ESM-2 35M frozen residue-mean embeddings + fixed ridge (Rewire): report

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 1 configuration
• 1 evaluation
• 2 results
• 4 evidence claims
• 1 use case

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• ESM-2 35M frozen residue-mean embeddings + fixed ridge (Rewire) (FLIP2 Rhomax by_wild_type): facets.method_types: Facets: method types ["foundation_model"] (Frozen ESM-2 embeddings with a ridge probe)
• Evaluated ESM-2 checkpoint belongs to ESM-2: Family ESM-2 (model_configuration.checkpoint; model.name)
• Evidence summary: Assess rhodopsin wavelength prediction across sequence backgrounds: Summary RhoMax, which adds structural binding-pocket inputs, had the lowest wavelength error on each of its own wild-type holdout splits. In separate Rewire runs on the FLIP2 split, two amino-acid composition probes outranked two frozen ESM-2 probes, and both ESM-2 probes fell below a constant control. Averaged over four RhoMax splits, mean absolute error was 10.45 nm for RhoMax, 13.46 nm for RhoMax plus retinal and 14.90 nm for the BLASSO comparator, with standard deviations of 10.9 to 11.7 nm; per-split RhoMax error ran from 7.28 to 16.68 nm. In electronvolts the three methods are closer, and on split 3 all three print the same mean absolute error of 0.045 eV. On the FLIP2 Rhomax by_wild_type split of 184 held-out sequences, Rewire's runs gave a Spearman correlation of 0.418 for a 40-feature amino-acid composition probe and 0.418 for a 22-feature one, against -0.146 for a frozen ESM-2 8M probe and -0.222 for a 35M probe; NDCG was 0.955, 0.955, 0.896 and 0.907 against 0.921 for a training-mean control, whose Spearman is undefined. The two split definitions are different and their scores must not be pooled. The RhoMax values are reported by its authors with no independent reproduction, and the Rewire runs are single runs with no seed variability, with the 35M checkpoint chosen after the 8M outcome was known. (uc20260930-source-rhomax-2024: per-split and mean rows of the wild-type holdout table, nm and eV columns; rewire-local-20260920 and rewire-local-20260921 run outputs: metrics.spearman and metrics.ndcg)
• Relevance of FLIP2 Rhomax by_wild_type to "Assess rhodopsin wavelength prediction across sequence backgrounds": Assessed by FLIP2 Rhomax by_wild_type (rewire-local-20260921-instructions-esm2-35m: Matched Rhomax evaluations; Procedure and evidence; Re-run or inspect; rewire-local-20260921-source-esm2-35m: /metrics; /coverage; /model_configuration; /protocol_configuration; /execution; /provenance; rewire-local-20260921-source-esm2-8m: /metrics; /coverage; /model_configuration; /protocol_configuration; /execution; /provenance; rewire-local-20260921-source-composition22: /metrics; /model_configuration; /protocol_configuration; rewire-local-20260920-source-flip2-train-mean: /metrics; /coverage; /model_configuration; rewire-local-20260920-source-flip2-composition: /metrics; /coverage; /model_configuration; /provenance)

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.