How to create a natural language search backend for BED files

The metadata of each BED file is needed to build a natural language search backend. BED files embedding vectors are created by Region2Vec model, and metadata embedding vectors are created by FastEmbed, SentenceTransformers, or other text embedding models.

Vec2VecFNN, a feedforward neural network (FNN), is trained to maps vectors from the embedding space of natural language to the embedding space of BED files. When a natural language query string is given, it will first be encoded to a vector by the text embedding model, and then created vector will be encoded to a query vector by the FNN. search backend can perform k-nearest neighbors (KNN) search among the stored BED file embedding vectors, and the BED files whose embedding vectors are closest to that query vector are the search results.

Store embedding vectors

It is recommended to use geniml.search.backend.HNSWBackend to store embedding vectors. In the HNSWBackend that stores each BED file embedding vector, the payload should contain the name or identifier of BED file. In the HNSWBackend that stores the embedding vectors of each metadata string, the payload should contain the original string text and the names of BED files that have that string in metadata.

Train the model

Training a Vec2VecFNN needs x-y pairs of vectors (x: metadata embedding vector; y: BED embedding vector). A pair of a metadata embedding vector with the embedding vectors of BED files in its payload is a target pair, othersie a non-target pair. Non-target pairs are sampled for contrastive loss. Here is sample code to generate pairs from storage backend and train the model:

# target is an array of 1 (target) and -1 (non-target) 
X, Y, target = vec_pairs(
    nl_backend,  # HNSWBackend that store metadata embedding vectors
    bed_backend,  # HNSWBackend that store BED embedding vectors
    "name",  # key to file name in BED backend payloads
    "files",  # key to matching files in metadata backend payloads
    True,  # sample non-target pairs
    1.0  # number of non-target pairs /number of target pairs = 1
)

# train without validate data
v2v_torch_contrast.train(
    X,
    Y,
    folder_path="path/to/folder/for/checkpoint",
    loss_func="cosine_embedding_loss",  # right now "cosine_embedding_loss" is the only contrastive loss function available
    training_target=target,
)

Search interface

A search interface consists of a storage backend where vectors are stored, and a module (geniml.search.query2vec) that embed the query. geniml.search supports two types of queries: region set query and text query.

Region set query

BED2Vec embed the query region set with a Region2VecExModel, and the embedding vector is used to perform KNN search within the backend.

from geniml.search import BED2BEDSearchInterface, BED2Vec

# init BED2Vec with a hugging face repo of a Region2VecExModel
bed2vec = BED2Vec("databio/r2v-ChIP-atlas-hg38-v2")

# the search_backend can be QdrantBackend or HNSWBackend
search_interface = BED2BEDSearchInterface(search_backend, bed2vec)

# the query cam be a RegionSet object (see geniml.io) or path to a BED file in disk
file_search_result = search_interface.query_search("path/to/a/bed/file.bed", 5)

Text query

Text2Vec embed the query string with a with a natural language embedding model first (default: FlagEmbedding), and then maps the text embedding vector into the embedding space of region sets through a trained Vec2VecFNN.

from geniml.search import Text2BEDSearchInterface, Text2Vec

text2vec = Text2Vec(
    "sentence-transformers/all-MiniLM-L6-v2",  # either a hugging face repo or an object from geniml.text2bednn.embedder
    "databio/v2v-geo-hg38"  # either a hugging face repo or a Vec2VecFNN
)

search_interface = Text2BEDSearchInterface(search_backend, text2vec)
text_search_result = search_interface.query_search("cancer cells", 5)

With a dictionary that contains query strings and id of relevant query results in search backend in this format:

{
    <query string>: [
        <id of relevant result in backend>,
        ...    
    ],
    ...
}

Text2BEDSearchInterface can return mean average precision, average AUC-ROC, and average R-Precision, here is example code:

query_dict = {
    "metadata string 1": [2, 3],
    "metadata string 12": [1],
    "metadata string 3": [2, 4, 5],
    "metadata string 1": [0]
}

MAP, AUC, RP = search_interface.eval(query_dict)