Research map / Graph learning / Graph algorithms and clustering
Improving protein optimization with smoothed fitness landscapes
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 GGS, which smooths a protein fitness landscape with Tikhonov graph-signal regularization over a mutation-based kNN graph and then samples improved sequences with Gibbs-with-gradients, evaluated on GFP and AAV optimization benchmarks.
Problem
Protein sequence optimization with a noisy, non-smooth fitness landscape and limited data, where prior methods are restricted to small mutational radii and may get stuck in local optima or produce false positives.
Approach
Formulate sequences as a kNN graph (Levenshtein distance) with fitness as a graph signal, smooth fitness via Tikhonov regularization (closed form (I+γL)^-1 Y), retrain a 1D CNN on smoothed labels, and run Gibbs with Gradients on that smooth energy; graph smoothing is also shown to improve several baselines.
Key results
- On GFP hard, GGS reaches 0.74 normalized fitness versus 0.18 for the best unsmoothed baseline (CbAS/AdaLead), and CbAS+GS rises to 0.57.
- On AAV hard, GGS achieves 0.60 fitness versus 0.40 for AdaLead and 0.36 for CbAS.
- Adding graph smoothing to CbAS raises GFP medium fitness from 0.14 to 0.66 and to 0.57 on GFP hard.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply graph-based smoothing to GGS (ablation) | GGS · GFP hard | Median Fitness: 0.10 -> 0.74 | Helped |
| apply graph-based smoothing to GGS (ablation) | GGS · GFP medium | Median Fitness: 0.51 -> 0.76 | Helped |
| apply graph-based smoothing to GGS (ablation) | GGS · AAV hard | Median Fitness: 0.33 -> 0.60 | Helped |
| apply graph-based smoothing to GGS (ablation) | GGS · AAV Harder2 | Median Fitness: 0.28 -> 0.27 | Hurt |
| apply graph-based smoothing to GGS (ablation) | GGS · GFP Harder3 | Median Fitness: 0.00 -> 0.00 | No clear effect |
| apply graph-based smoothing to CbAS (ablation) | CbAS · GFP hard | Fitness: 0.18 -> 0.57 | Helped |
| apply graph-based smoothing to AdaLead (ablation) | AdaLead · GFP hard | Fitness: 0.18 -> 0.39 | Helped |
| apply graph-based smoothing to CbAS (ablation) | CbAS · AAV hard | Fitness: 0.36 -> 0.4 | Helped |
| apply graph-based smoothing to BOqei (ablation) | BOqei · GFP medium | Fitness: 0.20 -> 0.08 | Hurt |
| apply graph-based smoothing to GFN-AL (ablation) | GFN-AL · AAV medium | Fitness: 0.2 -> 0.18 | No clear effect |
| set sampling temperature to 0.1 rather than 0.01 (ablation) | GGS · GFP hard | Fitness: 0.65 -> 0.74 | Helped |
| set sampling temperature to 1.0 rather than 0.1 (ablation) | GGS · GFP hard | Fitness: 0.74 -> 0.0 | Hurt |
Surprising findings
- Plain GWG without smoothing achieves near-zero fitness on GFP hard (0.0), showing gradient-based sampling depends on a smooth energy function.
- The transformer evaluator performed worse than a simple CNN, so the CNN was used as the evaluator.
Limitations
- Evaluation relies on a trained in-silico evaluator rather than wet-lab validation.
- Choice of smoothing weight gamma and graph size is task-dependent and selected by grid search.
- Smoothing gains are smaller on AAV, and the method fails on the hardest GFP task (10th percentile, gap 8).
Experimental setup
Datasets: GFP fluorescence dataset AAV dataset
Benchmarks: GFP medium GFP hard AAV medium AAV hard
Metrics: normalized fitness diversity novelty
Baselines: GFN-AL CbAS AdaLead BO-qei CoMs PEX GWG
Builds on: Gibbs With Gradients CbAS AdaLead PEX CoMs
Closest papers by meaning
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.90)
- A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.89)
- Robust Model-Based Optimization for Challenging Fitness Landscapes (ICLR 2024 · similarity 0.89)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.89)
- Augmenting Biological Fitness Prediction Benchmarks with Landscapes Features from GraphFLA (NeurIPS 2025 · similarity 0.88)
- Tree Search-Based Evolutionary Bandits for Protein Sequence Optimization (AAAI 2024 · similarity 0.88)
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.