KnowEvo: Knowledge Evolution for Protein Optimization
Organizations: Dalian University of Technology
Abstract
Large language models (LLMs) have shown promise in scientific discovery, yet systematically accumulating and refining reusable scientific knowledge remains challenging. Existing approaches primarily improve LLM-based problem solving through parameter optimization or feedback-driven refinement of prompts, programs, and solution trajectories, leaving acquired expertise either embedded in model weights or scattered across iterative problem-solving processes. As a result,knowledge is difficult to systematically inspect, evaluate, and refine. We propose KnowEvo, a framework that treats externalized knowledge itself as the optimization target. Our key insight is that scientific knowledge is inherently modular: multiple interacting mechanisms jointly determine the design outcomes. This makes knowledge evolution fundamentally challenging, as improvements cannot be attributed to individual components in isolation. To address this, KnowEvo represents knowledge as executable expertise blocks and organizes their refinements into strategy mutation trees, enabling structured exploration over a combinatorial knowledge space. The framework combines a Thompson-style forward allocation strategy that adaptively distributes evaluation budgets across strategy variants with a backward revision process, in which a code agent operates in an isolated sandbox to translate rollout evidence into localized program updates. This design enables persistent accumulation, evaluation, and refinement of mechanistic expertise across iterative experimentation. We evaluate KnowEvo on protein optimization tasks including stability and solubility. Results show that knowledge evolution consistently improves over initial strategy blocks and outperforms monolithic program evolution and prior LLM-guided evolution baselines across metrics such as pass@k, MRR, MAP,and NDCG@5. Code is available at https://github.com/recynie/KnowEvo.
Figures & tables
| Method | Hit rate | Ranking quality | ||||
|---|---|---|---|---|---|---|
| pass@1 | pass@3 | pass@5 | MRR | MAP | NDCG@5 | |
| Stability / Rosetta | ||||||
| Direct LLM proposal | 37.9 | 70.2 | 82.3 | 54.7 | 54.7 | 61.7 |
| Pro-1 8B | 42.4 | 69.2 | 77.3 | 56.4 | 53.2 | 61.0 |
| Monolithic program | 43.4 | 72.2 | 83.8 | 58.7 | 55.6 | 64.2 |
| GEPA | 37.9 | 70.7 | 83.8 | 55.0 | 54.1 | 62.7 |
| Method | Hit rate | Ranking quality | ||||
|---|---|---|---|---|---|---|
| pass@1 | pass@3 | pass@5 | MRR | MAP | NDCG@5 | |
| Random node allocation | 51.0 | 78.5 | 88.2 | 65.4 | 61.8 | 69.9 |
| Single-step LLM revision | 49.0 | 80.1 | 84.2 | 64.2 | 62.6 | 69.3 |
| Chain-only revision | 43.4 | 54.1 | 54.1 | 48.7 | 48.7 | 50.1 |
| KnowEvo (Full method) | 51.5 | 83.2 | 90.8 | 67.8 | 64.2 | 72.8 |
| Initialization | Hit rate | Ranking quality | ||||
|---|---|---|---|---|---|---|
| pass@1 | pass@3 | pass@5 | MRR | MAP | NDCG@5 | |
| Human-written | 51.5 | 83.2 | 90.8 | 67.8 | 64.2 | 72.8 |
| Cold start | 46.4 | 79.1 | 93.9 | 64.2 | 60.1 | 70.6 |
| Warm start | 54.6 | 81.1 | 91.8 | 69.3 | 62.6 | 72.4 |
| LLM-generated | 51.5 | 85.7 | 92.4 | 68.7 | 66.8 | 74.5 |
| Strategy | pass@1 | pass@3 | pass@5 | MRR | MAP | NDCG@5 |
|---|---|---|---|---|---|---|
| Initial | 9.1 | 31.3 | 36.9 | 20.5 | 20.5 | 24.7 |
| KnowEvo | 15.7 | 32.8 | 37.4 | 23.7 | 23.1 | 26.9 |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Task | Metric | Ours–Root (pp) | 95% bootstrap CI (pp) |
|---|---|---|---|
| Stability/Rosetta ( ) | pass@1 | ||
| pass@3 | |||
| pass@5 | |||
| MRR | |||
| MAP | |||
| NDCG@5 |
| Metric | Rosetta | ThermoMPNN | ||||
|---|---|---|---|---|---|---|
| Initial | KnowEvo | (pp) | Initial | KnowEvo | (pp) | |
| pass@1 | 42.4 | 50.5 | +8.1 | 9.1 | 15.7 | +6.6 |
| pass@3 | 74.2 | 82.8 | +8.6 | 31.3 | 32.8 | +1.5 |
| pass@5 | 85.4 | 88.4 | +3.0 | 36.9 | 37.4 | +0.5 |
| MRR | 59.1 | 66.4 | +7.3 | 20.5 | 23.7 | +3.2 |
| MAP | 56.3 | 63.2 | +6.9 | 20.5 | 23.1 | +2.6 |
| Step | Edge | Executable refinement |
|---|---|---|
| E1 | Reverse the hydrophobic preference in charge-free contexts; enforce diversity across edit families. | |
| E2 | Protect charge-free contexts and early N-terminal positions; increase spacing between edits. | |
| E3 | Shift candidate-family preference toward charge edits and away from generic aromatic relief. | |
| E4 | Gate K/R edits by local charge sign and the presence of nearby basic residues. | |
| E5 | Restrict K/R edits to globally and locally negative contexts outside early N-terminal positions. | |
| E6 | Require local polar support and protect early sequence positions. |
| Rule component | Positive examples | Negative examples |
|---|---|---|
| Deleted terminal residue | I, Y, M, A, C | G, D, E, K, H |
| Newly exposed C-terminus | Q, T, E, N, D, G | W, K, hydrophobic residues |
| Rule type | Revised program behavior | Intended effect |
|---|---|---|
| High-flex loop gate | Expands the risky-loop condition from to , and increases the penalty from to . | Suppress borderline interior or near-terminal loop edits. |
