ESM3

ESM-3 ESM 3

Peer-reviewed Prospective flag

Summary card

Поле	Значение
Год / version	2025 / NR
Last verified	2026-08-13
Что генерирует	Sequence, structure tokens, function annotations; can co-generate modalities
Форматы	Primarily proteins/monomers; complex design evidence is limited
De novo или optimization	Both de novo and local completion/redesign
DOI	10.1126/science.ads0018
Paper	https://doi.org/10.1126/science.ads0018
GitHub / repo	https://github.com/Biohub/esm
Service / docs	NR
Лицензия	Code/model-specific terms; open small model is non-commercial
Доступность / commercial-use	Small model local; larger models/API access vary

Задача и generated object

Used in these tasks: Генерация белковых sequences, Joint sequence–structure generation, Дизайн функционального сайта

Part of these workflows: Sequence-first de novo generation, Joint / multimodal sequence–structure generation

Поле	Значение
Required input	Any partial combination of sequence, structure and function prompts
Conditioning	Residue masks, partial sequence/structure, functional keywords/motifs
Fixed	NR
Editable	NR
Generated	Sequence, structure tokens, function annotations; can co-generate modalities
Single mutations	Yes via masks
Combinatorial mutations	Yes
Multi-objective	Prompt-level multi-conditioning, not validated Pareto optimization

Architecture and operating principle

Author-reported (matrix)

Поле	Значение
Base architecture	NR
Model type	1 + 4 Joint sequence–structure generator
Representation	NR
Generation unit	Sequence, structure tokens, function annotations; can co-generate modalities
Conditioning	Residue masks, partial sequence/structure, functional keywords/motifs
Output	Sequence, structure tokens, function annotations; can co-generate modalities
Sampling	NR
Internal scorer	Model likelihood/confidence; no experimentally calibrated universal fitness ranker
Structure generation	Structure decoding is part of the same multimodal model

Input representation

Поле	Значение
Node features	NR
Edge / pair features	NR

Encoder и decoder

Поле	Значение
Encoder	NR
Decoder	NR

Latent or generative state

Conditioning mechanism

Residue masks, partial sequence/structure, functional keywords/motifs

Generation mechanism

Masked multimodal Transformer; iterative unmasking of sequence, structure and function tracks

Sampling: NR

Training objectives

Losses: NR

Mathematical details: NR

Integrated evaluation

Model likelihood/confidence; no experimentally calibrated universal fitness ranker

Architectural limitations

Commercial-use terms differ by asset.
multimodal confidence is not functional evidence.
complex/antibody proof is limited.

Uses these architecture components: Protein language models, Masked-token infilling, Joint sequence–structure generation

Generator versus evaluation

Component	Role	Architecture class	Integrated or external	Required or optional	Influences generation or post hoc	Runtime included or excluded
ESM3 multimodal LM	primary_generator	Masked multimodal Transformer; iterative unmasking of sequence, structure and function tracks	integrated	required	generation	included
Model likelihood/confidence; no experimentally calibrated universal fitness ranker	integrated_scorer	NR	integrated	optional	post_hoc	unknown
Structure decoding is part of the same multimodal model	integrated_structure_predictor	NR	integrated	optional	post_hoc	included
Independent folding/complex prediction and experimental assays	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	4 prospective functional validation
Retrospective	esmGFP and additional design demonstrations; exact total by modality: NR
Prospective binding	Functional fluorescent protein demonstrated; not antibody evidence
Prospective function	Functional fluorescent protein demonstrated; not antibody evidence
Developability	NR
Tested design count	esmGFP and additional design demonstrations; exact total by modality: NR
Success criterion	Functional fluorescent protein demonstrated; not antibody evidence
Independent validation	NR
Summary	Functional fluorescent protein demonstrated; not antibody evidence

Limitations and failure modes

Commercial-use terms differ by asset.
multimodal confidence is not functional evidence.
complex/antibody proof is limited.

Deployment: Limited pilot; legal review required

Reproducibility, license, commercial use

Поле	Значение
Repo	https://github.com/Biohub/esm
License	Code/model-specific terms; open small model is non-commercial
Commercial-use status	Small model local; larger models/API access vary

Alternative tools: NR

Complementary: NR

Related: NR

Not directly comparable with

Efficient evolution with ESM-1v, Structural evolution (ESM-IF1 complex)

Сравнение допустимо только внутри совместимых категорий. Runtime и quality AntiFold vs RFantibody нельзя читать как одну шкалу: разные starting points и generated objects.

Depends on these external modules

Primary paper и полный список источников

Simulating 500 million years of evolution with a language model

Source matrix: landscape xlsx, last verified 2026-08-13.

