Sequence analysis · Property prediction

ESMC

Protein language-model embeddings, amino-acid probabilities, and substitution scores.

ESM Cambrian analyzes protein sequences with local 300M, 600M, or 6B models. Exports per-residue and mean-pooled embeddings, optional hidden states and logits, position-level predictions, and optional masked-marginal single-substitution log-odds. Representations support downstream property prediction; scores are not calibrated measurements of stability, fitness, or function.

Documentation ↗ Input parameters ↗ Examples ↗ Home page ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Input mode

Input

Job name Names the output directory.

Proteins

One independent protein per box (up to 26). Amino-acid letters only; no complexes, modifications, MSAs or templates. Maximum 2046 residues per protein, reserving two positions for start/end tokens.

Protein sequence or FASTA Enter one bare sequence (line wrapping allowed) or up to 26 FASTA records with unique headers. Each protein is analyzed independently. Accepts standard amino acids plus X, B, Z, U and O. Maximum 2046 residues per protein; sequences are never truncated.

Sequence file Upload FASTA or a plain-text protein sequence. The file contents are sent to the compute node.

Inference

Model Current Biohub MIT-licensed checkpoints, downloaded on first use. 300M is the lightest; 6B requires substantially more RAM/VRAM.

Outputs

Save residue and mean embeddings Export FP32 embeddings [residues, dimensions] and mean_embedding [dimensions] in one compressed NPZ per sequence. Excludes start/end tokens and padding.

Save full logits and probabilities Export [residues, 64] raw logits and softmax probabilities in NPZ. summary.json maps vocabulary indices; unused slots are null. Position predictions and CSV summaries are always saved.

Additional hidden-state layer -1 disables additional hidden-state export. Otherwise upstream indices are 0 through 30 (300M), 36 (600M), or 80 (6B). Index 0 is the first block input; the last is the final normalized output. Upstream collects all layers in memory when this is enabled; reduce batch size for long sequences.

Scoring

Single substitutions (optional) Exactly one input protein. Enter up to 100 substitutions such as M1A, separated by spaces, commas or newlines. Positions are one-based and the reference residue must match. Each substitution is scored independently by masking its position and computing log P(mutant) - log P(reference). Positive means model preference, not measured fitness or stability. Saves mutation_scores.csv.

Runtime

Batch size Number of independent proteins per inference batch. Smaller batches reduce memory use; sequences are padded only within a batch.

Device Auto uses CUDA when available, otherwise CPU. An explicitly requested unavailable CUDA device fails clearly.

Precision Model weight dtype. BF16 reduces memory on supported hardware. All exported floating arrays use FP32.

Ready