Research map / Probabilistic and scientific modeling / Molecules and proteins
A Variational Perspective on Generative Protein Fitness Optimization
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 VLGPO, a variational latent generative framework that combines a VAE latent space, a flow-matching prior, and a CNN fitness predictor with manifold-constrained classifier guidance to optimize protein fitness on AAV and GFP benchmarks.
Problem
Discovering protein sequences with higher fitness in a vast, sparse, discrete fitness landscape when only a small training set of mutants is available.
Approach
Embed protein sequences in a continuous VAE latent space, learn a flow-matching prior there, and guide sampling with the gradient of a fitness predictor evaluated on the denoised estimate x̂1 (manifold-constrained guidance), giving an explicit variational prior-times-likelihood posterior.
Key results
- On GFP medium, VLGPO with predictor g_phi reaches 0.87 median fitness vs. 0.76 for GGS and 0.10 for GWG.
- On GFP hard, VLGPO with smoothed predictor reaches 0.78 median fitness, best among methods (GGS 0.74, gg-dWJS 0.61).
- On AAV hard, VLGPO with smoothed predictor reaches 0.61 median fitness, best over GGS (0.60) and gg-dWJS (0.33); on AAV medium VLGPO with g_phi reaches 0.58 vs. GGS 0.51.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VLGPO with predictor gφ instead of GGS | protein fitness optimization · GFP medium | fitness: 0.76 $\pm$ 0.0 -> 0.87 $\pm$ 0.0 | Helped |
| Use VLGPO with smoothed predictor instead of GGS | protein fitness optimization · GFP hard | fitness: 0.74 $\pm$ 0.0 -> 0.78 $\pm$ 0.0 | Helped |
| Use VLGPO with predictor gφ instead of GGS | protein fitness optimization · AAV medium | fitness: 0.51 $\pm$ 0.0 -> 0.58 $\pm$ 0.0 | Helped |
| Use VLGPO with smoothed predictor instead of GGS | protein fitness optimization · AAV hard | fitness: 0.60 $\pm$ 0.0 -> 0.61 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with predictor gφ · GFP medium | fitness: 0.81 $\pm$ 0.0 -> 0.87 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with predictor gφ · GFP hard | fitness: 0.73 $\pm$ 0.0 -> 0.75 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with smoothed predictor · GFP medium | fitness: 0.84 $\pm$ 0.0 -> 0.84 $\pm$ 0.0 | No clear effect |
| Add manifold-constrained gradients (ablation) | VLGPO with smoothed predictor · GFP hard | fitness: 0.67 $\pm$ 0.1 -> 0.78 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with predictor gφ · AAV medium | fitness: 0.55 $\pm$ 0.0 -> 0.58 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with predictor gφ · AAV hard | fitness: 0.47 $\pm$ 0.0 -> 0.51 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with smoothed predictor · AAV medium | fitness: 0.52 $\pm$ 0.0 -> 0.53 $\pm$ 0.0 | Helped |
| Add manifold-constrained gradients (ablation) | VLGPO with smoothed predictor · AAV hard | fitness: 0.58 $\pm$ 0.0 -> 0.61 $\pm$ 0.0 | Helped |
Surprising findings
- On AAV medium, the smoothed-predictor variant of VLGPO is outperformed by the non-smoothed predictor variant, and on GFP medium, vanilla GWG (0.10) is far below its own flow-prior variant (0.87), which suggests the latent prior matters more than the predictor smoothing.
- Learned fitness-conditioned posterior (direct flow matching) fails to extrapolate on GFP hard (0.44 vs 0.75 for VLGPO).
Limitations
- Limited to only AAV and GFP benchmarks; not tested on other proteins (FLIP, ProteinGym suggested).
- Requires sensitive hyperparameter tuning (alpha_t, J), especially for GFP hard.
- Evaluation relies on an in-silico oracle rather than experimental validation.
Experimental setup
Datasets: AAV (FLIP) GFP (FLIP)
Benchmarks: FLIP AAV medium FLIP AAV hard FLIP GFP medium FLIP GFP hard
Metrics: median normalized fitness diversity (Levenshtein) novelty (Levenshtein)
Baselines: GFN-AL CbAS AdaLead BOqei CoMS PEX gg-dWJS GWG GGS
Builds on: flow matching beta-VAE gradient-guided MCMC (GWG) graph-smoothed GWG (GGS) FLIP benchmark
Closest papers by meaning
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.91)
- Improving protein optimization with smoothed fitness landscapes (ICLR 2024 · similarity 0.89)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.88)
- Variable-Length Generative Protein Design via Generalized Poisson Flow (NeurIPS 2026 · similarity 0.88)
- Tree Search-Based Evolutionary Bandits for Protein Sequence Optimization (AAAI 2024 · similarity 0.88)
- Variational Search Distributions (ICLR 2025 · 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.