Research map / Large language models / Agents, code and math
KnowEvo: Knowledge Evolution for Protein Optimization
Zijie Xing, Xingyue Liu, Runze Wang, Luoming Hu, Yanming Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KnowEvo, a framework that makes knowledge for LLM-driven protein optimization explicit, persistent, and reusable rather than implicit in weights or transient solution histories.
Problem
LLMs for scientific discovery such as protein optimization lack a mechanism for persistent, reusable knowledge accumulation; existing methods keep knowledge implicit in model weights or scattered across transient solution histories.
Approach
Makes accumulated optimization knowledge an explicit, persistent artifact that evolves across iterations and can be reused, instead of relying on parameter updates or trajectory refinement. Only the abstract was available, so finer mechanism details are unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evolve executable expertise blocks with KnowEvo | Root seed blocks · Rosetta stability | pass@1: 42.4 -> 50.5 | Improved |
| Evolve executable expertise blocks with KnowEvo | Root seed blocks · SaProt solubility | pass@1: 31.6 -> 51.5 | Improved |
| Replace Thompson-style allocation with uniform random node sampling (ablation) | KnowEvo · SaProt solubility | MRR: 67.8 -> 65.4 | Hurt |
| Replace sandbox agentic revision with single-step LLM revision (ablation) | KnowEvo · SaProt solubility | MRR: 67.8 -> 64.2 | Hurt |
| Replace mutation tree with chain-only growth (ablation) | KnowEvo · SaProt solubility | pass@3: 83.2 -> 54.1 | Hurt |
| Give the same sandboxed code agent to AdaEvolve | AdaEvolve · SaProt solubility | pass@1: 49.0 -> 42.3 | Hurt |
| Start from cold-start random roots instead of human-written (ablation) | KnowEvo · SaProt solubility | pass@1: 51.5 -> 46.4 | No clear effect |
| Use warm-start roots (ablation) | KnowEvo · SaProt solubility | pass@1: 51.5 -> 54.6 | No clear effect |
| Re-score Rosetta-evolved strategies with ThermoMPNN | KnowEvo · Stability cross-predictor | pass@1: 9.1 -> 15.7 | Improved |
| Increase accepted revisions from 10 to 100 (ablation) | KnowEvo · SaProt solubility | pass@3: 59.5 -> 74.2 | Improved |
| Allow more agent-work budget per revision (ablation) | KnowEvo · SaProt solubility | endpoint gain | Improved |
| Use DeepSeek instead of MiniMax as revision backbone (ablation) | KnowEvo · SaProt solubility | endpoint performance | Improved |
Limitations
- Only the abstract was available; experimental details are unverified.
Closest papers by meaning
- Knowledge-aware Reinforced Language Models for Protein Directed Evolution (ICML 2024 · similarity 0.92)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.89)
- Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents (ICML 2026 · similarity 0.88)
- Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks (NeurIPS 2026 · similarity 0.87)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.87)
- ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
