esm

Name: esm
Author: adaptyvbio

31installs20forks

How to install esm?

adaptyvbio/protein-design-skills/esm

$npx -y skills add adaptyvbio/protein-design-skills --skill esm

Installs into the current project.

Summary

ESM protein language models for embeddings, sequence scoring, structure prediction, and binder design. Use this skill when: (1) Computing pseudo-log-likelihood (PLL) or mutation-effect scores, (2) Getting protein embeddings for clustering or filtering, (3) Predicting complex stru

Files · 1

View on GitHub

SKILL.md

1	# ESM Protein Language Models
2
3	The ESM line is maintained at [github.com/Biohub/esm](https://github.com/Biohub/esm)
4	(Chan Zuckerberg Biohub, MIT license; the older `evolutionaryscale/esm` URL
5	redirects here). The current generation ships three artifacts: ESM C (language
6	model), ESMFold2 (structure prediction), and ESM Atlas (a map of predicted
7	structures). Weights are on [huggingface.co/biohub](https://huggingface.co/biohub);
8	the hosted API is at `biohub.ai`.
9
10	This skill covers ESM C, ESMFold2, and legacy ESM2. ESM3 is not covered because its
11	open weights are non-commercial.
12
13	## Which model to use
14
15	\| Task \| Model \|
16	\|------\|-------\|
17	\| Embeddings, PLL, mutation scoring \| ESM C (ESMC-6B), or ESM2 for a lighter run \|
18	\| Complex structure prediction \| ESMFold2 \|
19	\| High-throughput single-sequence folding \| ESMFold2 fast mode \|
20	\| Binder design \| ESMFold2 inversion (see below), or the `mosaic` / `bindcraft` skills \|
21	\| Variant effect / zero-shot scoring \| ESM C or ESM2 \|
22
23	## Prerequisites
24
25	\| Requirement \| Minimum \| Recommended \|
26	\|-------------\|---------\|-------------\|
27	\| Python \| 3.10+ \| 3.11 \|
28	\| PyTorch \| 2.0+ \| Latest \|
29	\| CUDA \| 12.0+ \| 12.1+ \|
30	\| GPU VRAM \| 24GB (ESM2 / small ESMC) \| 80GB (ESMC-6B, ESMFold2) \|
31
32	## ESM C: embeddings and scoring
33
34	ESM C is the successor to ESM2. It improves long-range structural understanding as
35	model scale grows and is the default choice for embeddings, pseudo-log-likelihood,
36	and mutation-effect scoring.
37
38	### Python (Hugging Face)
39	```python
40	from transformers import AutoModelForMaskedLM, AutoTokenizer
41	import torch
42
43	model_id = "biohub/ESMC-6B"
44	tok = AutoTokenizer.from_pretrained(model_id)
45	model = AutoModelForMaskedLM.from_pretrained(
46	model_id, output_hidden_states=True, torch_dtype=torch.bfloat16
47	).eval().cuda()
48
49	batch = tok(["MKTAYIAKQRQISFVK..."], return_tensors="pt").to("cuda")
50	with torch.no_grad():
51	out = model(**batch)
52
53	logits = out.logits # for PLL / mutation scoring
54	embeddings = out.hidden_states[-1] # per-residue representations
55	```
56
57	Install the package with `pip install esm@git+https://github.com/Biohub/esm.git@main`.
58
59	### Hosted API
60	```python
61	from esm.sdk import esmc_client
62	from esm.sdk.api import ESMProtein, LogitsConfig
63
64	model = esmc_client(model="esmc-600m-2024-12", url="https://biohub.ai", token="<API token>")
65	tensor = model.encode(ESMProtein(sequence="MKTAYIAKQRQISFVK..."))
66	out = model.logits(tensor, LogitsConfig(sequence=True, return_embeddings=True))
67	```
68
69	ESMC-6B has open weights; `esmc-600m` is the smaller API model. For mutation
70	scoring and fine-tuning, see the `esmc_mutation_scoring` and `esmc_finetune`
71	notebooks under [cookbook/tutorials](https://github.com/Biohub/esm/tree/main/cookbook/tutorials).
72
73	## ESMFold2: complex structure prediction
74
75	ESMFold2 is built on ESMC-6B with a diffusion structure head. Unlike the original
76	ESMFold, it predicts complexes (protein, DNA, ligand, and modified residues), takes
77	an optional MSA, and has a single-sequence fast mode for high-throughput screening.
78	It is validated for protein-protein interaction design and leads DockQ pass-rate on
79	Foldbench protein-protein and antibody-antigen complexes.
80
81	### Modal (biomodals)
82	```bash
83	printf '>protein\|A\nMKTAYIAKQRQISFVK...\n' > target.faa
84	uv run --with modal modal run modal_esmfold2.py --input-faa target.faa
85	```
86
87	The FASTA header tags `protein\|`, `dna\|`, `rna\|`, and `ligand\|` (SMILES) let you fold
88	complexes. GPU defaults to A100-40GB (set with `MODAL_GPU`).
89
90	### Python (local weights)
91	```python
92	from transformers.models.esmfold2.modeling_esmfold2 import ESMFold2Model
93	from esm.models.esmfold2 import ProteinInput, StructurePredictionInput, ESMFold2InputBuilder
94
95	model = ESMFold2Model.from_pretrained("biohub/ESMFold2").cuda().eval()
96	spi = StructurePredictionInput(sequences=[ProteinInput(id="A", sequence="BINDER_SEQ")])
97	result = ESMFold2InputBuilder().fold(model, spi, num_loops=20, num_sampling_steps=100)
98	# result.plddt, result.ptm, result.iptm, result.complex.to_mmcif()
99	```
100
101	For single-sequence high-throughput folding, the fast variant is the SDK model string
102	`esmfold2-fast-2026-05` (HF repo `biohub/ESM

