ESMFold2

Model overview

Synthyra/ESMFold2 packages the biohub/ESMFold2 checkpoint with the FastPLMs runtime for Hugging Face Transformers. It accepts raw amino-acid sequences or typed molecular-complex specifications; low-level forward accepts prepared feature tensors.

The repository uses the standard Transformers loading interface with trust_remote_code=True. See Technical details for each registered class and whether its weights come from the checkpoint.

The sequence- and token-classification classes reuse the pretrained backbone, but their task heads are newly initialized. Fine-tune those heads before interpreting their logits as predictions.

Install and platform requirements

Install the direct dependencies published with this model:

python -m pip install -r \
  "https://huggingface.co/Synthyra/ESMFold2/resolve/main/requirements.txt"

The FastPLMs implementation itself is embedded in the model repository. Transformers loads it through trust_remote_code=True.

This model requires Python 3.11-3.14, PyTorch 2.13, and Transformers 5.13.

The artifact requirements include the structure dependencies.

The release contract requires a CUDA device. The current validated target is the exact NVIDIA GH200 on Linux aarch64. Linux x86-64, CPU-only, Windows, and macOS structure runs are not release evidence.

The Hub quick start needs network access for the first download. For an air-gapped run, build the manifest-pinned local artifact first and use the offline example.

Quick start

from transformers import AutoModel

model_id = "Synthyra/ESMFold2"
model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    attn_implementation="sdpa",
).eval()

For offline validation, replace model_id with the manifest-built dist/hub/ESMFold2 path. Pass local_files_only=True.

Attention backends

The quick start uses sdpa.

Available backends are eager, sdpa, flex_attention. Requesting an unavailable backend raises instead of silently changing implementation.

output_attentions=True can use the documented one-call eager fallback to materialize attention tensors. The configured backend does not change.

Downstream prediction

The sequence and token prediction AutoClasses use the checkpoint backbone and create a new, untrained classifier. Sequence labels have shape (b,). Residue labels have shape (b, l) and use -100 outside biological positions. The folding trunk is skipped. The classifier uses the checkpoint's learned pLM state mixture and projection, followed by one trainable transformer probe.

import torch
from transformers import (
    AutoModelForSequenceClassification,
    AutoModelForTokenClassification,
)

model_id = "Synthyra/ESMFold2"
sequence_model = AutoModelForSequenceClassification.from_pretrained(
    model_id, num_labels=2, trust_remote_code=True
).eval()
token_model = AutoModelForTokenClassification.from_pretrained(
    model_id, num_labels=3, trust_remote_code=True
).eval()
sequences = ["MSTNPKPQRKTKRNT", "MKTIIALSYIFCLVFA"]
batch = sequence_model.prepare_classifier_inputs(sequences)
biological = batch["attention_mask"].bool()

sequence_labels = torch.zeros(len(sequences), dtype=torch.long)
token_labels = torch.full_like(batch["input_ids"], -100)
token_labels[biological] = 0

with torch.inference_mode():
    sequence_output = sequence_model(**batch, labels=sequence_labels)
    token_output = token_model(**batch, labels=token_labels)
print(sequence_output.logits.shape)  # (b, 2)
print(token_output.logits.shape)     # (b, l, 3)

PEFT fine-tuning

Install the training dependencies. Then attach LoRA to the loaded checkpoint:

python -m pip install "datasets>=4.8,<5" "peft>=0.19,<0.20"

from peft import LoraConfig, TaskType, get_peft_model

peft_model = get_peft_model(
    sequence_model,
    LoraConfig(
        task_type=TaskType.SEQ_CLS,
        r=8,
        lora_alpha=16,
        target_modules="all-linear",
        modules_to_save=["classifier"],
    ),
)

This checkpoint advertises a classification head. Save the separately trained classifier with the adapter. All FastPLMs checkpoints follow the Transformers PreTrainedModel contract and can use PEFT. The ESM2-specific shipped CLI is an example, not a support boundary. Record the target modules, base revision, data identity, and trainable parameter scope.

Alignment-conditioning contract

This is a full 48-block ESMFold2 checkpoint. It supports single-sequence inference and optional MSA-conditioned inference. Typed multichain and multimolecule inputs can attach an MSA to each applicable protein chain.

Protein folding

The single-protein helper returns typed structure and confidence outputs:

result = model.fold_protein(
    "MSTNPKPQRKTKRNT",
    num_loops=1,
    num_sampling_steps=200,
    num_diffusion_samples=1,
    seed=7,
)
pdb_text = model.result_to_pdb(result)
cif_text = model.result_to_cif(result)
print(result.ptm, result.plddt.mean().item())

No target structure is required. For complexes, construct the input from the types exposed by the loaded artifact:

types = model.input_types
complex_input = types.StructurePredictionInput(
    sequences=[
        types.ProteinInput(id="A", sequence="MSTNPKPQRKTKRNT"),
        types.ProteinInput(id="B", sequence="MKTIIALSYIFCLVFA"),
        types.DNAInput(id="C", sequence="ATGC"),
        types.LigandInput(id="L", smiles="O"),
    ]
)
complex_result = model.fold(
    complex_input,
    num_loops=1,
    num_sampling_steps=200,
    seed=7,
)
print(complex_result.ptm, complex_result.plddt.mean().item())

The typed interface also supports RNA, protein MSAs, modifications, and covalent bonds. The public schema recognizes PocketConditioning and DistogramConditioning, but the pinned official forward consumes neither. Its feature builder hard-codes a zero pocket feature and constructs distogram tensors that the released model ignores. FastPLMs therefore rejects non-null pocket and distogram conditioning instead of silently ignoring scientific inputs. Prepared ref_pos values are component reference geometries created during featurization, not target coordinates. Predicted coordinates and confidence scores are outputs and do not establish biochemical activity.

