geniml

You can convert BED files into ML-ready genomic embeddings by building a consensus peak universe, tokenizing the genomic intervals, and training unsupervised models like Region2Vec to generate region representations for similarity search and downstream modeling.

Generating cell embeddings from scATAC-seq data for clustering requires using scEmbed to learn representations from cell-level accessibility signals. You first need a reference universe to tokenize the scATAC-seq dataset before training the model for annotation and visualization.

Consensus peak universe building aggregates genomic intervals from multiple BED file collections into a unified reference. This universe standardizes tokenization across datasets, ensuring consistent preprocessing when training region or cell embeddings for cross-experiment comparisons.

Yes, you can create shared embeddings for genomic regions and metadata labels using BEDspace. This supports cross-modal queries, allowing you to compare BED file intervals and experimental metadata within a unified embedding space for downstream analysis.

Yes, a reference universe is required before tokenizing BED files for machine learning. Building a consensus peak universe establishes the standardized genomic coordinate vocabulary necessary for consistent tokenization and model training across different datasets.

Unsupervised region embeddings depend entirely on the quality of the consensus peak universe and tokenization process. If the reference universe poorly represents your scATAC-seq accessibility signals, the resulting cell embeddings may produce inaccurate clustering and annotation results.