BoevaLab/Gene-Embedding-Hub-dataset

Unified Human Gene Embeddings

A standardized, multimodal collection of 68 gene-level representation matrices aligned to a common universe of 20,097 human protein-coding genes.

The resource brings together representations derived from protein, DNA, RNA and mature mRNA sequences; bulk and single-cell expression; perturbational and functional-genomics data; biomedical literature; protein-protein interaction networks; ontologies; pathways; knowledge graphs; regulatory prediction models and single-cell foundation-model checkpoints.

Useful scripts and documentation can be found in: https://github.com/DanielTabaresLopez/Gene-Embeddings-ETH-SSRF

Summary

Item	Current release
Core embeddings	68
Common master-gene universe	20,097 genes
Total mapped gene rows across all core matrices	1,235,842
Mean coverage of the master universe	90.43%
Coverage range	20.25%–100.00%
Embedding dimensionality range	20–20,421
Matrix dtype	float32
Approximate staged repository size	4.1 GB

Objetive of the project

Gene embeddings are fixed numerical vectors representing gene identity. Different construction methods encode different biological priors: amino-acid sequence, nucleotide sequence, transcript structure, expression, function, literature, interaction networks, pathways, regulatory activity or knowledge learned by foundation models.

These representations are difficult to compare directly because they are distributed across many sources, use different identifier systems, cover different subsets of genes, have different dimensions, and are released in incompatible formats. This project creates a common interface so that the same genes can be compared across otherwise unrelated representation spaces.

The main scientific questions are:

1. Which embeddings and modalities learn similar gene geometries?
2. Which representations contain redundant versus complementary information?
3. Can one modality be predicted from another?
4. Do stronger or more general representations converge toward a shared geometry?
5. Which embeddings are most useful for particular downstream biological tasks?

Planned analyses include centered kernel alignment, nearest-neighbor overlap, linear predictability across modalities, dimensionality reduction, clustering, and downstream benchmarking. The present release is the standardized resource required for those analyses; it does not claim that any one embedding is universally best.

The work included:

1. Building a common gene-identifier backbone.
   A 20,097-row enriched human protein-coding master table was assembled and cross-checked across Ensembl, HGNC, Entrez, UniProt, RefSeq, CCDS, STRING, transcript, and protein identifiers.

2. Finding and evaluating embedding sources.
   Original releases were preferred. Official model repositories, Hugging Face, Zenodo, Figshare, and benchmark archives were used when appropriate. Each source was recorded with its method, modality, identifier type, checkpoint or release, file format, generation route, and known caveats.

3. Downloading, extracting, or generating vectors.
   Depending on the method, vectors were:

   • taken from official precomputed releases;
   • extracted from learned gene-token or position-embedding tables in checkpoints;
   • generated locally by running sequence models on canonical human CDS, mature transcripts, proteins, or TSS-centered genomic windows;
   • reconstructed from public network, ontology, expression, or knowledge-graph data;
   • derived from summarized model output tracks, as in the AlphaGenome representation.

4. Mapping every source to the same gene universe.
   Source identifiers were resolved to one final Ensembl gene ID using explicit, source-specific rules. Ambiguous mappings were excluded rather than guessed.

5. Standardizing output files.
   Every retained embedding was converted to a float32 matrix and packaged with row metadata and machine-readable provenance.

6. Running final quality control.
   All core outputs were checked for shape consistency, finite values, row-order agreement, duplicate Ensembl IDs, zero vectors, and expected dimensionality.

7. Documenting reproducibility and caveats.
   Every embedding folder contains a dedicated README with source links, mapping details, processing decisions, final shape, coverage, and interpretation notes.

Repository layout

.
├── README.md
├── data/
│   ├── master_gene_table.csv
│   └── embeddings/
│       └── <embedding_id>/
│           ├── embeddings.npz
│           ├── genes.tsv
│           ├── metadata.json
│           └── README.md
└── scripts/

Root-level data files

• data/master_gene_table.csv — the standardized human protein-coding gene universe and identifier backbone.

