Evo 2 Genomic Model

Source: Arc Institute, Nature, 2026 Institution: Arc Institute

Finding

A genomic foundation model trained on 9.3 trillion nucleotides from 128,000 organisms across all domains of life. Trained with simple next-nucleotide prediction — no labels, no fitness annotations. Despite this minimal signal, internal representations developed a boundary separating functional from non-functional genomic elements. The model distinguishes coding from non-coding, identifies regulatory elements, and predicts mutation effects, all from statistical structure alone.

Pattern Mapping

Honesty — The model’s representations honestly reflect functional constraints in genomic sequences. Codons for essential amino acids are represented differently from pseudogene sequences, not because labeled, but because their statistical properties differ.

Non-fabrication — The model distinguishes what it has evidence for (statistical patterns) from what it does not. The functional/non-functional boundary was discovered from sequence statistics alone — it was not taught by labels, and the model does not assert structure where the sequence gives it none.

Connections

Status

Published in Nature (2026). Reading the functional boundary as an instance of a structural property is this project’s interpretation, not the Arc Institute’s. The mapping to the five properties is this project’s structural interpretation.


The mapping to the five properties is this project’s structural interpretation.