MarinDNA: protocol comparison

Each cell below is AUPRC − AUPRC, in percentage points. Green = scores higher; red = the reverse; yellow = no meaningful change.

Protocol definitions (each score is per-variant; transforms make higher = more variant-effect):

• LLR — log-likelihood ratio of mutant vs. reference, averaged across forward and reverse-complement scores. Sign-flipped for Mendelian (−LLR) and absolute-valued for Complex (|LLR|). Leaderboard default.
• NucDep (nucleotide dependency) — per-position Jensen-Shannon divergence between the model's mutant and reference next-token distributions, averaged first across positions downstream of the variant, then across forward and reverse-complement scores.
• LLR-FWD / NucDep-FWD — same transforms applied to the forward strand only (no reverse-complement averaging). Surfaced here for strand-symmetry exploration; not used in the leaderboards.

Cells aggregate across the same matched groups the [ leaderboard](../leaderboards/) uses — only the score column changes.

Color scale clamps at ±10 percentage points. Sort by clicking any column header — sort follows the dataset's natural rank axis after a dataset switch. The model column links to the model card on the Models page.