Files in every embedding folder

• embeddings.npz — compressed NumPy archive. The matrix is stored under the key embeddings.
• genes.tsv — authoritative row metadata. Its row order matches the matrix exactly.
• metadata.json — machine-readable provenance, source, processing, mapping, dimensionality, coverage, and QC information.
• README.md — human-readable documentation for that specific embedding.

genes.tsv schemas may contain source-specific columns, but every core embedding includes an ensembl_gene_id column. Never assume that an embedding follows the full master-table row order: always use the corresponding genes.tsv.

Quick start

The examples below use the organization repository:

REPO_ID = "BoevaLab/Gene-Embedding-Hub-dataset"

Authentication is required while the dataset remains private. Run hf auth login once on the machine where the files will be accessed.

Download and load one embedding

from huggingface_hub import hf_hub_download
import numpy as np
import pandas as pd

repo_id = "BoevaLab/Gene-Embedding-Hub-dataset"
embedding_id = "esm2"
folder = f"data/embeddings/{embedding_id}"

matrix_path = hf_hub_download(
    repo_id=repo_id,
    filename=f"{folder}/embeddings.npz",
    repo_type="dataset",
)
genes_path = hf_hub_download(
    repo_id=repo_id,
    filename=f"{folder}/genes.tsv",
    repo_type="dataset",
)

with np.load(matrix_path, allow_pickle=False) as archive:
    X = archive["embeddings"]

genes = pd.read_csv(genes_path, sep="\t", dtype=str)

assert X.dtype == np.float32
assert X.shape[0] == len(genes)
assert genes["ensembl_gene_id"].is_unique

print("Matrix:", X.shape)
print(genes.head())

Align two embeddings by Ensembl gene ID

import numpy as np
import pandas as pd

def align_embeddings(X_a, genes_a, X_b, genes_b):
    row_a = pd.Series(
        np.arange(len(genes_a)),
        index=genes_a["ensembl_gene_id"],
    )
    row_b = pd.Series(
        np.arange(len(genes_b)),
        index=genes_b["ensembl_gene_id"],
    )

    common = row_a.index.intersection(row_b.index, sort=False)

    X_a_aligned = X_a[row_a.loc[common].to_numpy()]
    X_b_aligned = X_b[row_b.loc[common].to_numpy()]

    return common.to_numpy(), X_a_aligned, X_b_aligned

This alignment step is essential before computing cross-embedding similarity, regression, nearest neighbors, or concatenated multimodal features.

Standardization and mapping pipeline

1. Common master-gene universe

The master table contains one row per human protein-coding Ensembl gene and preserves a stable row order. It combines current and historical identifier evidence so that older or source-specific identifiers can be resolved conservatively.

Important identifier families include:

• Ensembl gene, transcript, and protein IDs;
• HGNC approved, previous, and alias symbols;
• Entrez Gene IDs;
• UniProt accessions;
• RefSeq and CCDS identifiers;
• STRING protein identifiers;
• canonical transcript/protein and GRCh38 genomic-coordinate information.

Embedding processing never adds new genes to this universe.

2. Identifier mapping rules

The general policy is unique mapping or exclusion:

• current or HGNC-approved symbols have priority over previous symbols and aliases;
• Entrez, UniProt, Ensembl protein, transcript, and STRING identifiers are accepted only when they resolve uniquely;
• ambiguous source identifiers are excluded and reported;
• unmapped identifiers remain explicit in mapping reports;
• duplicate source rows are averaged only when they safely resolve to the same final master-table gene;
• GenePT uses a stricter direct-first rescue policy so aliases do not overwrite or dilute a direct current-symbol vector;
• sequence-derived embeddings use one documented canonical sequence or genomic window per retained gene.

Coverage therefore reflects the biology and source vocabulary of each embedding, not artificial imputation.

