ProTEUS Turns Protein Mutation Search into a Closed-Loop World Model

Introduction

Protein engineering is often limited by a mismatch between the size of the search space and the amount of experimental data available. A protein may tolerate one substitution but fail after a nearby change, while combinations of individually modest mutations can produce large functional effects. Laboratory screening provides the most valuable evidence, yet it can inspect only a small part of the combinatorial landscape. ProTEUS approaches this problem as a closed-loop decision process rather than as a one-time sequence-ranking task.

What the system does

• It models mutations explicitly. ProTEUS is pretrained on deep mutational scanning assays and uses MutationFormer, an attention architecture designed to emphasize mutation-related information. Sequence, evolutionary context, and structural signals are mapped into a shared latent representation. In this space, a protein variant is treated as a state and a mutation as an action that can move the state.
• It predicts functional consequences. The goal is not merely to determine whether a sequence resembles known proteins. ProTEUS attempts to estimate what may happen after a particular mutation or combination of mutations. According to the paper, it outperforms established models on both zero-shot and supervised fine-tuning mutation-prediction tasks.
• It supports faster structural exploration. The authors introduce an all-atom predictor with explicit geometric constraint enforcement. The stated design objective is to produce physically valid structures while greatly reducing the cost of evaluating many candidates. The abstract reports a speed increase of more than tenfold relative to AlphaFold3.
• It learns from each experimental round. After candidates are tested, measured activity and other objectives are fed back into the system. The model can then adjust its understanding of the landscape and select candidates for the next round. The authors demonstrate this workflow in two multi-objective campaigns involving aldehyde dehydrogenase and levansucrase, reporting variants with more than ten times the activity of the wild type.

Why it matters

The central contribution is a workflow that connects virtual mutation evaluation, structural plausibility, and laboratory feedback. A state-action formulation is a natural fit for directed evolution: the current protein is the state, a mutation is an intervention, and the measured phenotype becomes evidence for the next decision. This can help models move beyond static sequence scoring and focus resources on experiments that are expected to be informative.

The available material also leaves important questions open. The abstract does not provide benchmark tables, the number of experimental rounds, candidate counts, or details on how competing objectives were balanced. The reported speed and activity gains therefore need to be assessed against the full paper, implementation, and experimental protocols. ProTEUS should be viewed not as a replacement for laboratory validation, but as a framework intended to make every experimental round more targeted and useful.

Source: Hugging Face Daily Papers