Research map / Probabilistic and scientific modeling / Molecules and proteins
GFlowNet Assisted Biological Sequence Editing
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 GFNSeqEditor, a GFlowNet-based biological sequence editor that identifies sub-optimal positions with a trained flow function and resamples them, yielding property gains with fewer edits on DNA and protein tasks.
Problem
Editing biological sequences to improve a property while minimizing the number of edits (for predictability and safety), rather than generating sequences de novo, which can drift far from natural sequences.
Approach
Uses a trained GFlowNet flow function to flag positions whose out-flow is small relative to the best action (with Gaussian noise for exploration), then resamples flagged positions from a flow-proportional policy mixed with the original base via a regularization coefficient lambda; provides theoretical bounds on reward, property improvement and edit count.
Key results
- On TFbinding, GFNSeqEditor reaches PI 0.14 at 24.27% edit percentage vs 0.12 for Directed Evolution at 25%.
- On CRE, GFNSeqEditor reaches PI 9.90 with 21.90% edits vs 4.24 for GFlowNet-E and 2.45 for MOGFN-AL at similar edit rates.
- On AMP, GFNSeqEditor achieves PI 0.33 with diversity 14.34 and GMDPI 2.17, versus 0.28 PI for GFlowNet-E.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GFNSeqEditor instead of DE | TFbinding sequence editing · TFbinding | PI: 0.12 -> 0.14 | Helped |
| Use GFNSeqEditor instead of GFlowNet-E | AMP sequence editing · AMP | PI: 0.28 -> 0.33 | Helped |
| Use GFNSeqEditor instead of GFlowNet-E | CRE sequence editing · CRE | PI: 4.24 -> 9.90 | Helped |
| Increase δ (ablation) | GFNSeqEditor · AMP and CRE | PI | Helped |
| Increase δ (ablation) | GFNSeqEditor · AMP and CRE | EP(%) | Mixed |
| Increase λ (ablation) | GFNSeqEditor · AMP and CRE | PI | Hurt |
| Increase λ (ablation) | GFNSeqEditor · AMP and CRE | EP(%) | Hurt |
| Increase σ (ablation) | GFNSeqEditor · AMP and CRE | PI | Hurt |
| Increase σ (ablation) | GFNSeqEditor · AMP and CRE | Diversity | Helped |
Surprising findings
- Seq2Seq produces a single deterministic edit per input, so its diversity is zero and it could not be applied to CRE due to too few training pairs.
Limitations
- Relies on a well-trained GFlowNet; performance depends on flow model quality.
- Theoretical bounds rely on assumptions (e.g., action probability bounded by 1-epsilon) and the bounds are loose relative to empirical results.
Experimental setup
Datasets: TFbinding AMP (DBAASP) CRE (MPRA, Malinois oracle)
Benchmarks: TFbinding AMP CRE
Metrics: property improvement edit percentage diversity GMDPI CDF of generated property
Baselines: Directed Evolution Ledidi LaMBO MOGFN-AL GFlowNet-E Seq2Seq
Builds on: GFlowNet trajectory balance GFlowNet-AL AdaLead
Closest papers by meaning
- gfnx: Fast and Scalable Library for Generative Flow Networks in JAX (NeurIPS 2026 · similarity 0.86)
- pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows (NeurIPS 2026 · similarity 0.83)
- Designing Biological Sequences without Prior Knowledge Using Evolutionary Reinforcement Learning (AAAI 2024 · similarity 0.83)
- Steering Protein Language Models (ICML 2025 · similarity 0.82)
- DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing (NeurIPS 2025 · similarity 0.82)
- PhyloGFN: Phylogenetic inference with generative flow networks (ICLR 2024 · similarity 0.81)
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.