Proto is not affiliated with Biohub. This toolkit is open source and builds on the implementation produced by this organization. Product names, logos, and trademarks are the property of their respective owners.

```bibtex theme={null} @misc{candido2026language, title={Language Modeling Materializes a World Model of Protein Biology}, author={Candido, Salvatore and Hayes, Thomas and Derry, Alexander and Rao, Roshan and Lin, Zeming and Verkuil, Robert and Wu, Bryan and Lee, Jin Sub and Bruguera, Elise S. and Keval, Jehan A. and Kopylov, Mykhailo and Pak, John E. and Wu, Wesley and Thomas, Neil and Mataraso, Samson and Hsu, Alvin and Trotman-Grant, Ashton C. and Fatras, Kilian and dos Santos Costa, Allan and Badkundri, Rohil and Ak{\i}n, Halil and Oktay, Deniz and Deaton, Jonathan and Montabana, Elizabeth and Sitwala, Hrishita and Yu, Yue and Wiggert, Marius and Carlin, Dylan Alexander and Goering, Anthony W. and Blazejewski, Tomasz and Sandora, McCullen and Hla, Michael and Jia, Tina Z. and Kloker, Leon H. and Sofroniew, Nicholas J. and Uehara, Masatoshi and Pannu, Jassi and Bachas, Sharrol and Liu, Daniel S. and Sercu, Tom and Rives, Alexander}, year={2026}, url={https://www.biorxiv.org/content/10.64898/2026.06.03.729735}, note={Preprint} } @software{evolutionaryscale_2024, author={{EvolutionaryScale Team}}, title={evolutionaryscale/esm}, year={2024}, publisher={Zenodo}, doi={10.5281/zenodo.14219303}, url={https://doi.org/10.5281/zenodo.14219303} } ```

Copy citation

### ESM C Embeddings (`esmc-embedding`) Runs each input sequence through ESM C once and averages the per-residue representations, excluding the start and end tokens and any padding, into a single fixed-length vector per sequence. Per-position scores (logits) over the 20 standard amino acids are also returned when requested. #### API Reference

Source Protein sequence(s) to process. Can be provided as:

Source ESM C weights variant. `"esmc_300m"` (960-dim embeddings), `"esmc_600m"` (1152-dim), and `"esmc_6b"` (2560-dim). Larger checkpoints give richer representations at the cost of GPU memory; `"esmc_6b"` holds \~13 GB of bf16 weights. Available options: `esmc_300m`, `esmc_600m`, `esmc_6b` Include per-position logits in the output (large; disable to save memory). Transformer layer index for embeddings. `-1` returns the post-norm final-layer output (`outputs.embeddings`); other indices select from pre-norm per-block `outputs.hidden_states`. Range is checkpoint- dependent (esmc\_300m: 30 layers, esmc\_600m: 36 layers, esmc\_6b: 80 layers). Verbosity level (0=quiet, 1=info, 2=debug, 3=raw subprocess stderr). Device to run the model on. Maximum execution time in seconds. `None` waits indefinitely. Random seed. When set, tools run reproducibly up to small GPU float noise (see `BaseToolOutput.approx_equal`), and the seed participates in cache keys. When None, cacheable seed-sensitive tools skip cache until seeded. Number of sequences to process in parallel. Larger batches improve throughput but require more GPU memory.

Source Per-sequence embedding results. Each `SequenceEmbedding` contains: Mean-pooled embedding vector for one sequence. Binary mask indicating valid positions (1) vs padding (0). Optional per-position amino acid logits for one sequence. Optional 2D coordinate from a UMAP projection of all embeddings in the same call. Populated when `n_sequences >= 4`; `None` otherwise (single-point or 2-3-point UMAP is meaningless).

#### Applications The averaged embedding is a learned numerical representation of a protein, suitable for machine-learning tasks such as clustering, classification, and property prediction, and for similarity search by comparing these vectors (for example with cosine similarity). The optional per-position scores give the model's predicted amino-acid preference at each site, useful for conservation analysis and for examining the model's expectations at specific positions. ESM C is embedding-focused, so it is the lighter-weight choice when you need embeddings or per-position scores but not sequence generation or scoring. #### Usage Tips * **`model_checkpoint` selects the model size.** `esmc_300m` (the default) has embedding size 960, `esmc_600m` has 1152, and `esmc_6b` has 2560. Larger checkpoints give richer representations but cost more GPU memory and time — `esmc_6b` loads about 13 GB of bf16 weights, before activations that grow with sequence length and `batch_size`. * **`repr_layer` selects which internal model layer the embedding is taken from.** The default `-1` uses the final layer; other values select earlier layers. * **Per-position scores are large.** Enabling `return_logits` adds an array of size (sequence length by 20) per sequence, which dominates runtime and memory for long inputs. Leave it set to `False` unless you need the per-position scores.