The abstract is read at the publisher; the summary is JClub's.
On Swiss-Prot remote-homology retrieval, embedding-based protein language model nearest-neighbour transfer recovers at most 15% of homologs at a 1% false-positive rate, versus 85% for MMseqs2.
- Why it matters: Embedding annotation lacks a null model, so cosine thresholds for controlling false positives are not well grounded, unlike alignment tools that use E-values.
- What they did: The study benchmarks final-layer, mean-pooled embeddings from twelve pLMs against MMseqs2 and Foldseek on Swiss-Prot and SCOPe, using per-model thresholds fixed at 1% false positives and synthetic decoy ladders.
- The result: ESM-2 and ESM-C show extreme anisotropy (fewer than 15 effective dimensions out of 640–2560), and ZCA whitening fitted on the reference database boosts threshold-free Top-1 superfamily retrieval, e.g., ESM-2 650M on SCOPe from 59% to 88%.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.