Agents · arXiv 2026

VenusRAR

Rank-and-Reason — a two-stage multi-agent framework that turns PLM ensembles into wet-lab-ready mutant shortlists, with biophysical chain-of-thought auditing.

Yang Tan*, Yuanxi Yu*, Can Wu, Bozitao Zhong, Mingchen Li, Guisheng Fan, Jiankang Zhu, Yafeng Liang, Nanqing Dong, Liang Hong
(* equal contribution)

arXiv Code ProteinGym

Problem

Zero-shot mutation predictors often return statistically confident rankings that still violate basic biophysical constraints. Selecting candidates for the wet lab then falls to manual expert review — slow, subjective, and hard to scale.

Rank-and-Reason

VenusRAR automates that workflow in two stages:

  • Rank-Stage — a Computational Expert and Virtual Biologist aggregate a context-aware multi-modal ensemble of PLM signals, reaching Spearman ρ ≈ 0.551 on ProteinGym (vs. 0.518 for the prior VenusREM-class baseline).
  • Reason-Stage — an agentic Expert Panel uses chain-of-thought reasoning to audit candidates against geometric and structural constraints, improving Top-5 Hit Rate by up to 367% on ProteinGym-DMS99.

Wet-lab validation on Cas12i3 nuclease reported a 46.7% positive rate (14/30), including novel mutants with 4.23× and 5.05× activity gains.

Relation to other Venus tools

VenusRAR sits above single-model scorers such as VenusREM, ProSST, and ProtSSN: it ensembles their signals, then applies LLM agents for biophysical triage. Deployment paths can go through VenusFactory2.

Citation

@misc{tan2026venusrar,
  title={Rank-and-Reason: Multi-Agent Collaboration Accelerates Zero-Shot Protein Mutation Prediction},
  author={Tan, Yang and Yu, Yuanxi and Wu, Can and Zhong, Bozitao and Li, Mingchen and Fan, Guisheng and Zhu, Jiankang and Liang, Yafeng and Dong, Nanqing and Hong, Liang},
  year={2026},
  eprint={2602.00197},
  archivePrefix={arXiv},
  primaryClass={q-bio.QM},
  url={https://arxiv.org/abs/2602.00197}
}