ESM-IF1
ESM-IF1 ESM Inverse Folding
ICML 2022 main track (A*)Summary card
| Поле | Значение |
|---|---|
| Год / version | 2022 / NR |
| Last verified | 2026-08-27 |
| Что генерирует | Sequence and sequence likelihood conditioned on backbone |
| Форматы | Monomer and protein complex |
| De novo или optimization | Design or masked redesign |
| DOI | 10.48550/arXiv.2204.07148 |
| Paper | https://proceedings.mlr.press/v162/hsu22a.html |
| GitHub / repo | https://github.com/facebookresearch/esm |
| Service / docs | NR |
| Лицензия | MIT |
| Доступность / commercial-use | Code + weights |
Задача и generated object
Used in these tasks: Fixed-backbone inverse folding, Предложение single mutations
Part of these workflows: Fixed-backbone inverse folding
| Поле | Значение |
|---|---|
| Required input | N–Cα–C backbone coordinates; partial masks allowed |
| Conditioning | Partial masks, complexes, interfaces, multistate use cases |
| Fixed | NR |
| Editable | NR |
| Generated | Sequence and sequence likelihood conditioned on backbone |
| Single mutations | Yes |
| Combinatorial mutations | Yes |
| Multi-objective | No explicit multi-objective optimizer |
Architecture and operating principle
Author-reported (matrix)| Поле | Значение |
|---|---|
| Base architecture | NR |
| Model type | 2 Fixed-backbone inverse-folding generator |
| Representation | NR |
| Generation unit | Sequence and sequence likelihood conditioned on backbone |
| Conditioning | Partial masks, complexes, interfaces, multistate use cases |
| Output | Sequence and sequence likelihood conditioned on backbone |
| Sampling | NR |
| Internal scorer | Conditional log-likelihood; can score mutations |
| Structure generation | No |
Input representation
NR
| Поле | Значение |
|---|---|
| Node features | NR |
| Edge / pair features | NR |
Encoder и decoder
NR
| Поле | Значение |
|---|---|
| Encoder | NR |
| Decoder | NR |
Latent or generative state
NR
Conditioning mechanism
Partial masks, complexes, interfaces, multistate use cases
Generation mechanism
GVP geometric encoder + autoregressive Transformer
Sampling: NR
Training objectives
NR
Losses: NR
Mathematical details: NR
Integrated evaluation
Conditional log-likelihood; can score mutations
Architectural limitations
- Fixed-backbone assumption.
- scores fold compatibility more directly than therapeutic function.
- legacy repository.
Uses these architecture components: Inverse folding, Geometric transformers, SE(3)-equivariant networks, Invariant models
Generator versus evaluation
| Component | Role | Architecture class | Integrated or external | Required or optional | Influences generation or post hoc | Runtime included or excluded |
|---|---|---|---|---|---|---|
| ESM-IF1 inverse-folding model | primary_generator | GVP geometric encoder + autoregressive Transformer | integrated | required | generation | included |
| Conditional log-likelihood; can score mutations | integrated_scorer | NR | integrated | optional | post_hoc | unknown |
| Backbone/complex quality must be independently checked | mandatory_external_dependency | NR | external | required | post_hoc | excluded |
| Experimental binding/function assays | experimental_validation | NR | external | required | post_hoc | excluded |
Training data и biases
| Поле | Значение |
|---|---|
| Training data | NR |
| Preprocessing | NR |
| Negative examples | NR |
| Fine-tuning | NR |
| Label requirements | NR |
| Training hardware | NR |
| Training time | NR |
Inference hardware и runtime
| Поле | Значение |
|---|---|
| Candidate count | NR |
| Batch size | NR |
| Generation steps | NR |
| Recycles | NR |
| Size | NR |
| Reported runtime | NR |
| Runtime boundary | NR |
| CPU | NR |
| GPU | NR |
| VRAM | NR |
| RAM | NR |
| Disk | NR |
| Databases | NR |
Experimental evidence
| Поле | Значение |
|---|---|
| Evidence class | 2 retrospective in primary paper; 4 when embedded in later structural-evolution workflow |
| Retrospective | Primary paper: no prospective wet lab |
| Prospective binding | NR |
| Prospective function | NR |
| Developability | NR |
| Tested design count | Primary paper: no prospective wet lab |
| Success criterion | Retrospective recovery/generalization; later antibody evolution evidence belongs to separate workflow |
| Independent validation | NR |
| Summary | Retrospective recovery/generalization; later antibody evolution evidence belongs to separate workflow |
Limitations and failure modes
- Fixed-backbone assumption.
