2026-09-01
RegimeFormer embeds 202.6M proteins in a fragile-to-adaptive coordinate and scores 407M residues. Spearman 0.698 on 217 substitution assays, matching the best supervised 0.700.
Protein language models already estimate the effect of a point mutation from sequence. Deep mutational scanning (DMS) can map that landscape densely, but only for a handful of proteins at a time. Public databases now hold hundreds of millions of sequences. The missing object is a protein-level coordinate: a way to say why the same class of substitution is tolerated in one background and destructive in another.
RegimeFormer supplies that coordinate. Each protein is placed on a continuous axis from fragile to adaptive, and the axis is then used as a prior for residue-level and substitution-level prediction.
RegimeAtlas starts from 424.8 million raw records drawn from UniProtKB, RefSeq, Ensembl, MGnify and related sources. After alphabet, length (at least 30 amino acids), low-complexity and 90% identity near-duplicate clustering, 202,556,313 non-redundant proteins remain, 47.7% of the raw set. All of them receive ESM-family embeddings and a protein-level regime index. High-resolution training and residue inference run on a diversity-preserving one-million-protein subset; 995,995 proteins pass, covering 407,048,356 residues. The production backbone is ESM2 t33 650M, with a streaming window of 1,022 residues, trained and scanned on 32 A100 GPUs over about five months.
For each residue and candidate mutant amino acid, the model fuses residue PLM features, wild-type and mutant embeddings, the protein regime vector, assay context, and optional structure or evolution features. Six regime-aware attention blocks process the fused input. Assay type enters through a FiLM adapter; regime state is mixed with a soft gate. Three heads emit a continuous effect, a predictive standard deviation, and a damaging / tolerated / beneficial classification. Residue fragility and adaptability are tails of the substitution distribution; residue uncertainty aggregates substitution-level standard deviations.
Full L×19 matrices are not stored. On-demand queries reproduce the offline reference at Pearson and Spearman 0.997, using 0.35% of full-matrix storage.
The coordinate is not a relabeling of taxonomy. On 24,918 proteins with complete covariates, adjusted mutual information with taxonomy is 0.0063. PLM features add 0.331 balanced accuracy over taxonomy, length and Pfam. Mean absolute Spearman against downstream perturbational readouts is 0.38 for the real regime, 0.08–0.11 above length, shuffled-embedding and composition controls.
High-fragility residues enrich at active sites (odds ratio 11.3), ligand-binding sites (6.1), interfaces (3.2) and disease missense positions (4.0). Mutation sensitivity falls along the axis (Pearson r = −0.71). Existing predictors fail along the same axis: ESM-650M's fragile-to-adaptive MAE gap is 1.38. A failure-forecast model that sees regime plus metadata reaches AUROC 0.861.
| Setting | Metric | RegimeFormer | Comparator |
| 217 substitution assays | Spearman ρ | 0.698 | AUTOSCIENTISTS 0.700; Kermut 0.662 |
| Full model | substitution Spearman | 0.723 | no mutant embedding 0.548; PLM-only 0.579; no regime 0.701 |
| Unseen protein / family / <30% identity | retained fraction | 82% / 62% / 38% | largest relative gain in the low-homology band |
| 5% screening budget | HitRate | 0.540 | strongest baseline 0.434 |
| PRISM drug response | R² | 0.833 random split | 0.038 unseen compounds |
Against Kermut on 87 proteins, RegimeFormer wins 81, with a mean paired gain of +0.036. A frozen post-cutoff check barely moves, from 0.591 to 0.585. Structural Spearman with AlphaFold pLDDT is 0.375; evolutionary Spearman with phyloP is 0.5895. On SciPlex3, Pearson rises from 0.253 (txPert/GEARS) to 0.294, with RMSE essentially unchanged.
The headline is not a new DMS leaderboard champion. In-distribution, dropping the regime coordinate costs only 0.022 Spearman. The coordinate earns its keep under unseen families, low homology and small experimental budgets: a 5% screen recovers about 24% more beneficial variants than the strongest baseline. Storing residue summaries and computing substitutions on demand is what makes a 200-million-protein map queryable.
The cell and drug transfer is a second-scale demo. Unseen cell lines still reach R² 0.746. Unseen chemistry does not. Use the atlas to rank residues and to forecast where existing variant-effect models will err. Do not treat it as a potency model for novel small molecules.
Sequence length is the production bottleneck: success stays above 99% through 1,000 residues and falls to 83.4% beyond 10,000. Unseen-compound drug response is an explicit failure (R² 0.038, Pearson 0.222). Low-homology proteins keep only about 38% of in-distribution substitution performance.
Several claims run ahead of the evidence. 0.698 versus 0.700 is a tie with the top displayed supervised aggregate. Mutant amino-acid identity, not the regime prior, carries the in-distribution signal. The expert-utility study measures an interface that adds coordinates and virtual DMS (HitRate@20 0.41 vs 0.29), not wet-lab confirmation of 407 million virtual maps. Code availability names a retained analysis environment and checkpoints, with no public repository in the main text. This is a bioRxiv preprint. The 202.6 million entries are protein-level coordinates; residue scans cover one million proteins. Treating those two layers as one proteome-wide mutational atlas overstates resolution.