ProTEUS: A mutation-resolved world model for closed-loop protein evolution
Yikun Zhang, Minghao Guo, Guangxue Xu, Peter Yichen Chen, Wojciech Matusik, Wengong Jin, Jingjing Chai, Kaiyi Jiang
Foundry Bioscience Inc. · Massachusetts Institute of Technology · Stanford University School of Medicine · The University of British Columbia · Northeastern University · Broad Institute of MIT and Harvard
Published November 9, 2026 · on the feed November 11, 2026 · ▲ 13 · arXiv 2611.13348
The abstract
Protein engineering requires navigating rugged fitness landscapes, where small sequence changes can produce large functional differences, but experimental screening can only sample a small fraction of sequence space. Thus, a central challenge is to efficiently traverse the large combinatorial space of mutations in silico. We introduce ProTEUS, a mutation-resolved world model for closed-loop protein evolution. Pretrained on deep mutational scanning assays, ProTEUS learns a joint representation of sequence, evolutionary, and structural information using MutationFormer, a mutation-biased attention architecture. Variants are represented as states in this joint latent space, and mutations define actions whose functional consequences are predicted. To support exploration of many mutants, we developed a fast, physics-constrained all-atom predictor that utilizes explicit geometric constraint enforcement to achieve >10 fold increase in speed than AlphaFold3 with guaranteed physical validity. Lastly, ProTEUS refines its prediction and understanding of protein activity through closed loop experimental feedback to guide evolution toward a desired phenotype. ProTEUS surpasses established models in both zero-shot and supervised fine-tuning mutation prediction tasks. We experimentally demonstrate its utility in two closed-loop multi-objective engineering campaigns on aldehyde dehydrogenase and levansucrase, yielding variants with >10 fold higher activity than wild type. ProTEUS provides a world-model framework that connects virtual evaluation of mutational outcomes with experimental learning for efficient protein evolution.
The Brief
2:20A short podcast summary of the paper. Written and voiced by AI.
Transcript
ProTEUS. A mutation-resolved world model for closed-loop protein evolution. By Yikun Zhang and seven others, from Foundry Bioscience, MIT, and five other institutions. Published on the ninth of November.
Protein engineering means navigating a rugged fitness landscape. Small changes to a sequence can produce large differences in function, so the surface being walked over is full of cliffs rather than gentle slopes. Experimental screening is the honest way to read that surface, and it can only ever sample a small fraction of sequence space. The combinatorial space of possible mutations is simply too large. That leaves a central challenge: traversing that space efficiently in silico, before any pipette is picked up.
ProTEUS frames the problem as a world model, resolved at the level of single mutations. A variant is a state in a joint latent space. A mutation is an action, and the model predicts the functional consequence of taking it. That framing is what makes exploration tractable: instead of scoring sequences in isolation, the system has a space it can move through and a notion of what each move does. Its pretraining comes from deep mutational scanning assays, and the representation it learns is a joint one, holding sequence, evolutionary and structural information together.
Three pieces carry that. The representation is learned by MutationFormer, a mutation-biased attention architecture. Then, because exploring many mutants needs structures for many mutants, the authors built a fast physics-constrained all-atom predictor, which enforces geometric constraints explicitly. That enforcement buys two things: more than a ten fold increase in speed over AlphaFold3, and physical validity that is guaranteed rather than hoped for. The third piece closes the loop. Experimental feedback comes back in, and ProTEUS refines its predictions and its understanding of protein activity from it, steering the next round of evolution toward a desired phenotype.
On mutation prediction, ProTEUS surpasses established models both zero-shot and after supervised fine-tuning. The experimental demonstration is two closed-loop engineering campaigns, each with multiple objectives: one on aldehyde dehydrogenase, one on levansucrase. Both yielded variants with more than a ten fold higher activity than the wild type.
What the authors offer is a connection rather than a single model. Virtual evaluation of mutational outcomes on one side, experimental learning on the other, and a world-model framework that lets each inform the next cycle of the other.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free