Protein Embeddings · DSIMB
PATHOS-PLM-EMBEDDINGS
Precomputed protein language model embeddings for wild-type and mutant human SwissProt proteins.
What PATHOS-PLM-EMBEDDINGS is
PATHOS-PLM-EMBEDDINGS provides precomputed protein language model representations across the human proteome, covering 20,416 human SwissProt proteins. Developed by researchers at DSIMB, the resource captures wild-type residues alongside all possible single amino acid missense substitutions to support pathogenicity prediction in tools such as PATHOS.
Deriving high-dimensional contextual embeddings across millions of variants requires substantial compute.
How it was made
The dataset derives embeddings from three protein language models: ESM-C 600M (1,152 dimensions), ESM-2 650M (1,280 dimensions), and Ankh2 Large (1,536 dimensions). For each human SwissProt protein, all single amino acid substitutions are generated and evaluated alongside the original wild-type sequence. Two embedding vectors are retained in float32 precision for every entry: a position-specific embedding at the residue or mutation site, and a mean-pooled sequence embedding.
For sequences exceeding 1,024 residues, the input to the protein language model is trimmed to a 1,024-residue window centred on the variant position. For these extended sequences, the position embedding corresponds to the mutated residue within the window, and the mean embedding is pooled across the 1,024-residue slice. Files are organised in Parquet format, indexed by an auxiliary shard manifest recording the inclusive UniProt accession boundaries for each shard.
Key facts
- Proteins covered
- 20,416 human SwissProt proteins
- Total rows
- 682,732,968
- Mutation rows per model
- 216,198,775
- Wild-type rows per model
- 11,378,881
- Supported PLMs
- ESM-C 600M, ESM-2 650M, Ankh2 Large
- Embedding dimensions
- 1,152 (ESM-C), 1,280 (ESM-2), 1,536 (Ankh2)
- Storage format
- Parquet
- Numeric precision
- float32
- Total Parquet size
- 6.50 TiB
What you can use it for
- Pathogenicity classification
- Supplying feature vectors to machine learning architectures predicting the deleterious effects of human missense mutations.
- Variant impact quantification
- Comparing mutant embeddings to wild-type embeddings using vector differences or cosine similarities to assess local structural or functional shifts.
- Multi-PLM ensemble modelling
- Concatenating representations from ESM-C, ESM-2, and Ankh2 Large to capture complementary sequence patterns for downstream predictors.
- Macromolecular ML training
- Training sequence-based classification or regression models via streaming data loaders without the computational overhead of running PLM inference.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from DSIMB/PATHOS-PLM-EMBEDDINGS at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 587,100 rows of esmc_600m/train
mutant
string_label- H5%
- W5%
- C5%
position
int- min
- 1
- median
- 78
- mean
- 130.98
- max
- 993
protein_id
string_label- A0A0J9YWL93%
- A0A0J9YY542%
- A0A0J9YXQ42%
variation
string_text- min
- 3
- median
- 4
- mean
- 4.30
- max
- 5
wild_type
string_label- S10%
- L10%
- G8%
Strengths and limits
Good at
- Comprehensive coverage of all possible single amino acid substitutions across 20,416 human SwissProt proteins.
- Precomputed representations from three distinct protein language models at full float32 precision.
- Includes both residue-specific embeddings at the variant site and sequence-wide mean-pooled embeddings.
- Sorted shard structure with an index manifest enabling targeted single-protein lookups without downloading entire configurations.
Watch out for
- Restricted to single amino acid missense substitutions and wild-type sequences, excluding insertions, deletions, and non-human species.
- Sequences longer than 1,024 amino acids use a 1,024-residue centred window rather than the complete protein context.
- The 6.50 TiB total volume requires targeted shard querying or streaming workflows to avoid severe local storage demands.
Questions
Can I look inside PATHOS-PLM-EMBEDDINGS without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is PATHOS-PLM-EMBEDDINGS under?
Its authors declare mit. Read the terms on the original page before reusing the data.
Which protein language models were used to generate the embeddings?
Representations were generated using ESM-C 600M (1,152-dimensional), ESM-2 650M (esm2_t33_650M_UR50D, 1,280-dimensional), and Ankh2 Large (1,536-dimensional).
How are proteins longer than 1,024 residues handled?
For proteins exceeding 1,024 residues, a 1,024-residue window centred on the variant site is passed to the language model, with the mean-pooled vector computed across that window.
What columns are included in the mutation files?
Each mutation record contains the UniProt accession, residue position, wild-type amino acid, substituted amino acid, compact mutation notation, position-specific embedding, and mean-pooled sequence embedding.
How are single proteins queried efficiently?
A metadata manifest file indexes the inclusive range of UniProt identifiers in each shard, allowing queries to identify and download only the specific shards containing the protein of interest.
How to cite it
Radjasandirane, R., Cretin, G., Diharce, J., de Brevern, A. G., & Gelly, J.-C. (2026). PATHOS: Predicting variant pathogenicity by combining protein language models and biological features. Artificial Intelligence in the Life Sciences, 9, 100165. http://dx.doi.org/10.1016/j.ailsci.2026.100165
Written from PATHOS-PLM-EMBEDDINGS's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.