- scores fold compatibility more directly than therapeutic function.
- legacy repository.
Deployment: Immediate retrospective evaluation
Reproducibility, license, commercial use
| Поле | Значение |
|---|---|
| Repo | https://github.com/facebookresearch/esm |
| License | MIT |
| Commercial-use status | Code + weights |
Related tools
Alternative tools: NR
Complementary: NR
Related: NR
Not directly comparable with
EvoDiff, FrameDiff, ProGen / ProGen2, RFdiffusion, RFdiffusion All-Atom (RFdiffusionAA)Depends on these external modules
NR
Related glossary terms
NRPrimary paper и полный список источников
Learning inverse folding from millions of predicted structuresSource matrix: landscape xlsx, last verified 2026-08-13.
Детальный source-grounded разбор
Недоступные или неполные источники
| Источник | Почему недоступен | Что использовано вместо |
|---|---|---|
| PMLR PDF Table A.1 полная сетка гиперпараметров | В bioRxiv HTML таблица A.1 не развёрнута численно | Текст Appendix: 8 enc/8 dec Transformer, 8 heads, d=512, 142M; GVP k=30 в маскированном режиме |
| Line-by-line esm/inverse_folding/* на указанном commit | В этом выпуске читался paper + repo SHA, не каждый .py | Архитектура из Hsu §2 / Appendix A.3 |
ESM-IF1 (GVP-Transformer) — inverse-folding generator $p(Y\mid X)$ по координатам N/Cα/C. Sequence recovery и zero-shot likelihood ratio не равны experimental $K_d$. Released модель — гибрид 4 GVP-GNN + 8+8 Transformer (~142M), обученный на CATH + 12M AF2-предсказанных структур.
Происхождение источников
| Метка | Источник | Ссылка | Примечание |
|---|---|---|---|
| [PAPER] | Hsu et al., ICML 2022 / PMLR 162:8946–8970; bioRxiv 10.1101/2022.04.10.487779 | https://proceedings.mlr.press/v162/hsu22a.html | Детали Methods — по bioRxiv HTML |
| [PAPER] | Jing et al., ICLR 2021 — GVP layers | https://openreview.net/forum?id=1YLJDvSx6J4 | NR |
| [CODE] | facebookresearch/esm (ESM-IF1 weights live here historically) | https://github.com/facebookresearch/esm | commit 2b369911bb5b4b0dda914521b9475cad1656b2ac |
| Поле | Значение |
|---|---|
| Репозиторий | https://github.com/facebookresearch/esm |
| Commit | 2b369911bb5b4b0dda914521b9475cad1656b2ac |
| Ветка / tag | main |
| Лицензия | MIT (esm repository; verify COPYING) |
Проверенные файлы реализации:
README.md (IF1 usage)esm/inverse_folding (не построчно в этом выпуске)
1. Что именно делает инструмент
[PAPER]Вход: backbone N, Cα, C coordinates (не side chains). Комплексы: цепи конкатенируются с 10 mask tokens между цепями. Частично masked backbone допустим (span mask train). bioRxiv: complexes; span masking up to 30 AA until 15% coords masked
[PAPER]Выход: авторегрессивная последовательность $Y$ (sampling или argmax при $T=10^{-6}$ для recovery). Backbone не генерируется.
[PAPER]Fixed-backbone inverse folding, не de novo fold. Zero-shot mutation scoring через likelihood ratio — evaluator mode той же сети, не отдельный affinity head.
2. Архитектура в одном абзаце
[PAPER]GVP-GNN encoder извлекает геометрию (rotation-equivariant vectors + invariant scalars). Затем change of basis в локальные residue frames → invariant scalars → generic Transformer encoder–decoder (learned positional embeddings, pre-LN). Лучшая модель: 4 GVP-GNN encoder layers, 8 Transformer encoder, 8 decoder, 8 heads, $d=512$, 142M. Инвариантность $p(Y\mid X)=p(Y\mid TX)$ для $T\in\mathrm{SE}(3)$ (Appendix A.3). §GVP-Transformer; Appendix A.3
4. Математическая модель
| Символ | Тип | Определение |
|---|---|---|
| $X$ | coordinates | backbone atom xyz, 3 atoms × n residues |
| $Y$ | token sequence | amino acids |
| $T$ | SE(3) transform | глобальный поворот/сдвиг входа |
5. Физическое устройство состояния
[PAPER]Атомы остова в $\mathbb{R}^3$; GVP vectors — геометрические направления; scalars — invariant. Sequence — AA identity. pLDDT AF2 — входной признак (RBF) для predicted structures, не выход ESM-IF1. 10 gap tokens — цепной разделитель, не физический линкер.
