ProTEUS Paper

ProTEUS A mutation-resolved world model
for closed-loop protein evolution

Yikun Zhang1,5#,6#,*, Minghao Guo2,*, Guangxue Xu3, Peter Yichen Chen4,
Wojciech Matusik2, Wengong Jin5,6, Jingjing Chai1, Kaiyi Jiang1,7,†

1 Foundry Bioscience2 MIT3 Stanford4 UBC5 Northeastern6 Broad Institute7 Princeton

* Equal contribution# Present addresses† Corresponding author

Paper Code ↗ VideoBibTeX

ProTEUS is a mutation-resolved world model for closed-loop protein evolution. It predicts mutation effects, selects variants for wet-lab experiments, and uses measured function to guide the next round.

14.03×SacB · 90 variants / 3 rounds6.85×ALDH · 150 variants / 5 rounds

Best screening activity relative to wild type · 30 variants per round

Watch the video 2 min 23 sec
Download video ↓

Recursive improvement with wet-lab feedback

Figures 4–5 ↗

30 variants per round

Rank candidates, measure their activity, and refit the target predictor for the next batch.

Refitting uses accumulated measurements; pretrained representations stay fixed.

Levansucrase SacB

14.03× WT screening activity

90 variants · 3 rounds

G23A · Q34N · G240W

A second lineage reached 12.38× WT screening activity.

View larger ↗

Aldehyde dehydrogenase ALDH

6.85× WT screening activity

150 variants · 5 rounds

S103V · A116K · H332N · L358I

View larger ↗

Cumulative top 10 through each round (Figures 4C and 5B). Dots show variant means; variants can recur. Violins use Gaussian KDE (bandwidth factor 0.35). Activity is relative to WT.

A mutation-resolved world model

Figure 1 ↗

Pretraining on functional measurements

Figure 2 ↗

1.32M variant measurements · 161 protein backgrounds.

A scaling law for functional data

80× training data56.3% lower test error
Test MSE and Spearman correlation as functional training data increases. Values are described alongside the chart and in the paper.

Test MSE: 0.262 → 0.115 on held-out variants within pretraining assays; model architecture fixed.

Mean zero-shot Spearman for all 15 models; selection Spearman, AUROC, AUPR and top-K hit fractions with SEM.

Zero-shot prediction and first-round selection.

19 prediction assays; 13 single-mutant selection assays. No target fitness labels. Mean ± SEM. MULTI-evolve: the paper’s rank-fusion adaptation (Figures 2E–F).

Per-assay data ↓

Fast structures for each candidate

Figure 3 ↗

200 diffusion steps → 2

A diffusion model initialized from Boltz-1 is post-trained through differentiable Gauss–Seidel projection.

AF3 uses 200 steps in this comparison. Validity is assessed against seven specified constraint families.

Inference is 7.8 to 10.8 times faster than AlphaFold 3 across five atom-count bins. Values are described alongside the chart and in the paper.

7.8–10.8× faster

Than AlphaFold 3 across five CASP15 atom-count bins on an NVIDIA RTX A6000.

Projected structures passed all seven specified validity checks in the reported benchmarks.

Citation

@misc{zhang2026proteusmutationresolvedworldmodel,
  title = {ProTEUS: A mutation-resolved world model for closed-loop protein evolution},
  author = {Yikun Zhang and Minghao Guo and Guangxue Xu and Peter Yichen Chen and Wojciech Matusik and Wengong Jin and Jingjing Chai and Kaiyi Jiang},
  year = {2026},
  eprint = {2611.13348},
  archivePrefix = {arXiv},
  primaryClass = {q-bio.BM},
  doi = {10.48550/arXiv.2611.13348},
  url = {https://arxiv.org/abs/2611.13348}
}