Learned representation and ESMC precision

ESMFold2 applies its learned state mixture and projection as H: (b, l, 81, 2560) -> Z: (b, l, 256). Retrieve Z through the public embedding API:

representations = model.embed_dataset(
    ["MSTNPKPQRKTKRNT", "MKTIIALSYIFCLVFA"],
    batch_size=2,
    full_embeddings=True,
)
print(representations[0].tensor.shape)  # (sequence_length, 256)

model.embed_dataset(..., full_embeddings=True) returns one (l, 256) residue tensor per single-chain input. It rejects complexes, ligands, MSAs, chain-separated inputs, cls, and parti in the embedding path.

Set esmc_precision to auto, bf16, fp32, or fp8 when loading. auto always resolves to BF16. Explicit FP8 is experimental, inference-only, and strict:

model.reload_esmc(precision="fp8", device="cuda:0")
print(model.esmc_precision_status)

FP8 raises when the validated CUDA and Transformer Engine path is unavailable. Canonical BF16 weights are retained, and transient quantization state is never serialized.

The ESMC backbone uses SDPA as the recommended highest-fidelity path. Flex Attention is supported and non-experimental but can be numerically divergent; ESMFold2 does not advertise FlashAttention for the folding interface.

Backend	Support	Measurement status
sdpa	Recommended fidelity path	Pending release measurement
eager	Supported	Pending release measurement
flex_attention	Supported, numerically divergent	Pending release measurement

Detailed backend measurements, release guardrails, and the GH200 package compatibility exception are maintained in the attention backend guide and release evidence manifest.

Verified CCD runtime asset

Structure preparation requires ccd.pkl from biohub/ESMFold2. The manifest pins its repository, revision, size, content identity, and MIT terms. This is a trusted-deserialization boundary. FastPLMs accepts only the pinned snapshot link inside the repository blob directory. User-supplied asset and cache_dir symlinks are rejected. The loader verifies a private temporary snapshot before deserialization, protecting against path-replacement and in-place source-write races. Offline execution requires the exact cached object and never downloads a replacement.

Optional folding TTT

The standard and Fast checkpoints expose opt-in folding TTT on their ESMC backbone:

adapted = model.fold_protein_ttt(
    "MSTNPKPQRKTKRNT",
    num_loops=1,
    num_sampling_steps=50,
    seed=7,
    ttt_config={"steps": 3, "batch_size": 1, "seed": 7},
)
print(adapted.ttt_metrics)

Entering a gradient-enabled path reloads canonical BF16 ESMC weights. TTT adds latency and memory and can worsen a prediction. It does not calibrate confidence or show biological validity. Folding TTT is result-scoped. Its transient ESMC adapter modules are excluded from checkpoint state. It is not a generic save_pretrained adapter-persistence path.

Technical details

• Inputs: Raw amino-acid sequences or typed molecular-complex specifications; low-level forward accepts prepared feature tensors
• Transformers classes: AutoConfig, AutoModel, AutoModelForSequenceClassification, AutoModelForTokenClassification
• Checkpoint weights: AutoConfig = FastPLMs extension, AutoModel = pretrained, AutoModelForSequenceClassification = base weights + untrained task head, AutoModelForTokenClassification = base weights + untrained task head
• Attention backends: eager, sdpa, flex_attention
• Precision: auto, fp32, bf16, fp8 (experimental)
• BF16 execution: fp32_parameters_autocast
• Generation contract: not_applicable
• Dependencies: core + structure
• Weight publication allowed: true
• Weight license status: resolved
• Redistributable: true
• Complete weight publication required: false

Validation and provenance

FastPLMs pins the checkpoint, upstream source revisions, state transformation, and required files in models.toml. Built artifacts record exact source identities and conversion details in source-record.json.

• FastPLMs checkpoint: Synthyra/ESMFold2
• Runtime revision: recorded separately in the built artifact and published commit
• Runtime source identities: recorded in source-record.json
• Official checkpoint: biohub/ESMFold2
• Artifact source: fast
• State transform: identity
• Pinned upstreams: biohub-esm, biohub-transformers, protein-ttt
• Release tiers: check, compliance, structure, feature, artifact, benchmark
• Unresolved required file identities: 0

Release validation includes the compliance tier. Its evidence identifies the checkpoint, backend, dtype, hardware, inputs, and reference revision.

Declared tiers compare configuration, tokenizer behavior, state, and representative inference with the pinned reference. A nonzero unresolved count blocks release. Metadata alone does not show that a build passed, that a backend is faster, or that an output is biologically valid.

License

Checkpoint terms: MIT. The Hub model-card identifier is mit. The local artifact contains applicable source licenses, notices, attribution, and conversion records. Review them before use.