Rank-and-Reason: Multi-Agent Collaboration Accelerates Zero-Shot Protein Mutation Prediction

TL;DR

The paper tackles zero-shot protein mutation prediction in low-resource settings, where traditional PLMs often produce high-confidence outputs that violate biophysical constraints. It introduces VenusRAR, a two-stage, multi-agent framework that first ranks mutations with a context-aware ensemble (Rank-Stage) and then audits them with physics-informed reasoning (Reason-Stage) to filter out biophysically invalid candidates. The approach achieves state-of-the-art global ranking on ProteinGym ($\rho=0.551$) and substantially improves selection quality under budget constraints (Top-5 Hit Rate up to $367\%$ improvement), with strong wet-lab validation on Cas12i3 (46.7% hits, including two super-mutants with $5.05$- and $4.23$-fold activity). This work demonstrates that integrating interpretable scientific reasoning with multi-modal predictions can translate computational gains into tangible experimental success, offering a practical, plug-and-play framework for autonomous protein design and beyond.

Abstract

Zero-shot mutation prediction is vital for low-resource protein engineering, yet existing protein language models (PLMs) often yield statistically confident results that ignore fundamental biophysical constraints. Currently, selecting candidates for wet-lab validation relies on manual expert auditing of PLM outputs, a process that is inefficient, subjective, and highly dependent on domain expertise. To address this, we propose Rank-and-Reason (VenusRAR), a two-stage agentic framework to automate this workflow and maximize expected wet-lab fitness. In the Rank-Stage, a Computational Expert and Virtual Biologist aggregate a context-aware multi-modal ensemble, establishing a new Spearman correlation record of 0.551 (vs. 0.518) on ProteinGym. In the Reason-Stage, an agentic Expert Panel employs chain-of-thought reasoning to audit candidates against geometric and structural constraints, improving the Top-5 Hit Rate by up to 367% on ProteinGym-DMS99. The wet-lab validation on Cas12i3 nuclease further confirms the framework's efficacy, achieving a 46.7% positive rate and identifying two novel mutants with 4.23-fold and 5.05-fold activity improvements. Code and datasets are released on GitHub (https://github.com/ai4protein/VenusRAR/).

Figures (5)

• Figure 1: The VenusRAR multi-agent framework. (a) Users define engineering objectives and multi-modal contexts. (b)Rank-Stage: A Computational Expert aggregates raw PLM scores, while a Virtual Biologist dynamically calibrates ensemble weights based on structural and evolutionary data quality. (c)Reason-Stage: A Virtual Expert Panel conducts CoT-based auditing to filter biophysically invalid candidates. (d) Wet-lab experimental validation confirming high-fold activity improvements in Cas12i3.
• Figure 2: Comparison of Rank-Stage (solid bars) versus Reason-Stage (hashed bars) across budgets $N \in \{10, \dots, 40\}$. Panels display Top 1% Precision (top), Top 5 Hit Rate (middle), and Normalized Max Score (bottom). The red dashed line marks the Static-Ensemble.
• Figure 3: Relative improvement on reasoning capability of DeepSeek-Reasoner vs. DeepSeek-Chat, averaged across different budgets ($N \in \{10, \dots, 40\}$). Error bars denote the standard deviation.
• Figure 4: Wet-lab validation on Cas12i3. (a) The 3D distribution of the identified hits. Beneficial mutations are dispersed across diverse structural domains, verifying the system's global search capability. (b) Activity landscape of the 30 candidates relative to Wild Type (WT=1.0). VenusRAR achieves a $46.7\%$ positive rate ($14/30$), with the top variant Mutant A ($5.05$-fold).
• Figure 5: Exploration frontier of Normalized Max Score vs. Selection Budget ($N$). Solid lines represent Reason-Stage, and dashed lines represent Rank-Stage. Error bars denote standard deviation. Stronger models (e.g., Deepseek-Reasoner) demonstrate higher exploration efficiency.