Learning inverse folding from millions of predicted structures
ICML 2022 main track (A*)| Поле | Значение |
|---|---|
| Authors | Chloe Hsu, Robert Verkuil, Jason Liu, Zeming Lin, Brian Hie, Tom Sercu, Adam Lerer, Alexander Rives |
| Year | 2022 |
| Venue / status | ICML 2022 main track (A*) |
| DOI | 10.48550/arXiv.2204.07148 |
| URL | https://proceedings.mlr.press/v162/hsu22a.html |
| Preprint URL | https://doi.org/10.1101/2022.04.10.487779 |
| Tools | ESM-IF1 |
Задача и claimed novelty
| Поле | Значение |
|---|---|
| Main biological task | Fixed-backbone inverse folding; interface/multistate sequence design |
| Claimed novelty | NR |
Архитектура
| Поле | Значение |
|---|---|
| Model architecture | GVP geometric encoder + autoregressive Transformer |
| Input representation | NR |
| Encoder / decoder | NR |
| Generative state | Sequence and sequence likelihood conditioned on backbone |
| Conditioning | Partial masks, complexes, interfaces, multistate use cases |
| Sampling | NR |
Training
| Поле | Значение |
|---|---|
| Objectives / losses | NR |
| Dataset | NR |
| Preprocessing | NR |
Inference
NR
Baselines и retrospective benchmarks
| Поле | Значение |
|---|---|
| Baselines | NR |
| Retrospective | Retrospective recovery/generalization; later antibody evolution evidence belongs to separate workflow |
Prospective experimental validation
| Поле | Значение |
|---|---|
| Prospective | NR |
| Tested designs | Primary paper: no prospective wet lab |
| Success criterion | Retrospective recovery/generalization; later antibody evolution evidence belongs to separate workflow |
| Main results | Retrospective recovery/generalization; later antibody evolution evidence belongs to separate workflow |
Limitations
Fixed-backbone assumption; scores fold compatibility more directly than therapeutic function; legacy repository.
Что статья показывает
- Retrospective recovery/generalization; later antibody evolution evidence belongs to separate workflow
Чего статья не показывает
- Predicted confidence, likelihood, pLDDT, docking or developability scores are not experimental affinity.
Размеченные утверждения
| Утверждение | Тип | Локатор |
|---|---|---|
| Learning inverse folding from millions of predicted structures | Author-reported | DOI/URL from landscape matrix |
| Publication status: ICML 2022 main track (A*) | Author-reported | landscape matrix, publication status column |
Source-grounded разбор статьи
ICML paper по inverse folding с 12M AF2 structures. Основной evidence — CATH recovery, ablations и zero-shot DMS correlations. Не путать Spearman 0.69 на Starr RBD с designed-binder validation.
Исследовательский вопрос
[PAPER]Улучшает ли обучение на миллионах predicted structures inverse folding (GVP vs GVP-Transformer) и zero-shot mutation scoring?
Архитектура
[PAPER]GVP-Transformer 142M: 4 GVP + 8+8 Transformer; $p(Y\mid X)=p(Y\mid TX)$. См. /tools/esm-if1.html.
Утверждения и evidence
[PAPER]Recovery 51.6% vs 38.3% CATH-only; buried 72%. Table 3 Starr Spearman 0.69 retrospective mutational scan, требует ACE2+RBM coords. Нет prospective designed proteins.
Что evidence не поддерживает
- Экспериментально подтверждённые de novo sequences от ESM-IF1 в этой статье
- Что likelihood ratio = $\Delta\Delta G$ или Kd
Ограничения авторов
[PAPER]Predicted structures can leak topology if not filtered (Gene3D/Foldseek); single-point SKEMPI may not improve with extra AF2 data; 100% recovery impossible because many sequences share a fold.
Источники
- [1] Hsu et al. PMLR 162:8946–8970 (2022). · источник
Нерешённые научные вопросы
- Table A.1 numeric hyperparameters