In brief: ESMC turns protein sequences into learned representations. These embeddings can help downstream models learn useful relationships, but the output is not by itself a statement of protein function.
What is a protein embedding?
A protein embedding is a numeric vector derived from a sequence by a pretrained model. Different model layers may encode different properties. Researchers can use these representations for clustering, retrieval or supervised prediction if the associated task has suitable labels.
Typical workflow
- Collect sequences with clear provenance and permissions.
- Normalize input formats and track duplicates.
- Use the currently documented ESMC tokenizer and model.
- Specify layer selection and aggregation method.
- Train a downstream baseline and assess out-of-distribution behavior.
Avoid accidental leakage
Highly homologous proteins may appear in both development and testing. A sequence-family split may better estimate whether an approach generalizes. Compare embeddings against sequence identity and simpler baselines rather than attributing all performance to a sophisticated latent representation.
What has changed since 2024?
Biohub now documents updated ESMC variants and an ESM Atlas built around representations from this family. Use current model cards because package APIs and available checkpoints have evolved since the first release.
Primary sources and further reading
Editorial review: October 2026. This guide is educational. Model availability and APIs may change; confirm exact versions in the linked documentation. Generated structures and sequences require independent scientific validation.