3. Source-specific extraction or generation

The collection contains several representation types:

• Released gene vectors standardized without changing their biological values.
• Checkpoint embedding tables extracted as one fixed vector per gene token or position.
• Sequence-model outputs generated from canonical protein, CDS, mature transcript, or TSS-centered DNA inputs.
• Feature and prediction vectors derived from expression, CRISPR dependency, localization, perturbation, splicing, or regulatory-track outputs.
• Network and knowledge-graph representations regenerated or transformed from public graphs and annotations.

4. Final output contract

Every core embedding satisfies:

embeddings.npz["embeddings"].shape == (number_of_rows_in_genes.tsv, embedding_dim)
embeddings.npz["embeddings"].dtype == float32
genes.tsv["ensembl_gene_id"] is unique

5. Quality-control criteria

A core output is retained only after checking:

• matrix rows equal genes.tsv rows;
• expected embedding dimension;
• no NaN values;
• no infinite values;
• no duplicated final Ensembl gene IDs;
• no unexpected zero vectors;
• deterministic or documented aggregation;
• recorded coverage and missing-gene information;
• source and processing provenance in metadata.json.

Embedding inventory

Each name below links to the embedding-specific README. Coverage is relative to the 20,097-gene master universe.

Biomedical literature and text

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
genept_ada	GenePT-ADA-v1.1	biomedical literature / text	19,099	1,536	95.03%
genept_model3_gene_protein	GenePT-Model3-GeneProtein-v1.1	biomedical literature + protein text	19,103	3,072	95.05%
bioconceptvec_cbow	BioConceptVec-CBOW-v1.1	biomedical literature / text	17,040	100	84.79%
bioconceptvec_skipgram	BioConceptVec-SkipGram-v1.1	biomedical literature / text	17,040	100	84.79%
bioconceptvec_glove	BioConceptVec-GloVe-v1.1	biomedical literature / text	17,040	100	84.79%
bioconceptvec_fasttext	BioConceptVec-fastText-v1.1	biomedical literature / text	17,040	100	84.79%
pubmedbert_gene_protein_summary	PubMedBERT-GeneProteinSummary-v1.1	biomedical literature + protein text / BERT summary embedding	19,168	768	95.3774%
biobert_gene_protein_summary	BioBERT-GeneProteinSummary-v1.1	biomedical literature + protein text / BERT summary embedding	19,168	768	95.3774%

Expression, perturbation, localization, and functional genomics

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
gene2vec	Gene2Vec-v1.1	coexpression / expression	17,916	200	89.15%
frogs_archs4_256	FRoGS-ARCHS4-256-v1.1	co-expression / ARCHS4 functional expression embedding	16,265	256	80.93%
gtex_tissue_median_log1p_zscore	GTEx-tissue-median-log1p-zscore-v1.1	bulk RNA expression / normal tissue median expression	19,353	68	96.30%
depmap_crispr_gene_effect_pca256	DepMap-CRISPRGeneEffect-PCA256-v1.1	functional genomics / CRISPR dependency	18,562	256	92.36%
hpa_normal_ihc_tissue_level_zscore	HPA-normal-IHC-tissue-level-zscore-v1.1	protein expression / normal tissue immunohistochemistry	13,420	59	66.78%
hpa_subcellular_location_multihot_zscore	HPA-subcellular-location-multihot-zscore-v1.1	protein subcellular localization	13,581	50	67.58%
cmap_gse106127_cgs_mean_landmark978_zscore	CMap-GSE106127-CGS-mean-landmark978-zscore-v1.1	perturbational transcriptomics / L1000 consensus genetic perturbation signatures	4,070	978	20.25%
mahi_global	Mahi-global-v1.1	tissue/context-aware gene embedding / global context	18,918	3,666	94.13%
mahi_all_contexts_mean	Mahi-all-contexts-mean-v1.1	tissue/context-aware gene embedding / mean across non-global contexts	18,918	3,666	94.13%

