Few-shot learning for protein evolution with EVOLVEpro
Peer-reviewed| Поле | Значение |
|---|---|
| Authors | NR |
| Year | 2025 |
| Venue / status | Peer-reviewed journal; Science (Q1) |
| DOI | 10.1126/science.adr6006 |
| URL | https://doi.org/10.1126/science.adr6006 |
| Preprint URL | NR |
| Tools | EVOLVEpro |
Задача и claimed novelty
| Поле | Значение |
|---|---|
| Main biological task | Few-shot active-learning evolution of protein activity, including antibody binding |
| Claimed novelty | NR |
Архитектура
| Поле | Значение |
|---|---|
| Model architecture | Protein-LM embeddings + lightweight regression/random forest + iterative acquisition |
| Input representation | NR |
| Encoder / decoder | NR |
| Generative state | Ranked single or multi-mutant variants for next round |
| Conditioning | User-defined mutational space; single/multi-mutant workflows |
| Sampling | NR |
Training
| Поле | Значение |
|---|---|
| Objectives / losses | NR |
| Dataset | NR |
| Preprocessing | NR |
Inference
NR
Baselines и retrospective benchmarks
| Поле | Значение |
|---|---|
| Baselines | NR |
| Retrospective | NR |
Prospective experimental validation
| Поле | Значение |
|---|---|
| Prospective | Up to 100× property improvement across tasks; cannot generalize this number to antibodies |
| Tested designs | Six proteins across reported applications; minimal rounds can use ~10 labels, but campaign sizes vary |
| Success criterion | Up to 100× property improvement across tasks; cannot generalize this number to antibodies |
| Main results | Up to 100× property improvement across tasks; cannot generalize this number to antibodies |
Limitations
Not zero-shot; performance depends on assay noise/library definition; active-learning rounds and experimental turnaround dominate.
Что статья показывает
- Up to 100× property improvement across tasks; cannot generalize this number to antibodies
Чего статья не показывает
- Predicted confidence, likelihood, pLDDT, docking or developability scores are not experimental affinity.
Размеченные утверждения
| Утверждение | Тип | Локатор |
|---|---|---|
| Few-shot learning for protein evolution with EVOLVEpro | Interpretation | DOI/URL from landscape matrix |
| Publication status: Peer-reviewed journal; Science (Q1) | Author-reported | landscape matrix, publication status column |
Review queue
- Bibliographic title must be verified against the PDF.
Source-grounded разбор статьи
Недоступные или неполные источники
| Источник | Почему недоступен | Что использовано вместо |
|---|---|---|
| Science PDF/HTML 10.1126/science.adr6006 | Paywall | bioRxiv 10.1101/2024.07.17.604015 |
| Data S1 / SI files | Не скачаны | Methods narrative in preprint HTML |
Primary question: можно ли few-shot supervised model on frozen PLM embeddings ускорить directed evolution при десятках experimental labels per round.
Полная библиография и статус
[PAPER]Jiang et al. Rapid protein evolution by few-shot learning with a protein language model. bioRxiv DOI 10.1101/2024.07.17.604015. Peer-reviewed counterpart: Rapid in silico directed evolution by a protein language model with EVOLVEpro. Science DOI 10.1126/science.adr6006 (2025). Tool: EVOLVEpro.
Исследовательский вопрос
[PAPER]Как выбрать следующие mutants при малом числе assay labels, используя pretrained PLM features, без fine-tune LM на fitness.
Архитектура и математическая формулировка
[CODE]Frozen embeddings $E(s)$ + top-layer regressor. Released code default selection $\mathrm{top}_k \hat y$, не UCB. См. active learning и tool page.
Training objective и данные
[PAPER]Grid search: 12 DMS. Prospective: Cas12f, PE2, Bxb1, T7 RNAP, REGN10987. Splits DMS — simulation over remaining variants, не независимый белок.
Inference
[PAPER]Refit RF each round; rank unlabelled embeddings. 16/round typical in DMS sim; antibody campaign 11/round × 5 (preprint).
Baselines, ablations, retrospective
[PAPER]Baselines: random; other PLMs; other regressors; PCA. Ablation-like grid search. Retrospective: DMS recovery of high activity.
Prospective evidence
[PAPER]Functional campaigns + antibody ELISA/IC50 (preprint numbers). Не developability panel как главная endpoint.
Главные утверждения авторов и что их поддерживает
[PAPER]Few-shot PLM+surrogate эффективен vs random на DMS (retrospective + ablation-like). Prospective property improvements на нескольких белках. Antibody example — one campaign, ELISA proxy.
Что evidence не поддерживает
- Универсальный 100× для всех antibodies (Science wording не сверен; не обобщать)
- UCB как released algorithm
- Zero-shot без labels
- De novo structure generation
- Likelihood PLM = activity (авторы показывают расхождение)
Ограничения авторов
NRЗависимость от library и assay; не zero-shot. Полный список Science Discussion — NR без PDF.
Дополнительные ограничения интерпретации
[CODE]Hardcoded cluster paths in code; Science vs preprint antibody name; SI unread.
SI и файлы репозитория
[PAPER]Data S1 упоминается в preprint (PCA). Repo files: evolvepro/src/model.py, evolve.py. Zenodo 10.5281/zenodo.13899358 не инспектировался.
Источники
- [1] bioRxiv 10.1101/2024.07.17.604015 · DOI · источник
- [2] Science 10.1126/science.adr6006 · DOI · источник
Нерешённые научные вопросы
- Постраничный diff Science vs preprint