Case study · Life sciences
The human proteome, read by a protein language model
Which proteins does a protein language model read as alike, and what do they share?
The Space
The human proteome, read by a protein language model
All 20,431 reviewed human proteins (UniProtKB/Swiss-Prot) placed by ESM-2 embeddings, so proteins the model reads as alike sit together, with families, compartments, structures and disease links from UniProt.
Source: UniProtKB/Swiss-Prot (CC BY 4.0) with ESM-2 embeddings (MIT); The UniProt Consortium, NAR 2025; Lin et al., Science 379, 1123 (2023).
20,431 records
Largest clusters in the map sample
- Mixed functions24%
- Cytoplasmic proteins18%
- Nuclear proteins13%
- Enzymes: hydrolases9%
- Secreted proteins8%
- Membrane proteins7%
- Transcription factors7%
- Zinc-finger transcription factors5%
On the map
What to look for
- Proteins sit by what ESM-2 reads in their sequences, with their UniProt families and compartments alongside.
- Annotations from UniProt (structures, compartments, disease links) colour the map without changing where proteins sit.
With Mantis
What to ask
- Which protein families sit closest to the kinases?
- Where do disease-linked proteins concentrate?
- Which clusters mix compartments, and why?
More case studies
- The immune response to severe COVID-19, cell by cellWhich immune cells carry the interferon response in severe COVID-19, and how does it differ from severe influenza?
- Which neurons Alzheimer’s disease takes firstAlong the disease continuum, which cell types are lost earliest in the middle temporal gyrus?
- Kinase inhibitors by structure and targetWhere do approved kinase drugs sit among the chemistry that hits the same families?
Work with us
Bring representation superintelligence to your team.
Start in the workspace today, or talk to us about your data, your deployment and your questions.