On November 9, 2026, a multi-institutional research team unveiled ProTEUS, an open computational world model designed to navigate epistatic fitness landscapes and steer closed-loop directed protein evolution. By pairing mutation-focused transformer representations with rapid physics-constrained structural modeling and iterative wet-lab experimental feedback, the platform achieves more than ten-fold activity gains across complex multi-site enzyme engineering campaigns while dramatically lowering computational overhead.
Navigating Rugged Fitness Landscapes in Protein Engineering
Modern protein science relies heavily on directed evolution to engineer enzymes, binding scaffolds, and therapeutic proteins with properties that rarely emerge under natural selection. However, wet-lab mutagenesis and deep mutational scanning (DMS) can sample only a microscopic fraction of the astronomical sequence space—roughly 20 raised to the power of the sequence length. While single-point mutations can be evaluated in high-throughput screens, synergistic or antagonistic interactions across multiple substituted sites—known as epistasis—create rugged fitness landscapes where single-site improvements fail to predict multi-mutation trajectories.
Computational screening methods have historically bridged this gap using zero-shot sequence likelihoods or structural compatibility scoring. Yet models trained solely on evolutionary sequence distributions frequently fail to prioritize functional activity under non-physiological conditions, as natural evolution selects for organismal survival rather than specific industrial or catalytic metrics. ProTEUS, developed collaboratively by researchers at Foundry Bioscience, the Massachusetts Institute of Technology, Stanford University School of Medicine, the University of British Columbia, Northeastern University, the Broad Institute, and Princeton University, reframes protein design as a world model where variants represent latent states and mutations constitute functional actions.
MutationFormer and Physics-Constrained Structural Inference
At the center of ProTEUS is MutationFormer, a 12-layer transformer architecture engineered to overcome the representation dilution common to whole-sequence pooling. In typical protein representations, averaging residue embeddings across hundreds of amino acids obscures the sparse substitutions that govern biochemical divergence. MutationFormer aligns amino acid sequence, evolutionary multiple sequence alignments (MSAs), and structural features residue by residue, using an explicit binary mutation mask and mutation-biased attention. In ablation experiments, setting the mutation-bias strength to 10 elevated the model Spearman rank correlation from 0.606 to 0.840, demonstrating that focusing model capacity on substituted positions while retaining contextual geometry is critical for fitness estimation.
To evaluate higher-order mutant candidates, the platform requires candidate-specific atomic coordinates without the computational bottleneck of standard diffusion pipelines. Rather than running hundreds of denoising steps through tools like the AlphaFold Protein Structure Database pipeline or AlphaFold 3, the authors implemented a physics-constrained all-atom predictor based on Boltz-1. The system pairs two denoising updates with a differentiable Gauss–Seidel projection that enforces seven stereochemical constraint families, including tetrahedral chirality, double-bond planarity, ligand internal distances, and nonbonded inter-chain separations.
This hybrid geometric framework achieved a median 9.4-fold speedup over 200-step Boltz-1 on the CASP15 benchmark and delivered between 7.81-fold and 10.81-fold faster runtimes than AlphaFold 3 across atom-count bins. In structural validation on complex PDB 7WCF, the predictor resolved unphysical steric clashes, expanding a compressed 1.41 Å ligand-side chain distance to a physically plausible 4.44 Å while maintaining competitive global complex LDDT scores (0.6321).
Pretraining on ProteinGym and Empirical Scaling Behavior
The ProTEUS latent space was pretrained on 182 substitution deep mutational scanning assays from the ProteinGym repository, spanning 161 protein backgrounds across human, prokaryotic, eukaryotic, and viral origins. To prevent data leakage, the authors applied rigorous homology filtering, excluding any assay matching downstream evaluation targets at 30% sequence identity or TM-scores of 0.5 or greater. On held-out test data, ProTEUS attained a Spearman correlation of 0.933 and a mean squared error (MSE) of 0.115, outperforming the PoET-2 baseline (0.861).
The study demonstrated clear power-law data scaling within biological activity representations. Expanding the functional pretraining dataset 80-fold reduced test MSE by 56.3% (from 0.262 to 0.115), corresponding to an approximately 12.1% error reduction with each doubling of supervised functional data. In higher-order mutation generalization benchmarks across 15 multi-mutant datasets (K-vs-rest evaluation), ProTEUS attained a mean Spearman correlation of 0.448 compared to 0.295 for MULTI-evolve, achieving superior accuracy in 80% of test comparisons.
Closed-Loop Wet-Lab Evolution in Bacterial Systems
To validate the world-model framework in live experimental pipelines, the investigators coupled ProTEUS to an agentic workflow that manages candidate generation, LLM-based physicochemical rationale synthesis, and automated ranking updates based on plate-reader assays in Escherichia coli.
In a multi-objective engineering campaign on Bacillus subtilis levansucrase (SacB)—optimizing the product of enzymatic activity and expression—an initial zero-shot batch of 30 single mutants identified Q34N, which displayed 4.50-fold higher catalytic activity than wild type. Subsequent rounds guided by experimental feedback identified double mutants Q34N:G240W (8.94-fold) and Q34N:S359F (8.91-fold), culminating in a triple mutant (G23A:Q34N:G240W) achieving 14.03-fold wild-type activity. Notably, all three substituted residues in the top-performing variant were uncommon among natural homologs (introduced-residue frequencies below 3.8%), illustrating the model capacity to explore non-consensus functional variations without phylogenetic conservatism.
A second campaign on Cupriavidus pinatubonensis aldehyde dehydrogenase (ALDH) evaluated 150 variants across five successive rounds of 30 nominations. Beginning with zero-shot identification of L358I (2.02-fold wild-type activity), the model iteratively incorporated lysate absorbance data to discover S103V:A116K:H332N:L358I. This quadruple mutant demonstrated a 6.85-fold increase in total activity and a 7.90-fold enhancement in concentration-normalized catalytic rate, confirmed via purified protein kinetic assays. Structural mapping revealed that all four mutations were located more than 13 Å away from annotated active-site residues, underscoring the system ability to capture long-range allosteric interactions that evade rational active-site remodeling.
Governance, Biosecurity, and Institutional Research Oversight
The transition from open-loop generative design to closed-loop autonomous evolution marks an important shift for academic research administration. As biological foundation models become capable of directing automated synthesis and screening workflows with small experimental budgets, research institutions face new compliance and risk-assessment demands.
University compliance offices and an Institutional Biosafety Committee (IBC) must consider how agentic closed-loop platforms interact with dual-use research of concern (DURC) guidelines. Because ProTEUS can rapidly optimize enzyme activity and expression across multi-mutant pathways in under 100 physical assays, safety protocols must account for epistatic gains of function that cannot be anticipated from single-mutation screening. Transparent documentation through comprehensive model cards, rigorous screening of automated synthesis orders, and computational audit trails will be essential as universities integrate closed-loop machine learning into biotechnology pipelines.
Sources
- arXiv:2611.13348 preprint: Y. Zhang, M. Guo, G. Xu, P. Y. Chen, W. Matusik, W. Jin, J. Chai, and K. Jiang, “ProTEUS: A mutation-resolved world model for closed-loop protein evolution,” submitted November 9, 2026. URL: https://arxiv.org/abs/2611.13348 (checked 2026-11-14).
- ProteinGym benchmark repository: Functional deep mutational scanning benchmarks for protein fitness prediction. URL: https://proteingym.org (checked 2026-11-14).
