What problem does it solve?
Genomic interval machine learning turns BED-region collections into useful vector representations so you can perform similarity search, clustering, and other ML tasks without hand-engineering features for peaks and single-cell chromatin accessibility.
Core Features & Use Cases
- Region2Vec: Learn unsupervised embeddings of genomic regions from tokenized BED intervals for region similarity and downstream ML features.
- BEDspace: Train joint embeddings for region sets and metadata labels (via StarSpace) to enable metadata-aware searches across regions and experimental conditions.
- scEmbed: Generate cell-level embeddings for single-cell ATAC-seq (scATAC-seq) to support clustering, annotation, and scanpy workflows.
- Consensus peaks (universes): Build standardized reference peak sets from multiple BED files to define a consistent vocabulary for tokenization.
- Utilities: Support caching (BBClient), null-model randomization (BEDshift), embedding evaluation, and region tokenization/search backends.
Use case example: You have multiple ATAC-seq BED files and want a shared peak vocabulary plus region embeddings; you then build a universe from the BED collection, tokenize regions against that universe, and train Region2Vec (or scEmbed for single-cell) to cluster similar regulatory regions across experiments.
Quick Start
Use the geniml skill to train genomic region embeddings from your BED files by building a consensus universe, tokenizing into a model-ready format, and running the appropriate Region2Vec or scEmbed training pipeline from the provided references.
