in upgrades from ESM2 — a 300M ESM C rivals the old 650M model with faster inference. Alongside them, the generative ESM3 reasons jointly over sequence, structure and function, enabling controllable design.

Genomes: Evo 2

Arc Institute's Evo 2, published in Nature, is trained on nine trillion DNA bases across every domain of life with a one-million-token context window. It predicts the functional impact of variants — including noncoding and clinically significant mutations — without task-specific fine-tuning, and it generates coherent sequences at genome scale. Crucially, it is fully open: weights, code and the OpenGenome2 dataset.

Cells: virtual cell models

The newest frontier is the cell itself. Arc's State model is trained on 167 million observational and over 100 million perturbed cells, and predicts how populations respond to drugs, cytokines and genetic perturbations. It is one of the first models to consistently beat simple linear baselines on perturbation prediction — a reminder that this field is young and benchmarking matters.

Where this is heading

Each scale is becoming programmable. The interesting work now is connecting them: using a DNA model to propose edits, a protein model to design the resulting machinery, and a cell model to predict the phenotype. Tools like scvi-tools remain the glue that turns raw single-cell data into something these models can learn from.