MarinDNA Observatory

Public, version-controlled benchmarks and interpretation for genomic language models trained under MarinDNA. Two pillars: how well each model ranks variants, and what each model has learned.

Benchmarks

Variant-effect leaderboards:

A model family's AUPRC depends on which score you compute it from — the protocol pages compare scoring approaches head-to-head on the same models and dataset:

Interpretation

Visual analyses of what the trained models have internalized:

Reference