license: pddl
tags:
- biology
- bioinformatics
- llm
- embeddings
- protein
pretty_name: ESM2-15B Human/Mouse embedding
ESM2-15B Human and Mouse protein embeddings
This dataset contains protein embeddings obtained through the ESM2-15B model for the Human and Mouse species.
The model used can be found here: https://huggingface.co/facebook/esm2_t48_15B_UR50D
Input sequences
Protein sequences were obtained from Swiss-Prot/Uniprot, meaning they were curated beforehand. The sequences were obtained from the following link in the month of May, 2025.
Only proteins with at least one GO annotation were kept. An initial total of 552512 entries were initially obtained. Embeddings were generated only for Human and Mouse species.
Output embeddings
The ESM2-15B model generated embeddings with 5120 dimensions. A total of 36132 unique protein embeddings were generated belonging to both species.
Downstream analysis
In order to be able to do relevant analysis with the embeddings, the corresponding relevant metadata was kept, including the sequence accessions, their names, source organism and GO annotations.
What does this dataset contain?
- Protein embeddings (ESM2_15B_Human_Mouse_Embeddings.npy): in .npy format, these are the 5120-dimensional embeddings for each protein belonging to Human or Mouse.
- Parsed metadata (ESM2_15B_Human_Mouse_Metadata.csv.gz): in .csv fomat, it contains the relevant metadata directly associated with the embeddings specified above.
- Entry: this column contains the name of the protein.
- Accession: this column contains the accession number.
- Organism: this column contains the source organism. At this moment, it can be either Human or Mouse.
- GO annotations: this column contains a semicolon-separated string of GO annotations associated with that sequence.
- Raw metadata (Original_Metadata.tsv.gz): in .tsv format, this file contains extended metadata directly obtained from Swiss-Prot/Uniprot.
- Source script (ESM2_script.py): in .py format, this script generated the embeddings directly from the source data.
- Classification analysis (ESM2_Classification.ipynb): Jupyter Notebook that parses the embeddings and the metadata in order to perform a classification task on the 20 most common GO annotations for the Human sequences. A Random Forest classifier or a LightGBM classifier is fitted for every annotation and a prediction is made on a test split, obtaining relevant metrics. It also contains PCA and UMAP analysis of the embeddings.
What is the purpose of this dataset?
The main purpose is to facilitate downstream tasks that involve protein embeddings and their analysis together with their associated metadata by the scientific community. The included embeddings are highly dimensional and obtained through one of the latest protein embedding models.
It should also be possible to add external metadata to each embedding for the purpose of increasing the information within models.
Why Human and Mouse?
Since they are the most studied species, embeddings belonging to their protein sequences were generated first. Also, generating them was extremely computationally intensive, taking up to two weeks.
The possibility of generating embeddings for other species might be explored in the future. Available species are named in the included Python script.
Usage and citation
You may save, use and redistribute this dataset freely as you see fit. If you feel like citing this repository, it will be appreciated.
Requests
If you see any errors or corrections, by all means point them out so they can be fixed.
Final notes
The research field is quickly approaching a chapter where reproducibility will not be guaranteed. I am sure you already feel it. A crisis is fast approaching, and very few people look beyond their own ego and focus only on personal gains.
This is the reason why the field has become extremely tainted. Built on the backs of overworked students, generating results is more important than fact-checking them or giving the experimental design a little more thought.
When growing your own credentials is your ultimate goal, this is the result.
Science should be open and free.
Sources
- ESM2 model: https://huggingface.co/facebook/esm2_t48_15B_UR50D
- Input sequences: https://ftp.uniprot.org/pub/databases/uniprot/current_release/knowledgebase/complete/uniprot_sprot.dat.gz (as found by May, 2025)