People who liked this also viewed

6 picks

Type

esm

31installs20forks

View on GitHub

How to install esm?

adaptyvbio/protein-design-skills/esm

$npx -y skills add adaptyvbio/protein-design-skills --skill esm

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# ESM Protein Language Models
2
3	The ESM line is maintained at [github.com/Biohub/esm](https://github.com/Biohub/esm)
4	(Chan Zuckerberg Biohub, MIT license; the older `evolutionaryscale/esm` URL
5	redirects here). The current generation ships three artifacts: ESM C (language
6	model), ESMFold2 (structure prediction), and ESM Atlas (a map of predicted
7	structures). Weights are on [huggingface.co/biohub](https://huggingface.co/biohub);
8	the hosted API is at `biohub.ai`.
9
10	This skill covers ESM C, ESMFold2, and legacy ESM2. ESM3 is not covered because its
11	open weights are non-commercial.
12
13	## Which model to use
14
15	\| Task \| Model \|
16	\|------\|-------\|
17	\| Embeddings, PLL, mutation scoring \| ESM C (ESMC-6B), or ESM2 for a lighter run \|
18	\| Complex structure prediction \| ESMFold2 \|
19	\| High-throughput single-sequence folding \| ESMFold2 fast mode \|
20	\| Binder design \| ESMFold2 inversion (see below), or the `mosaic` / `bindcraft` skills \|
21	\| Variant effect / zero-shot scoring \| ESM C or ESM2 \|
22
23	## Prerequisites
24
25	\| Requirement \| Minimum \| Recommended \|
26	\|-------------\|---------\|-------------\|
27	\| Python \| 3.10+ \| 3.11 \|
28	\| PyTorch \| 2.0+ \| Latest \|
29	\| CUDA \| 12.0+ \| 12.1+ \|
30	\| GPU VRAM \| 24GB (ESM2 / small ESMC) \| 80GB (ESMC-6B, ESMFold2) \|
31
32	## ESM C: embeddings and scoring
33
34	ESM C is the successor to ESM2. It improves long-range structural understanding as
35	model scale grows and is the default choice for embeddings, pseudo-log-likelihood,
36	and mutation-effect scoring.
37
38	### Python (Hugging Face)
39	```python
40	from transformers import AutoModelForMaskedLM, AutoTokenizer
41	import torch
42
43	model_id = "biohub/ESMC-6B"
44	tok = AutoTokenizer.from_pretrained(model_id)
45	model = AutoModelForMaskedLM.from_pretrained(
46	model_id, output_hidden_states=True, torch_dtype=torch.bfloat16
47	).eval().cuda()
48
49	batch = tok(["MKTAYIAKQRQISFVK..."], return_tensors="pt").to("cuda")
50	with torch.no_grad():
51	out = model(**batch)
52
53	logits = out.logits # for PLL / mutation scoring
54	embeddings = out.hidden_states[-1] # per-residue representations
55	```
56
57	Install the package with `pip install esm@git+https://github.com/Biohub/esm.git@main`.
58
59	### Hosted API
60	```python
61	from esm.sdk import esmc_client
62	from esm.sdk.api import ESMProtein, LogitsConfig
63
64	model = esmc_client(model="esmc-600m-2024-12", url="https://biohub.ai", token="<API token>")
65	tensor = model.encode(ESMProtein(sequence="MKTAYIAKQRQISFVK..."))
66	out = model.logits(tensor, LogitsConfig(sequence=True, return_embeddings=True))
67	```
68
69	ESMC-6B has open weights; `esmc-600m` is the smaller API model. For mutation
70	scoring and fine-tuning, see the `esmc_mutation_scoring` and `esmc_finetune`
71	notebooks under [cookbook/tutorials](https://github.com/Biohub/esm/tree/main/cookbook/tutorials).
72
73	## ESMFold2: complex structure prediction
74
75	ESMFold2 is built on ESMC-6B with a diffusion structure head. Unlike the original
76	ESMFold, it predicts complexes (protein, DNA, ligand, and modified residues), takes
77	an optional MSA, and has a single-sequence fast mode for high-throughput screening.
78	It is validated for protein-protein interaction design and leads DockQ pass-rate on
79	Foldbench protein-protein and antibody-antigen complexes.
80
81	### Modal (biomodals)
82	```bash
83	printf '>protein\|A\nMKTAYIAKQRQISFVK...\n' > target.faa
84	uv run --with modal modal run modal_esmfold2.py --input-faa target.faa
85	```
86
87	The FASTA header tags `protein\|`, `dna\|`, `rna\|`, and `ligand\|` (SMILES) let you fold
88	complexes. GPU defaults to A100-40GB (set with `MODAL_GPU`).
89
90	### Python (local weights)
91	```python
92	from transformers.models.esmfold2.modeling_esmfold2 import ESMFold2Model
93	from esm.models.esmfold2 import ProteinInput, StructurePredictionInput, ESMFold2InputBuilder
94
95	model = ESMFold2Model.from_pretrained("biohub/ESMFold2").cuda().eval()
96	spi = StructurePredictionInput(sequences=[ProteinInput(id="A", sequence="BINDER_SEQ")])
97	result = ESMFold2InputBuilder().fold(model, spi, num_loops=20, num_sampling_steps=100)
98	# result.plddt, result.ptm, result.iptm, result.complex.to_mmcif()
99	```
100
101	For single-sequence high-throughput folding, the fast variant is the SDK model string
102	`esmfold2-fast-2026-05` (HF repo `biohub/ESM

People who liked this also viewed

6 picks

Type