Networks, ontologies, pathways, and knowledge graphs

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
mashup_string	Mashup-STRING-v1.1	PPI / network	16,331	800	81.26%
node2vec_consensus_ppi	Node2Vec-consensus-PPI-v1.1	PPI / network	18,356	128	91.34%
deepnf_string_v12_ppmi_svd_fusion	deepNF-STRING-v12-PPMI-SVD-fusion-v1.1	PPI / STRING multimodal network	19,071	256	94.89%
opa2vec_goa_human	OPA2Vec-GOA-human-v1.1	Gene Ontology / ontology annotation semantic embedding	18,660	200	92.85%
newt_go_256	NEWT-GO-256-v1.1	functional annotation / GO component embedding	13,788	256	68.61%
newt_archs4_256	NEWT-ARCHS4-256-v1.1	co-expression / ARCHS4 component embedding	16,265	256	80.93%
newt_go_graph	NEWT-GO-graph-v1.1	GO graph-derived component embedding	13,788	256	68.61%
newt_msigdb_bundle	NEWT-MSigDB-bundle-v1.1	pathway / gene-set component embedding	15,423	256	76.74%
newt_cellnet	NEWT-CellNet-v1.1	transcription-factor network component embedding	17,499	128	87.07%
primekg_nonppi_biomedical_context_tfidf_svd256	PrimeKG-nonPPI-biomedical-context-TFIDF-SVD256-v1.1	heterogeneous biomedical knowledge graph / non-PPI gene context	19,216	256	95.62%
hetionet_nonppi_biomedical_context_tfidf_svd256	Hetionet-nonPPI-biomedical-context-TFIDF-SVD256-v1.1	heterogeneous biomedical knowledge graph / non-PPI gene context	18,787	256	93.48%

DNA, regulatory sequence, and splicing

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
nt_v2_500m_multispecies_cds	Nucleotide-Transformer-v2-500M-multispecies-CDS-v1.1	DNA sequence / canonical CDS nucleotide language model	20,085	1,024	99.94%
hyenadna_medium_160k_cds	HyenaDNA-medium-160k-CDS-v1.1	DNA sequence / canonical CDS nucleotide language model	20,085	256	99.94%
dnabert2_117m_cds	DNABERT-2-117M-CDS-v1.1	DNA sequence / canonical CDS nucleotide language model	20,085	768	99.94%
deepsea_tss_1kb_epigenomic_probs	DeepSEA-TSS-1kb-epigenomic-probabilities-zscore-v1.1	regulatory DNA / TSS epigenomic probability model	20,097	919	100.00%
spliceai_canonical_splice_site_profile	SpliceAI-canonical-splice-site-profile-zscore-v1.1	splicing / canonical splice-site sequence model profile	20,097	606	100.00%
enformer_tss_393kb_human_central3	Enformer-TSS-393kb-human-central3-log1p-zscore-v1.1	regulatory DNA / long-range TSS sequence-to-activity model	20,097	5,313	100.00%
basenji2_tss_131kb_human_central3	Basenji2-TSS-131kb-human-central3-log1p-zscore-v1.1	regulatory DNA / TSS sequence-to-activity model	20,097	5,313	100.00%
alphagenome_tss1mb_regulatory_tracksummary_pca256	AlphaGenome-TSS1Mb-regulatoryTrackSummary-PCA256-v1.1	regulatory DNA / multimodal long-range sequence-to-function output embedding	20,084	256	99.94%

RNA and mature-mRNA sequence

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
helix_mrna_full_transcript	Helix-mRNA-full-transcript-v1.1	mRNA sequence / full mature transcript language model	19,976	256	99.40%
rnafm_full_transcript	RNA-FM-full-transcript-v1.1	RNA/mRNA sequence / full mature transcript language model	19,976	640	99.40%
orthrus_base_4track_full_transcript	Orthrus-base-4-track-full-transcript-v1.1	RNA/mRNA sequence / full mature transcript state-space model	19,976	256	99.40%