7. Training
[PAPER]CATH topology 80/10/10: 16153 / 1457 / 1797. Плюс 12M UniRef50 AF2 structures (Amber-relaxed; sequences ≤500; ranked by MSA Transformer distogram lDDT). Mix per epoch: experimental set + 10% random sample of 12M → ~1:80 experimental:predicted. pLDDT < 90 маскируется (~25% predicted coords). Binary token experimental vs predicted. Span masking: spans ≤30 AA until 15% coords masked.
[PAPER]Small GVP-GNN (1M, 3+3) деградирует на predicted data. Large GVP-GNN (21M, 8+8) и GVP-Transformer выигрывают. Overfit CATH-only при scaling 1M→21M (42.2% → 39.2% recovery). Table 1; Fig. 6
8. Inference
[PAPER]Sampling из $p(Y\mid X)$; для max recovery $T=10^{-6}$. Partial backbone: masked nodes excluded from GVP then zeros into Transformer (Appendix A.2). Multi-state: geometric mean of likelihoods $p(Y\mid A)$ and $p(Y\mid B)$ как proxy $p(Y\mid A,B)$.
NRDefault CLI temperature / batch в этом выпуске: NR (код IF1 не построчно). Не путать с ESM-2 masked LM.
10. Почему, по мнению авторов, метод эффективен
[PAPER]Mechanism. Predicted structures ×3 orders of magnitude снимают data bottleneck; GVP даёт геометрию, Transformer — long-range / masked spans, которые чистый GVP-GNN не держит (Fig. 4).
[PAPER]Author claim. 51% native recovery (72% buried) on structurally held-out backbones; ~10 pp over prior GVP-GNN on CATH-only; generalization to complexes, partial masks, multi-state, zero-shot mutation.
[PAPER]Evidence. Retrospective CATH topology split; ablations Fig. 6 / Table 1 / C.1; zero-shot Rocklin stability, SKEMPI, Starr ACE2–RBD Spearman 0.69 when given RBD+ACE2 coords. No prospective designed-and-tested binders in this paper.
[INTERPRETATION]Causal support. Predicted-structure mix — ablation Fig. 6a–b (directly supported). Spearman 0.69 — retrospective correlation, не affinity maturation campaign.
12. Ablation analysis
[PAPER]CATH-only vs +AF2: GVP-Transformer recovery 38.3% → 51.6%; GVP-GNN-large 39.2% → 50.8%; tiny GVP-GNN 42.2% → 38.6% (деградация). Span vs independent masking: span лучше на длинных масках (Fig. 4). Число GVP layers перед Transformer помогает (Fig. C.1). Complex coords vs single chain: substantial perplexity gain (Table 2). RBM coords / ACE2 coords necessary for RBD affinity correlation (Table 3).
13. What the model does not represent
- Растворитель, протонирование, гликозилирование, кофакторы
- Side-chain packing как отдельный decoder
- Настоящая $\Delta G$; Spearman на Starr scan ≠ designed antibody Kd
- De novo backbone
- Antibody-specific fine-tune (это ESM-IF1; AntiFold — отдельный FT)
14. Evidence and limitations
[INTERPRETATION]Авторы честно показывают, что IF likelihoods коррелируют с некоторыми mutational assays. Это не prospective design evidence. AntiFold позже FT эту архитектуру на SAbDab+OAS — см. /tools/antifold.html. Structural evolution (Shanker/Hie Science) использует IF1-complex scoring — отдельный инструмент каталога.
