A System One Approach to General Protein Evolution
Abstract
Protein engineering requires deciding which mutation to test next and when to stop. Existing approaches often require target-family measurements, encode candidates separately, or return scores that are not calibrated to the assay. We introduce Pev, a general mutation policy that explores a System One approach to protein evolution: learning from experiments across families to make decisions in a single parallel query. Given a parent sequence and an assay context, Pev encodes the parent once and returns calibrated probabilities for choosing an edit or stopping, together with each edit's probability of improving the assay outcome. One policy covers peptides, small folded domains, and antibodies across binding and stability objectives, without fitting model weights to the target family. In ten-well screens on held-out families, Pev finds 3.3× as many beneficial mutations as random selection and 1.8× as many as zero-shot ESM-2. It also outperforms a ridge surrogate refitted on target measurements in closed-loop search on the held-out GB1 landscape. Post-hoc calibration improves probability reliability without changing decisions, while shared encoding provides over 800× the throughput of per-mutant encoding at 1,000 candidates per request. These results suggest that a System One approach could support efficient experimental decisions across protein families, with Pev providing an initial demonstration.
Related
- Limitations of general protein language models and public TCRpMHC data in training specificity prediction models
- Carboxysome-Inspired Protein Coacervates for Light-Driven CO2 Reduction and H2 Evolution
- A General Harness for Protein Foundation Model Fitness Prediction
- When Metabolism Became Messaging: The Stepwise Evolution of the GABAergic Signaling System
- Experimental evolution reveals genetic routes for adaptive loss of the antibacterial type VI secretion system
BibTeX
@misc{wen2026system,
title = {{A System One Approach to General Protein Evolution}},
author = {Yibo Wen and Han Liu},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.05396},
eprint = {2611.05396},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.05396},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A System One Approach to General Protein Evolution
AU - Yibo Wen
AU - Han Liu
PY - 2026
DA - 2026/11/01
PB - arXiv
DO - 10.48550/arxiv.2611.05396
UR - https://arxiv.org/abs/2611.05396
AB - Protein engineering requires deciding which mutation to test next and when to stop. Existing approaches often require target-family measurements, encode candidates separately, or return scores that are not calibrated to the assay. We introduce Pev, a general mutation policy that explores a System One approach to protein evolution: learning from experiments across families to make decisions in a single parallel query. Given a parent sequence and an assay context, Pev encodes the parent once and returns calibrated probabilities for choosing an edit or stopping, together with each edit's probability of improving the assay outcome. One policy covers peptides, small folded domains, and antibodies across binding and stability objectives, without fitting model weights to the target family. In ten-well screens on held-out families, Pev finds 3.3× as many beneficial mutations as random selection and 1.8× as many as zero-shot ESM-2. It also outperforms a ridge surrogate refitted on target measurements in closed-loop search on the held-out GB1 landscape. Post-hoc calibration improves probability reliability without changing decisions, while shared encoding provides over 800× the throughput of per-mutant encoding at 1,000 candidates per request. These results suggest that a System One approach could support efficient experimental decisions across protein families, with Pev providing an initial demonstration.
ER -