esmfold2

ESMFold2 (Biohub)

All-atom diffusion co-folding from the Biohub ESM release (2026). ESMFold2 = 48 pair layers with MSA support; ESMFold2-Fast = 24 layers, single-sequence only, ~1.7x faster.

License: MIT (code github.com/Biohub/esm + weights HF biohub/*). Paper: "Language Modeling Materializes a World Model of Protein Biology" (2026).

Install

CUDA 12.x GPU (H100/A100-class); Python 3.12 only. Fresh venv; needs egress to HF Hub, GitHub, PyPI:

pip install --no-cache-dir uv
uv venv --python 3.12 /work/venv && source /work/venv/bin/activate
uv pip install \
  "torch>=2.5,<2.8" einops "biotite>=1.0" rdkit msgpack-numpy biopython \
  scikit-learn brotli attrs pandas cloudpathlib httpx tenacity zstd pydssp \
  pygtrie accelerate huggingface_hub safetensors "numpy<3" networkx \
  sentencepiece tokenizers regex packaging filelock pyyaml typing_extensions \
  "transformers @ git+https://github.com/Biohub/transformers.git@3a8956fb4d4ea16b0ec8e71deef2c2909b6a5cbf"
uv pip install --no-deps "esm @ git+https://github.com/Biohub/esm.git@f652b471"
# OPTIONAL — only affects ESMC attention; trunk speedup comes from set_kernel_backend("fused")
uv pip install ninja packaging wheel setuptools
MAX_JOBS=8 uv pip install --no-deps --no-build-isolation "flash-attn<3"
# Do NOT install transformer-engine — RuntimeError (not ImportError) on import
# slips ESMC's guard and kills ESMFold2Model import.

The bundled esmfold2_gpu Modal env (remote-compute-modal skill) is the canonical, version-pinned recipe.

Gotchas:

• Default kernel backend is None (reference PyTorch, ~12x slower than paper). Call model.set_kernel_backend('fused') after from_pretrained(). See section below.
• Match torch CUDA build to your driver; the pin <2.8 targets CUDA 12.2.
• Weights via Xet bridge ~300 MB/s: ESMFold2 1.36 GB, ESMFold2-Fast 0.76 GB. Set HF_HOME=/work/hf_cache.

Usage — local model

from esm.models.esmfold2 import (
    ESMFold2InputBuilder, StructurePredictionInput,
    ProteinInput, DNAInput, RNAInput, LigandInput, Modification,
)
from transformers.models.esmfold2.modeling_esmfold2 import ESMFold2Model

model = ESMFold2Model.from_pretrained("biohub/ESMFold2").cuda().eval()
# or "biohub/ESMFold2-Fast" (24 layers, no MSA, ~1.7x faster)
# or "biohub/ESMFold2-Experimental{,-Fast}{,-Cutoff2025}" (4 design-critic models)

spi = StructurePredictionInput(sequences=[
    ProteinInput(id="A", sequence=target_seq),
    ProteinInput(id="B", sequence=binder_seq),
    # DNAInput(id="C", sequence="ACGT", modifications=[Modification(position=5, ccd="C36")]),
    # RNAInput(id="D", sequence="ACGU"),
    # LigandInput(id="L", ccd=["SAH"]),  # or smiles="..."
])
# Homodimer: ProteinInput(id=["A","B"], sequence=seq)

results = ESMFold2InputBuilder().fold(
    model, spi,
    num_loops=10,             # paper FoldBench eval: 10; 20-loop variant: 20
    num_sampling_steps=68,    # paper eval: 68 (truncated EDM)
    num_diffusion_samples=5,  # paper eval: 5/seed
    seed=0,
)
# fold() returns list[Prediction], one per diffusion sample. Each carries
# .plddt [L], .ptm, .iptm, .pae [L,L], .pair_chains_iptm, .complex.to_mmcif().
# Rank by ipTM for complexes / mean pLDDT for monomers:
best = max(results, key=lambda r: float(r.iptm if r.iptm is not None
                                        else r.plddt.mean()))
open("pred.cif", "w").write(best.complex.to_mmcif())