Protein sequence and protein representations

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
esm2	ESM2-v1.1	protein sequence	18,690	1,280	93.00%
msa_transformer_depth1_uniprot_human	MSATransformer-depth1-UniProt-human-v1.1	protein sequence / MSA Transformer depth-1	19,472	768	96.8901%
probe_aac	PROBE-AAC-v1.1	protein representation / amino acid composition	18,873	20	93.91%
probe_albert	PROBE-ALBERT-v1.1	protein representation	18,453	4,096	91.82%
probe_apaac	PROBE-APAAC-v1.1	protein representation / amino acid composition	18,871	80	93.90%
probe_bert_bfd	PROBE-BERT-BFD-v1.1	protein representation	18,453	1,024	91.82%
probe_bert_pfam	PROBE-BERT-PFAM-v1.1	protein representation	18,461	768	91.86%
probe_blast	PROBE-BLAST-v1.1	protein representation / sequence similarity feature	18,873	20,421	93.91%
probe_cpc_prot	PROBE-CPC-PROT-v1.1	protein representation	18,453	512	91.82%
probe_esmb1	PROBE-ESMB1-v1.1	protein representation	18,453	1,280	91.82%
probe_gene2vec_uniprot	PROBE-GENE2VEC-UNIPROT-v1.1	protein-associated representation	17,564	200	87.40%
probe_hmmer	PROBE-HMMER-v1.1	protein representation / sequence-profile feature	18,873	20,421	93.91%
probe_ksep	PROBE-KSEP-v1.1	protein representation	18,801	400	93.55%
probe_learned_vec	PROBE-LEARNED-VEC-v1.1	protein representation	18,873	64	93.91%
probe_mut2vec	PROBE-MUT2VEC-v1.1	protein-associated representation	17,445	300	86.80%
probe_pfam	PROBE-PFAM-v1.1	protein domain feature	18,873	6,227	93.91%
probe_protvec	PROBE-PROTVec-v1.1	protein representation	18,873	100	93.91%
probe_seqvec	PROBE-SeqVec-v1.1	protein representation	18,873	1,024	93.91%
probe_t5	PROBE-T5-v1.1	protein representation	18,453	1,024	91.82%
probe_tcga_embedding	PROBE-TCGA-EMBEDDING-v1.1	protein-associated / cancer expression representation	18,221	50	90.67%
probe_unirep	PROBE-UniRep-v1.1	protein representation	18,873	5,700	93.91%
probe_xlnet	PROBE-XLNet-v1.1	protein representation	18,453	1,024	91.82%

Single-cell foundation-model gene embeddings

Public ID	Embedding	Modality	Genes	Dimensions	Coverage
geneformer_v2_316m	Geneformer-V2-316M-v1.1	single-cell expression foundation model	20,007	1,152	99.55%
genecompass_base_gene_token	GeneCompass-Base-gene-token-v1.1	single-cell expression foundation model / knowledge-informed gene-token embedding	16,578	768	82.49%
scgpt_whole_human	scGPT-whole-human-v1.1	single-cell expression foundation model	19,280	512	95.93%
scgpt_pancancer	scGPT-PanCancer-v1.1	single-cell expression foundation model / cancer-context gene-token embedding	19,071	512	94.89%
uce_33l	UCE-33L-v1.1	single-cell expression foundation model / protein-informed gene-token embedding	19,082	1,280	94.95%
scfoundation_gene_pos	scFoundation-gene-pos-v1.1	single-cell expression foundation model / fixed learned gene-position embedding	19,062	768	94.85%
scprint_medium_v1_5_gene	scPRINT-medium-v1.5-gene-v1.1	single-cell expression foundation model / learned gene-token embedding	19,995	256	99.49%

