ESM3 generation: esm 3.4.1.post1

Model and interface

Local ESM3.from_pretrained("esm3-sm-open-v1", device=torch.device("cpu")) loads the open 1.4B checkpoint. Hosted esm.sdk.client() uses Biohub, with esm3-small-2024-08, esm3-medium-2024-08 (7B), and esm3-large-2024-03 (98B) listed in the current ESM3 README. Access depends on the account; model listing is not an authenticated availability test. Do not rely on the old unverified esm3-medium-multimer-2024-09 example.

ESMProtein contains optional sequence, secondary-structure, SASA, function and coordinate tracks. Sequence masking uses _. Use consistent track lengths and explicit chain/residue mappings; plain single-chain recipes here exclude | chain breaks. Coordinates are torch.Tensor atom37 arrays (L,37,3) in angstroms, with missing atoms represented by NaN, not Python None entries in a tensor.

ESMProtein.from_pdb(path, chain_id="A") selects an author chain. The current default chain_id="all" combines chains, so specify the desired chain explicitly. to_pdb(path) writes a file; to_pdb_string() returns text. Preserve original residue numbers/insertion codes externally when sequence indices differ.

Generation controls

GenerationConfig supports tracks sequence, structure, secondary_structure, sasa, and function. Relevant released defaults are num_steps=20, temperature=1.0, temperature_annealing=True, top_p=1.0, strategy="random", schedule="cosine", and condition_on_coordinates_only=True. Source comments about earlier defaults can differ from the actual field values; record explicit settings.

• Use a positive step count suited to the missing positions. Local generation caps excessive steps at the number of masks; the hosted client also caps by sequence length and can mutate the config. Create a separate config per request.
• top_p is a float; 1.0 disables nucleus restriction. Do not pass None.
• Temperature has no maximum of 1.0. At zero, sampling can select argmax, but default annealing has a positive floor and random unmasking remains stochastic. Temperature zero alone is not a reproducibility guarantee.
• condition_on_coordinates_only selects coordinates instead of discrete structure tokens as structure conditioning. It does not clear the sequence or every other track. For inverse folding, remove the sequence explicitly.
• .generate() fills masked positions. It does not resample all fixed residues; a temperature schedule made from repeated calls on completed sequences is not refinement. Remask intended sites, or use the SDK's actual annealing option.

The unified low-level inference path is encoded = model.encode(protein) then output = model.logits(encoded, LogitsConfig(...)). Inspect output.logits.sequence for sequence logits. model.forward(encoded) followed by model.decode(output) is invalid: forward has lower-level batched tensor arguments and decode expects an ESMProteinTensor, not a logits object.

Inverse folding and independent checking

Illustrative pretrained inference; model is a loaded ESM3 inference client.

from esm.sdk.api import ESMProtein, ESMProteinError, GenerationConfig

original = ESMProtein.from_pdb("target.pdb", chain_id="A")
assert original.coordinates is not None and original.sequence
prompt = ESMProtein(coordinates=original.coordinates.clone())
design = model.generate(prompt, GenerationConfig(
    track="sequence", num_steps=min(20, len(original.sequence)),
    temperature=0.7, condition_on_coordinates_only=True,
))
if isinstance(design, ESMProteinError):
    raise design
assert design.sequence is not None and len(design.sequence) == len(original.sequence)

# Sequence-only input avoids conditioning the structural check on the target.
predicted = model.generate(ESMProtein(sequence=design.sequence), GenerationConfig(
    track="structure", num_steps=min(20, len(design.sequence)),
))
if isinstance(predicted, ESMProteinError):
    raise predicted
predicted.to_pdb("independently_predicted.pdb")

Compare matched, finite atoms after structural superposition; report alignment coverage and confidence, not just an RMSD number. Successful coordinate generation is not evidence that a design is stable or functional. A same-model round trip also is not independent experimental validation.

Function conditioning

FunctionAnnotation is exported from esm.sdk.api and defined in esm.utils.types. Its start and end are 1-based, inclusive. Labels must occur in the selected function/residue-annotation tokenizer vocabulary: supported InterPro IDs or exact keyword entries, not invented natural-language underscored labels. A range specifies where a function annotation applies; it is not a constraint fixing a catalytic residue or GFP chromophore.

For a local loaded ESM3 model, inspect the supported vocabulary before constructing annotations (illustrative; vocabulary assets were not downloaded in this review):

from esm.sdk.api import ESMProtein, FunctionAnnotation, GenerationConfig

label = selected_supported_label  # Chosen from this model's vocabulary.
tokenizer = model.tokenizers.function
assert label in tokenizer.interpro_labels or label in tokenizer.keyword_vocabulary
annotation = FunctionAnnotation(label=label, start=1, end=150)
prompt = ESMProtein(sequence="_" * 150, function_annotations=[annotation])
result = model.generate(prompt, GenerationConfig(track="sequence", num_steps=20))

Function-track predictions are annotation hypotheses. Low temperature is not calibrated confidence. Confirm with curated evidence and task-specific assays; function conditioning alone cannot guarantee fluorescence or enzyme activity.

Multitrack designs

A useful sequence is: complete masked sequence -> predict structure -> optionally remask selected sequence positions while retaining chosen structural conditions -> create a fresh sequence-only prompt for structural checking. Explicitly clear coordinates before predicting a new structure on an existing protein object. Keep the initial prompt, each generated candidate and the exact mutations. Avoid splitting long proteins into arbitrary independent chunks and claiming a coherent full-length fold.

Sources: ESM3 README, SDK types, generation, annotation coordinates, function tokenizer.