macwiatrak/bacformer-genome-embeddings-corpus

Corpus of >1.3M genome embeddings from Bacformer

A dataset of over 1.3M bacterial genome embeddings together with associated metadata.

The dataset was collated by combining publicly available bacterial genomes from MGnify, SPIRE and NCBI. The extracted genomes were then embedded with a pretrained Bacformer model (masked objective). The genome embeddings were calculated by taking the average of contextual protein embeddings from the last layer of the Bacformer encoder.

Usage

We outline a full example on how to use the dataset for scalable search of similar genomes in the Bacformer tutorial.

from datasets import load_dataset

# load the dataset
ds = load_dataset("macwiatrak/bacformer-genome-embeddings-corpus", split="test")

# add a FAISS index
ds.add_faiss_index(column="genome_embedding"

# get k most similar genomes
query_vector = np.array(ds[0]['genome_embedding'])
# set k (i.e. number of most similar genomes to retrieve)
k = 10
# get k most similar genomes
dists, examples = ds.get_nearest_examples(
        "genome_embedding",
        query_vector,
        k=k
    )
# convert the examples to a DF, exclude the first item as it's the genome embedding itself
examples = pd.DataFrame(examples)[1:].drop(columns=['genome_embedding'])
# add L2 distances
examples['l2_distance'] = dists[1:]
examples.head()

See github repository for more details.