Research map / Reinforcement learning / Visual and representation RL
Designing Biological Sequences without Prior Knowledge Using Evolutionary Reinforcement Learning
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 ERLBioSeq, which combines evolutionary algorithms and reinforcement learning with a CNN+BiLSTM fitness predictor (LGFEN) to design DNA, RNA, and protein sequences without prior knowledge.
Problem
Designing biological sequences with desired properties in huge discrete search spaces when little labeled data is available and wet-lab evaluation is costly.
Approach
Alternates EA mutation/crossover with an RL policy that generates sequences, with EA-generated candidates screened by a learned predictor combining local (CNN) and global (BiLSTM) sequence features, so each method enhances the other's exploration.
Key results
- ERLBioSeq achieves top-10 mean of 0.976 on RNA length 14 and 0.805 on RNA length 50, outperforming the baselines in Figure 2 and Table 1.
- On the protein 3MSI task, ERLBioSeq reaches top-10 mean 14.621 versus 12.294 for the best EA-combined variant (EVOBo) in Table 3.
- Replacing the RL module with Bo, CMA-ES, CBAS, or GFlowNet within the EA lowers performance, and LGFEN outperforms MLP, CNN, LSTM and BI-LSTM predictors across tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ERLBioSeq instead of AdaLead | TF Bind 8 design · TF Bind 8, Top10 mean | mean binding activity: 0.984±0.007 -> 0.989±0.006 | Helped |
| Use ERLBioSeq instead of AdaLead | TF Bind 8 design · TF Bind 8, Top50 mean | mean binding activity: 0.963±0.016 -> 0.970±0.013 | Helped |
| Use ERLBioSeq instead of AdaLead | TF Bind 8 design · TF Bind 8, Top100 mean | mean binding activity: 0.941±0.026 -> 0.953±0.022 | Helped |
| Use local and global sequence features in LGFEN rather than an MLP predictor (ablation) | ERLBioSeq · DNA, Top10 mean | mean binding activity: 0.981 ± 0.008 -> 0.989 ± 0.006 | Helped |
| Use local and global sequence features in LGFEN rather than a CNN predictor (ablation) | ERLBioSeq · RNA14, Top10 mean | mean binding activity: 0.942 ± 0.054 -> 0.976 ± 0.043 | Helped |
| Use local and global sequence features in LGFEN rather than an MLP predictor (ablation) | ERLBioSeq · RNA50, Top10 mean | mean binding activity: 0.728 ± 0.028 -> 0.805 ± 0.043 | Helped |
| Use local and global sequence features in LGFEN rather than an LSTM predictor (ablation) | ERLBioSeq · Protein 3MSI, Top10 mean | score: 12.477±1.829 -> 14.621±2.218 | Helped |
| Replace the RL module with Bayesian Optimization (ablation) | ERLBioSeq evolutionary process · Protein 3MSI, Top10 mean | score: 12.294±1.007 -> 14.621±2.218 | Hurt |
| Replace the RL module with CMA-ES (ablation) | ERLBioSeq evolutionary process · RNA14, Top10 mean | mean binding activity: 0.930 ± 0.058 -> 0.976 ± 0.043 | Hurt |
| Replace the RL module with Cbas (ablation) | ERLBioSeq evolutionary process · RNA50, Top10 mean | mean binding activity: 0.804 ± 0.038 -> 0.805 ± 0.043 | No clear effect |
| Replace the RL module with GflowNet (ablation) | ERLBioSeq evolutionary process · Protein 3MSI, Top10 mean | score: 12.976±1.077 -> 14.621±2.218 | Hurt |
Surprising findings
- Ablation suggests RL contributes most to the EA component, per the conclusion.
Limitations
- Evaluated only on energy-model or simulated oracles (ViennaRNA, PyRosetta, TF Bind 8) rather than wet-lab experiments.
- DNA task is relatively easy so differences between methods are small; top-k averages used instead of full curves.
Experimental setup
Datasets: ViennaRNA-based RNA binding task PyRosetta 3MSI protein folding task TF Bind 8 (Barrera et al. 2016)
Benchmarks: RNA binding (length 14 and 50) PyRosetta protein design 3MSI TF Bind 8 DNA design
Metrics: cumulative maximum fitness mean fitness of top-10/50/100 sequences
Baselines: GFlowNet AdaLead Dynappo CBAS CMA-ES Bayesian Optimization
Builds on: AdaLead Dynappo CMA-ES CBAS
Closest papers by meaning
- Improved Off-policy Reinforcement Learning in Biological Sequence Design (ICML 2025 · similarity 0.89)
- Tree Search-Based Evolutionary Bandits for Protein Sequence Optimization (AAAI 2024 · similarity 0.86)
- Regulatory DNA Sequence Design with Reinforcement Learning (ICLR 2025 · similarity 0.85)
- Knowledge-aware Reinforced Language Models for Protein Directed Evolution (ICML 2024 · similarity 0.85)
- Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks (ICML 2025 · similarity 0.84)
- BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments (ICLR 2025 · similarity 0.84)
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.