Datasets:
H2dddhxh
/
AI4Genome

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
80
================================================================================
gene_pooled_embeddings.npz — AlphaGenome gene-level DNA embeddings (pooled)
================================================================================
WHAT THIS FILE IS
-----------------
One row per gene: AlphaGenome sequence embeddings (128 bp windows, 3072 dims)
were averaged over all windows L_i for that gene (mean pooling along the
sequence). Model / layer match the raw shards (see SHARD_KEYS_AND_SAMPLE.md).
File: gene_pooled_embeddings.npz (NumPy compressed archive, ~500+ MB)
WHO HAS WHICH GENES
-------------------
Rows are protein_coding and lncRNA genes from the filtered AlphaGenome run
(54,901 genes in the version built with default pooling). Each gene_id appears
once (unique Ensembl IDs with version, e.g. ENSG00000000003.14).
CONTENTS (keys inside the .npz)
--------------------------------
pooled (G, 3072) float32 main matrix — one vector per gene
gene_ids (G,) str unique Ensembl gene ID (versioned)
gene_symbols (G,) str HGNC symbol (not globally unique)
gene_types (G,) str e.g. protein_coding, lncRNA
chroms (G,) str
starts (G,) str genomic coordinates (as stored in shards)
ends (G,) str
strands (G,) str + or -
seq_lengths (G,) int32 input DNA length in bp
G = number of genes. Row index i is consistent across all arrays.
HOW TO LOAD IN PYTHON
---------------------
import numpy as np
path = "gene_pooled_embeddings.npz" # or full path on cluster
data = np.load(path, allow_pickle=True)
X = data["pooled"] # (G, 3072) float32
ids = data["gene_ids"] # use for joins; compare as str(id)
# Example: map Ensembl -> row index
id_to_row = {str(g): i for i, g in enumerate(data["gene_ids"])}
data.close() # optional; good practice with np.load
Note: allow_pickle=True is required because string columns are stored as
numpy object arrays.
JOINING WITH OTHER TABLES
-------------------------
- Match on gene_ids as strings (same version string as in GTF/FASTA metadata).
- gene_symbols can repeat across genes; prefer gene_ids for unique keys.
REGENERATING OR OTHER POOL MODES
--------------------------------
Script: pool_embeddings.py (same directory)
- Default pool mode: mean -> (G, 3072)
- Other modes: max, cls, first window, last window, mean+max -> (G, 6144)
Batch job example (repo root): pool_alphagenome_embeddings.slurm
Raw per-window embeddings (much larger) live in shard_*_of_005.npz or
alphagenome_shards_and_doc.tar — see SHARD_KEYS_AND_SAMPLE.md.
QUESTIONS
---------
See SHARD_KEYS_AND_SAMPLE.md for raw shard layout and embedding geometry
(L_i = ceil(seq_length / 128) windows per gene).
================================================================================