Research map / Probabilistic and scientific modeling / Molecules and proteins
ProSpero: Active Learning for Robust Protein Design Beyond Wild-Type Neighborhoods
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 ProSpero, an active learning loop that guides a frozen EvoDiff generative model with a surrogate via targeted alanine-scan-inspired masking and charge-constrained Sequential Monte Carlo, to find high-fitness, novel protein sequences beyond the wild-type neighborhood.
Problem
Data-efficient protein sequence design that must find high-fitness sequences far from the wild type while avoiding surrogate misspecification and biologically implausible sequences.
Approach
Treats iterative design as inference-time guidance of a frozen pre-trained generative model (EvoDiff-OADM) by a surrogate updated from oracle feedback, combining alanine-scan-style targeted masking of fitness-relevant residues with SMC whose proposals are restricted to same-charge amino acids and reweighted with a perplexity correction.
Key results
- ProSpero achieves the highest maximum fitness on 5 of 8 tasks and second on the other 3, improving over wild type on all tasks (5 runs).
- On the E4B task, ProSpero's mean fitness of top 100 is 8.017 vs 7.948 for PEX, and novelty is highest on 6 of 8 tasks, reportedly 2 to 9 times PEX's novelty.
- Under severe covariate shift (75 mutations from wild type), ProSpero reaches max pTM 0.672 vs 0.652 for the next best (MLDE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProSpero instead of the strongest baseline | protein fitness optimization · E4B, AAV, GFP, UBE2I, and LGK | maximum fitness: 8.099 $\pm$ 0.017 -> 8.114 $\pm$ 0.037 (E4B); 0.686 $\pm$ 0.021 -> 0.720 $\pm$ 0.027 (AAV); 3.603 $\pm$ 0.003 -> 3.617 $ | Helped |
| Use ProSpero instead of the strongest baseline | protein fitness optimization · AMIE, TEM, and Pab1 | maximum fitness: 0.248 $\pm$ 0.007 -> 0.246 $\pm$ 0.006 (AMIE); 1.232 $\pm$ 0.000 -> 1.231 $\pm$ 0.002 (TEM); 1.978 $\pm$ 0.188 -> 1.527 | Mixed |
| Use ProSpero instead of leading baselines | protein design · moderate, severe, and low-data covariate shifts on UBE2I | maximum pTM: 0.810 $\pm$ 0.020 -> 0.822 $\pm$ 0.027; 0.652 $\pm$ 0.059 -> 0.672 $\pm$ 0.031; 0.806 $\pm$ 0.013 -> 0.808 $\pm$ 0.017 | Helped |
| Add SMC resampling (ablation) | ProSpero · AAV noisy-surrogate setting | fitness | Helped |
| Use targeted rather than random masked positions (ablation) | ProSpero · AAV noisy-surrogate setting | fitness | Helped |
| Add SMC resampling and targeted masking (ablation) | ProSpero · AAV noisy-surrogate setting | fitness | Helped |
| Add SMC resampling, targeted masking, and charge-compatible proposal restriction (RAA) (ablation) | ProSpero · AAV noisy-surrogate setting | fitness | Helped |
| Use surrogate guidance at very low SNR | ProSpero · AAV noisy-surrogate setting | fitness | Hurt |
| Use ProSpero rather than leading approaches | protein design · 8 protein engineering tasks | mean fitness of top 100 | Helped |
| Use ProSpero rather than leading approaches | protein design · 8 protein engineering tasks | novelty | Helped |
| Use ProSpero rather than leading approaches | protein design · E4B | validity | No clear effect |
Surprising findings
- Under very low SNR surrogate noise, guidance appears to slightly hinder performance, with ablated variants sometimes performing comparably.
Limitations
- Approximate SMC may reduce diversity; the authors note this as a possible risk.
- Relies on the EvoDiff ByteNet CNN backbone; deterministic runs are much slower.
- Charge-based reduced alphabet is a fixed heuristic and may not suit every protein; adaptive alphabets are left to future work.
Experimental setup
Datasets: AAV (FLEXS) GFP (FLEXS) AMIE TEM-1 E4B Pab1 UBE2I LGK
Benchmarks: AMIE TEM-1 E4B Pab1 AAV GFP UBE2I LGK
Metrics: maximum fitness mean fitness of top 100 novelty (Hamming distance to start) diversity validity pTM pLDDT scPerplexity
Baselines: CMA-ES DyNaPPO BO PEX AdaLead CbAS GFN-AL GFN-AL-δCS LatProtRL MLDE
Builds on: EvoDiff PEX AdaLead GFN-AL-δCS Sequential Monte Carlo
Closest papers by meaning
- Evolution-Inspired Loss Functions for Protein Representation Learning (ICML 2024 · similarity 0.88)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.88)
- A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.88)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.87)
- Tree Search-Based Evolutionary Bandits for Protein Sequence Optimization (AAAI 2024 · similarity 0.87)
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.87)
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.