Research map / Large language models / Retrieval and knowledge
Knowledge-aware Reinforced Language Models for Protein Directed Evolution
Read the paperCheck your idea against 108,967 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 KnowRLM, an RL-based machine-learning-assisted directed evolution method that uses an amino acid knowledge graph (AAKG) to guide a PLM-based policy via preferential random walks with a dynamic sliding window, and fine-tunes a fitness predictor as reward model in active learning.
Problem
Machine-learning-assisted directed evolution (MLDE) relies on data-driven sampling and ignores biochemical domain knowledge about amino acids, making exploration of the huge mutation space inefficient.
Approach
Encode amino acid physicochemical relationships as a knowledge graph (AAKG built from AAontology) and use preferential random walks on it, guided by PLM embeddings and a dynamic sliding window, as the mutation-type policy of an RL directed evolution loop with an actively fine-tuned fitness predictor as reward.
Key results
- On GB1 with 384 samples, KnowRLM reaches max 0.972, mean 0.562, NDCG 0.884, versus EvoPlay's 0.840 / 0.460 / 0.847.
- On PhoQ with 384 samples, KnowRLM reaches max 0.658 and mean 0.157, versus the best baseline max 0.474 (EvoPlay) and mean 0.148 (CLADE2.0).
- Ablation removing the AAKG lowers GB1 (384 samples) max from 0.972 to 0.930 and NDCG from 0.884 to 0.816.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AAKG to the policy (ablation) | KnowRLM · GB1, sample size 384 | max: 0.930 -> 0.972 | Helped |
| add AAKG to the policy (ablation) | KnowRLM · GB1, sample size 384 | mean: 0.504 -> 0.562 | Helped |
| add AAKG to the policy (ablation) | KnowRLM · GB1, sample size 384 | NDCG: 0.816 -> 0.884 | Helped |
| add AAKG to the policy (ablation) | KnowRLM · PhoQ, sample size 384 | max: 0.446 -> 0.658 | Helped |
| add AAKG to the policy (ablation) | KnowRLM · PhoQ, sample size 384 | mean: 0.143 -> 0.157 | Helped |
| add AAKG to the policy (ablation) | KnowRLM · PhoQ, sample size 384 | NDCG: 0.819 -> 0.819 | No clear effect |
| use a dynamic sliding window rather than a non-dynamic window (ablation) | KnowRLM · GB1, sample size 384 | max: 0.862 -> 0.972 | Helped |
| use a dynamic sliding window rather than a non-dynamic window (ablation) | KnowRLM · GB1, sample size 384 | mean: 0.513 -> 0.551 | Helped |
| use a dynamic sliding window rather than a non-dynamic window (ablation) | KnowRLM · GB1, sample size 384 | NDCG: 0.826 -> 0.878 | Helped |
| increase the sliding-window bounds w1 and w2 (ablation) | KnowRLM · GB1, sample size 384 | mean: 0.551 -> 0.543 | Hurt |
| decrease the sliding-window bounds w3 and w4 (ablation) | KnowRLM · GB1, sample size 384 | NDCG: 0.878 -> 0.869 | Hurt |
| increase w1 and decrease w2, w3, and w4 (ablation) | KnowRLM · GB1, sample size 384 | NDCG: 0.878 -> 0.892 | Helped |
Surprising findings
- KnowRLM with only one or two RL rounds (192 or 288 samples) already surpasses most baselines on GB1, suggesting strong sample efficiency.
Limitations
- Knowledge graph built only from amino acid-level properties, with limited structured biochemical knowledge.
- Sensitivity to dynamic window hyperparameters; extreme settings degrade performance.
- Evaluated only on two four-site mutation datasets (GB1, PhoQ).
Experimental setup
Datasets: GB1 PhoQ AAontology
Benchmarks: GB1 PhoQ
Metrics: max fitness mean fitness NDCG
Baselines: MLDE ftMLDE (EVmutation) ftMLDE (Transformer) CLADE CLADE2.0 EvoPlay
Builds on: EvoPlay MLDE ftMLDE CLADE2.0 AAontology
Closest papers by meaning
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.89)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.88)
- MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering (NeurIPS 2024 · similarity 0.88)
- Steering Protein Language Models (ICML 2025 · similarity 0.87)
- Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.87)
- Retrieval-Augmented Language Model for Knowledge-aware Protein Encoding (ICML 2025 · 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.