Self-evolving AI agents for protein discovery and directed evolution

Abstract

Protein scientific discovery is bottlenecked by the manual orchestration of information and algorithms, while general agents are insufficient in complex domain projects. VenusFactory2 provides an autonomous framework that shifts from static tool usage to dynamic workflow synthesis via a self-evolving multi-agent infrastructure to address protein-related demands. It outperforms a set of well-known agents on the VenusAgentEval benchmark, and autonomously organizes the discovery and optimization of proteins from a single natural language prompt.

Figures (6)

• Figure 1: The VenusFactory2 architecture and evaluation.a, Evolution of interaction paradigms of AI tools in protein science, shifting from static programmatic interfaces to dynamic, natural language-driven autonomous orchestration. b, The multi-agent architecture comprising five specialized functional roles for collaborative reasoning and execution. c, The four-phase operational workflow illustrates the iterative process of task decomposition and tool invocation. d, Comparative performance evaluation of VenusFactory2 against a suite of state-of-the-art general-purpose LLMs and domain-specific agents across three complexity tiers of the VenusAgentEval benchmark.
• Figure 2: Implementation mechanisms and biological validation.a, Overview of the integrated computational infrastructure, stratified into four functional quadrants ranging from structural mining to generative design. b, Workflow of the Research module for generating structured execution protocols via hierarchical retrieval and constraint verification. c, Mechanism of autonomous tool instantiation, exemplified by the runtime synthesis of an allergenicity predictor. d, Validation in targeted enzyme retrieval, recapitulating the identification of KbPETase via semantic search and multiparametric filtration ($T_m$, pH). e, Validation in VHH antibody directed evolution, demonstrating adaptive code generation to model combinatorial epistatic effects utilizing single-site priors.
• Figure 7: Performance distribution across datasets. Violin plots showing the normalized evaluation metrics (0 to 1) for all trials conducted by the agent. The red star indicates the optimal configuration identified by the agent for each respective dataset.
• Figure 8: Agent optimization trajectory. Scatter points represent individual training trials. The red solid line indicates the cumulative best performance achieved by the agent over the progression of trials. The grey dashed line denotes a baseline performance derived from standard human-selected configurations.
• Figure 9: Agent trial allocation. Stacked bar chart illustrating the total number of trials the agent allocated to each pre-trained model architecture, categorized by the selected fine-tuning method (freeze vs. ses-adapter).