| Deletion vetoes | Adds penalties for semi-buried deletions and near-terminal high-flex deletions . | Avoid deletions that create packing voids or local clashes. |
| DEL–INS conflict | Widens the risky DEL–INS proximity check from 10 to 25 residues. | Avoid paired edits that induce backbone strain. |
| Method | Total tokens (M) | Tokens/revision (M) | pass@1 | pass@3 | pass@5 | MRR | MAP | NDCG@5 |
|---|---|---|---|---|---|---|---|---|
| KnowEvo | 44.088 | 0.4550 | 51.5 | 83.2 | 90.8 | 67.8 | 64.2 | 72.8 |
| Single-call KnowEvo | 3.728 | 0.0373 | 49.0 | 80.1 | 84.2 | 64.2 | 62.6 | 69.3 |
| AdaEvolve | 3.300 | 0.1640 | 49.0 | 74.0 | 82.5 | 61.9 | 59.8 | 66.9 |
| Agentic AdaEvolve | 117.931 | 2.3590 | 42.3 | 79.6 | 80.6 | 58.0 | 58.1 | 64.2 |
| GEPA | 0.764 | 0.0347 | 40.3 | 76.0 | 87.2 | 59.4 | 58.2 | 66.7 |
| Block | Mechanism | Editing rule |
|---|---|---|
| Glycine entropy relief | Surface or loop glycines can increase backbone flexibility and raise the entropy cost of folding. Replacing selected glycines with alanine can reduce excess flexibility with small steric cost. | Rank glycine residues by low contact count, high local backbone motion, local polar support, and absence of dense glycine/proline clustering. Propose substitutions at the best ranked positions. |
| Proline rigidity relief | An isolated proline in a flexible, weakly packed region can introduce a local kink or restrict favorable backbone conformations. Removing this rigidity can sometimes improve local packing. | Rank prolines that are low-contact, locally isolated, and not part of a dense proline/glycine cluster. Replace selected prolines with alanine in more polar or flexible windows, or leucine in more hydrophobic windows. |
| Polar/acidic swap | Small polar and acidic residues can often be adjusted within the same chemical family without large structural disruption. Such edits can improve local polar packing or reduce unfavorable side-chain placement. | Rank low-contact polar or acidic residues in local polar clusters. Propose conservative swaps such as , , , , and . |
| Core packing rescue | A high-contact site occupied by a small or polar residue can indicate an under-packed core or weak hydrophobic neighborhood. A conservative bulkier replacement can improve packing. | Rank targetable residues by high contact count, hydrophobic neighbors, aromatic neighbors, low backbone motion, and distance from termini. Propose hydrophobic or packing-preserving substitutions, while spacing edits across distinct sites. |
| Loop micro-indel stabilizer | Flexible loops and termini can sometimes tolerate small length changes or conservative reshaping, but interior loop indels can also create packing voids or backbone strain. | Rank flexible or terminal residues by contact count, local flexible-residue density, terminal distance, and backbone motion. Depending on allowed mutation types, propose a small deletion, insertion, or conservative substitution, while spacing edits across the sequence. |
| Block | Mechanism | Editing rule |
|---|---|---|
| Hydrophobic patch relief | Local hydrophobic clusters can reduce solubility by favoring exposed self-association or aggregation-prone patches. | Rank hydrophobic residues in windows enriched for hydrophobic or aromatic residues and poor in charged residues. Replace selected residues with more polar or less aggregation-prone amino acids. |
| Charge redistribution | Weakly charged surface-like windows can often be made more soluble by adding charged or polar residues, especially when the local window is hydrophobic or charge-imbalanced. | Rank residues in locally hydrophobic or charge-poor windows. Propose substitutions that add charge or polarity, such as , , , or . |
| Aromatic patch breakup | Aromatic clusters can promote self-association through local aromatic stacking and hydrophobic patch formation. | Rank aromatic residues in windows with multiple aromatic residues. Replace selected aromatics with less aggregation-prone alternatives, such as , , , or . |
| Multi-site solubilizer | Some proteins need more than one low-risk solubility edit. Combining edits from different residue mechanisms can improve the shortlist without relying on a single residue class. | Pool candidates from hydrophobic-patch relief, charge redistribution, and aromatic-patch breakup. Select a spaced and mechanism-diverse set of edits. |
| Terminal low-complexity editor | Low-complexity or hydrophobic terminal tails can be poorly soluble and can sometimes be improved by a small terminal edit. | Inspect the N- and C-terminal windows. If a terminal window is low-complexity or hydrophobic, propose a one-residue deletion, a short polar insertion, or a fallback substitution, depending on the allowed mutation types. |
Large Language Model-Guided Optimization
Large Language Model-Guided Optimization uses language models to propose, steer, and refine solutions to hard search and design problems.
See all 654 papers