Детальный source-grounded разбор

1. Что именно делает инструмент

[MATRIX]Вход: Any partial combination of sequence, structure and function prompts

[MATRIX]Generated object: Sequence, structure tokens, function annotations; can co-generate modalities

[MATRIX]Фиксируется: NR. Изменяется: NR.

[MATRIX]De novo или parent-based: Both de novo and local completion/redesign.

2. Архитектура в одном абзаце

[MATRIX]Masked multimodal Transformer; iterative unmasking of sequence, structure and function tracks

4. Математическая модель

NRТочные уравнения, losses и sampling solver будут внесены после чтения paper/code. Не подставляйте родовой UCB/transformer/diffusion, если он не подтверждён для этого инструмента.

Поля NR: conditional probability, update equations, sampling algorithm

5. Физическое устройство состояния

[MATRIX]Seed generated state / output: Sequence, structure tokens, function annotations; can co-generate modalities. Structure generation: Structure decoding is part of the same multimodal model.

6. Физико-химическая интерпретация

[INTERPRETATION]Координаты, frames, PLM embeddings или residue tokens не являются расчётом ΔG, электростатики или растворителя, пока это не сделано внешней energy function. Likelihood / pLDDT / docking score ≠ экспериментальный binding.

7. Training

[MATRIX]Данные: NR. Fine-tuning: NR. Negative examples: NR.

8. Inference

[MATRIX]Sampling: NR. Generation steps: NR. Runtime boundary: NR.

10. Почему, по мнению авторов, метод эффективен

[MATRIX]Author-reported summary (матрица): Functional fluorescent protein demonstrated; not antibody evidence

NRCausal support не классифицирован, пока не прочитаны ablation и experimental sections.

12. Ablation analysis

NRAblation в этом выпуске не разобрана.

13. What the model does not represent

Явный растворитель и ионная сила — обычно отсутствуют в генераторе.
Протонирование, гликозилирование, кофакторы — обычно отсутствуют, если не оговорено.
Конформационная энтропия и долгая динамика — не моделируются одним forward pass.
Экспрессия, агрегация, иммуногенность — не равны likelihood генератора.
Истинная binding free energy — не равна internal score.

14. Evidence and limitations

[MATRIX]Evidence class (матрица): 4 prospective functional validation. Prospective binding: Functional fluorescent protein demonstrated; not antibody evidence. Function: Functional fluorescent protein demonstrated; not antibody evidence. Developability: NR.

Commercial-use terms differ by asset.
multimodal confidence is not functional evidence.
complex/antibody proof is limited.

15. Reproducibility

NRРепозиторий: https://github.com/Biohub/esm. Лицензия: Code/model-specific terms; open small model is non-commercial. Commit: NR, пока страница не имеет отдельного analysis YAML.

Модульная схема

Модуль	Класс модели	Вход	Выход	Обучается или frozen	Generator или evaluator
ESM3 multimodal LM	Masked multimodal Transformer; iterative unmasking of sequence, structure and function tracks	NR	NR	NR	primary_generator
Model likelihood/confidence; no experimentally calibrated universal fitness ranker	NR	NR	NR	NR	integrated_scorer
Structure decoding is part of the same multimodal model	NR	NR	NR	NR	integrated_structure_predictor
Independent folding/complex prediction and experimental assays	NR	NR	NR	NR	mandatory_external_dependency
Experimental binding/function assays	NR	NR	NR	NR	experimental_validation

Generator versus evaluator

Компонент	Роль	Integrated или external	Required или optional	Влияет на generation или только ранжирует после
ESM3 multimodal LM	primary_generator	integrated	required	generation
Model likelihood/confidence; no experimentally calibrated universal fitness ranker	integrated_scorer	integrated	optional	post_hoc
Structure decoding is part of the same multimodal model	integrated_structure_predictor	integrated	optional	post_hoc
Independent folding/complex prediction and experimental assays	mandatory_external_dependency	external	required	post_hoc
Experimental binding/function assays	experimental_validation	external	required	post_hoc

Claim–evidence matrix

Утверждение	Предполагаемый механизм	Тип evidence	Датасет / эксперимент	Comparator	Метрика / результат	Источник	Оговорка
Functional fluorescent protein demonstrated; not antibody evidence	Masked multimodal Transformer; iterative unmasking of sequence, structure and function tracks	5. prospective functional validation	esmGFP and additional design demonstrations; exact total by modality: NR	NR	Functional fluorescent protein demonstrated; not antibody evidence	[MATRIX] landscape matrix 2026-08-13	Не читать как полный разбор primary paper.