Research map / Probabilistic and scientific modeling / Molecules and proteins
Steering Generative Models with Experimental Data for 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.
Steers protein generative priors (continuous, discrete uniform/absorbing diffusion, and an autoregressive PLM) with a few hundred labeled fitness values via classifier guidance, DAPS posterior sampling, NOS, or DPO, and integrates guidance into a Thompson-sampling-like adaptive loop on TrpB, CreiLOV and GB1.
Problem
Protein fitness optimization in large combinatorial design spaces when only about 10^2–10^3 wet-lab fitness labels are available, where prior steered-generation work mostly used surrogate rewards or large labeled datasets.
Approach
Systematically compares plug-and-play guidance (classifier guidance, DAPS posterior sampling, NOS) on several MSA-trained generative priors with few labels, and adapts decoupled annealing posterior sampling to discrete diffusion models, combined with an ensemble of value functions sampled per round as Thompson sampling.
Key results
- Steering with 200 labeled sequences lets most models and methods generate sequences with higher fitness, with diffusion guidance outperforming DPO on ARLM in the low-data regime.
- DAPS and CG give the strongest steerability overall; the paper concludes DAPS with discrete diffusion has the highest overall performance.
- In adaptive Thompson-sampling-like campaigns, guided MDLM outperforms ARLM+DPO, unconditional prior sampling, and APEXGo in mean and maximum fitness on TrpB, CreiLOV and GB1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use plug-and-play diffusion guidance instead of ARLM+DPO | SGPO protein fitness optimization · TrpB, CreiLOV, and GB1 | fitness steerability | Helped |
| Use CG or DAPS guidance | Diffusion-based SGPO · TrpB and CreiLOV | fitness steerability | Helped |
| Use DAPS instead of CG | Continuous diffusion models · TrpB and CreiLOV | fitness steerability | Helped |
| Use uniform-noise diffusion instead of absorbing-state diffusion | D3PM and MDLM · TrpB and CreiLOV | fitness steerability | No clear effect |
| Use continuous diffusion instead of discrete diffusion | SGPO · TrpB and CreiLOV | fitness steerability | Hurt |
| Use NOS guidance instead of CG or DAPS | Diffusion-based SGPO · TrpB and CreiLOV | fitness steerability | Hurt |
| Use guided MDLM sampling instead of DPO, unconditional sampling, or APEXGo | Adaptive SGPO · TrpB, CreiLOV, and GB1 | mean and maximum fitness | Helped |
| Guide each new sample with a value function drawn from an ensemble instead of using a single value function | Adaptive MDLM guidance · TrpB, CreiLOV, and GB1 | maximum fitness | Helped |
| Use DAPS with discrete diffusion | SGPO · TrpB, CreiLOV, and GB1 | fitness | Helped |
| Use a Gaussian-process value function instead of the neural-network ensemble | Adaptive SGPO · Protein fitness optimization | optimization performance | Hurt |
| Use ESM-embedding latent diffusion instead of the other continuous diffusion representation | Continuous protein diffusion · Protein sequence generation | generation performance | No clear effect |
| Use UDLM instead of D3PM | Discrete diffusion protein prior · Protein sequence generation | generation diversity | Hurt |
Surprising findings
- The UDLM model suffers mode collapse, and continuous diffusion over ESM embeddings does not boost performance.
- Guidance with diffusion models requires no prior weight updates and each guidance experiment took minutes, versus hours for pretraining.
Limitations
- Evaluated only on in-silico oracles trained on existing data, not wet-lab validation.
- Only substitutions in fixed-length sequences; insertions and deletions not considered; TrpB and LM outputs mapped back to design space manually.
- Gaussian process value function performed poorly on high-dimensional protein representations; relative rather than absolute comparisons recommended.
Experimental setup
Datasets: TrpB enzyme activity CreiLOV fluorescence GB1 binding UniRef90 MSAs
Benchmarks: TrpB oracle CreiLOV oracle GB1 oracle
Metrics: mean fitness maximum fitness diversity (Shannon entropy at mutated positions) perplexity novelty/uniqueness
Baselines: ARLM+DPO APEXGo Unconditional prior sampling NOS
Builds on: DAPS D3PM MDLM UDLM NOS DPO
Closest papers by meaning
- A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.91)
- Steering Protein Language Models (ICML 2025 · similarity 0.90)
- Improving protein optimization with smoothed fitness landscapes (ICLR 2024 · similarity 0.90)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.89)
- Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design (ICLR 2025 · similarity 0.89)
- Diffusion Language Models Are Versatile Protein Learners (ICML 2024 · similarity 0.89)
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.