Important interpretation notes

• Coverage differs by design. A matrix with lower coverage is not automatically lower quality. CMap, for example, represents genes with available consensus genetic-perturbation signatures and therefore has intentionally narrower coverage.
• Not every matrix is an internal hidden-state embedding. The resource also contains learned lookup tables, biological feature vectors, pooled sequence outputs, model prediction profiles, and compact reductions of regulatory outputs.
• Dimensions are not directly comparable. Some representations are compact latent vectors, while BLAST, HMMER, PFAM, Enformer, and Basenji2 retain high-dimensional feature spaces.
• The MSA Transformer entry is a depth-1 baseline. It uses one UniProt protein sequence represented as a one-row MSA. It must not be described as a homolog-rich MSA embedding.
• The deepNF entry is a reproducible local derivative. The retained matrix uses public STRING v12 evidence channels with sparse PPMI and SVD fusion; it is not claimed to be byte-identical to an unrecovered original human deepNF release.
• The OPA2Vec entry is locally reconstructed and OPA2Vec-derived. It follows the ontology/annotation embedding concept in a reproducible Python 3 workflow rather than claiming an unchanged run of the historical implementation.
• The Node2Vec PPI entry is a local reproducible run. The recovered consensus graph was used, but the original benchmark's exact hyperparameters were not fully recoverable.
• AlphaGenome is a derived regulatory-output representation. It summarizes selected CAGE, RNA-seq, chromatin-accessibility, histone, transcription-factor, and PRO-cap prediction tracks around 1 Mb TSS-centered windows, then applies signed-log transformation, z-scoring, and PCA to 256 dimensions. It is not an AlphaGenome trunk or hidden-state embedding.
• Single-cell checkpoint entries are fixed gene representations. They are extracted gene-token or position embeddings, not cell-contextual outputs from a new expression dataset.
• The repository is not a substitute for source documentation. Use each folder's README and metadata.json before interpreting or redistributing an embedding.

Contributing an embedding

A proposed contribution should provide one fixed vector per gene and include:

data/embeddings/<embedding_id>/
├── embeddings.npz
├── genes.tsv
├── metadata.json
└── README.md

Minimum requirements:

1. map genes to the repository's master universe using explicit identifier rules;
2. retain only unique final Ensembl gene mappings;
3. make genes.tsv row order exactly match the matrix;
4. store the matrix as float32 under the NPZ key embeddings;
5. document the source, license, checkpoint or release, input data, pooling or aggregation, and all transformations;
6. report missing, unmapped, ambiguous, and duplicate source identifiers;
7. pass the same finite-value, shape, duplication, and zero-vector QC;
8. state clearly whether the vector is downloaded, checkpoint-extracted, locally generated, or derived from model outputs.

New contributions should not silently impute absent genes or treat aliases as separate genes.

Versioning and reproducibility

• Public folder IDs: simplified and stable, with internal version information retained in metadata.
• Master universe: 20,097 human protein-coding genes.
• Current core release: 68 QC-passed representations.
• Reproducibility: processing scripts and detailed per-embedding notes are included where redistribution permits. Exact third-party checkpoints and raw inputs may need to be downloaded from their original providers.

Changes to biological values, sequence selection, identifier mapping, dimensionality reduction, or the master universe should trigger a documented new release rather than silently replacing files.

Licensing, attribution, and citation

This repository aggregates outputs from many independent projects. Before using or redistributing a matrix:

1. read its embedding-specific README and metadata.json;
2. follow the original source's license and terms;
3. cite the original method, model, dataset, and benchmark release where applicable;
4. cite this unified resource or its accompanying report once a formal citation is available;
5. do not assume that access to this repository grants rights beyond the original source terms.

No project DOI is claimed in this README. A formal citation entry should be added after the project report or archival release receives a persistent identifier.

Project

Building and Exploring a Unified Resource for Gene Embeddings
ETH Zürich Summer Research Fellowship, 2026

• Researcher: Daniel Tabares López
• Supervisors: Prof. Valentina Boeva and Michael Bohl

Acknowledgements

This resource depends on the work of the researchers and organizations that released the original models, embeddings, checkpoints, sequence resources, expression datasets, knowledge graphs, ontologies, and benchmark archives. Their individual source pages and citations are recorded in the corresponding embedding folders.