15. Reproducibility
[CODE]Weights historically esm_if1_gvp4_t16_142M_UR50. Репозиторий esm на 2023-06-27 (commit выше) — maintenance; Meta перемещала модели. Точный inference example в этом выпуске не прогонялся. repo SHA only
Физико-химический смысл
| Признак или операция модели | Физико-химическая интерпретация | Прямо / косвенно / внешне | Важное ограничение |
|---|---|---|---|
| N/Cα/C coordinates | остов полипептида | прямо | нет side chain / solvent |
| GVP vector channels | локальные геометрические направления | equivariant inductive bias | не сила взаимодействия |
| Transformer over invariant scalars | sequence-structure long-range stats | learned | не SE(3) внутри attention |
| AF2 pLDDT as input RBF | confidence predicted structure | прямо как feature predicted examples | не выход IF1 и не affinity |
Модульная схема
| Модуль | Класс модели | Вход | Выход | Обучается или frozen | Generator или evaluator |
|---|---|---|---|---|---|
| GVP-GNN encoder (4 layers) | geometric GNN | backbone graph | scalar+vector features → local-frame invariant scalars | trained | encoder of generator |
| Transformer enc/dec 8+8 | AR seq2seq | invariant tokens + partial Y | next-AA logits | trained | generator |
| likelihood ratio | same AR model | WT structure + mutant sequence | scalar score | frozen | evaluator only |
Generator versus evaluator
| Компонент | Роль | Integrated или external | Required или optional | Влияет на generation или только ранжирует после |
|---|---|---|---|---|
| GVP-Transformer decoder | primary_generator | integrated | required | generation |
| log p / likelihood ratio | integrated_scorer | integrated | optional | post_hoc (zero-shot); не меняет sample path unless used as objective externally |
| AlphaFold2 (training data only) | structure predictor for pretraining corpus | external | required for the 12M set, not at IF1 inference | training distribution, not inference coords unless user supplies predicted backbone |
Claim–evidence matrix
| Утверждение | Предполагаемый механизм | Тип evidence | Датасет / эксперимент | Comparator | Метрика / результат | Источник | Оговорка |
|---|---|---|---|---|---|---|---|
| Predicted structures improve IF recovery for large models | 12M AF2 UniRef50 mix 1:80 | 3. ablation evidence | CATH 4.3 topology test | same architecture, CATH-only | GVP-Transformer 38.3% → 51.6% recovery | [PAPER] Table 1 | Recovery ≠ experimental fold of designed sequences in this paper |
| GVP-Transformer держит masked spans лучше GVP-GNN | Transformer after local GVP | 3. ablation evidence | variable mask lengths | GVP-GNN-large | GVP-GNN → background perplexity for masks >5; Transformer remains moderate | [PAPER] Fig. 4 | in silico |
| Zero-shot IF likelihood correlates with ACE2–RBD binding | log p on RBM given complex coords | 2. retrospective benchmark | Starr et al. 2020 deep mutational scan | ESM-1v / sequence-only; ablate ACE2 or RBM coords | Spearman 0.69 with full RBD+ACE2; weaker without ACE2; none if RBM masked | [PAPER] Table 3 | Корреляция на мутациях WT, не designed antibodies; не Kd новых молекул |
| Model is generally better at protein design | — | 1. только architectural rationale | n/a | n/a | нет prospective designs в Hsu 2022 | [INTERPRETATION] experimental evaluation is in silico / zero-shot DMS | Не утверждать wet-lab success ESM-IF1 как generator |
Различия между статьёй и реализацией
| Версия статьи | Commit репозитория | Checkpoint / model | Файл | Различие | Практическое следствие |
|---|---|---|---|---|---|
| ICML 2022 / bioRxiv v1 HTML | 2b369911bb5b4b0dda914521b9475cad1656b2ac | esm_if1_gvp4_t16_142M_UR50 (name from ESM docs; not re-opened this pass) | esm/inverse_folding (not line-audited) | Paper describes three architectures; released ESM-IF1 is the GVP-Transformer. Exact CLI defaults NR this pass. | Не описывать released IF1 как чистый GVP-GNN Jing. |
Схемы information flow
flowchart LR X["Backbone N/CA/C"] --> GVP["4× GVP-GNN"] GVP --> F["Local frames → invariant tokens"] F --> TE["8× Transformer encoder"] TE --> TD["8× Transformer decoder AR"] TD --> Y["Sequence Y"] Y -.->|"optional"| LR["likelihood ratio vs WT"]
Источники
- [1] Hsu C. et al. Learning inverse folding from millions of predicted structures. PMLR 162:8946–8970 (2022). · DOI · источник
- [2] Jing B. et al. Learning from protein structure with geometric vector perceptrons. ICLR 2021. · источник
- [3] facebookresearch/esm commit 2b369911. · источник
Нерешённые научные вопросы
- Table A.1 optimizer / batch / lr — NR numerically in HTML
- Построчный audit InverseFoldingModel.forward
- Совпадение checkpoint name с 142M GVP-Transformer — по документации ESM, не по hash файла весов