alterlab-geniml

What problem does it solve?

Geniml brings a complete toolkit to build unsupervised embeddings from genomic interval data, enabling similarity searches, clustering, and downstream ML analyses across BED files, scATAC-seq data, and consensus peak sets.

Core Features & Use Cases

• Region2Vec: learn embeddings for genomic regions to reduce dimensionality and enable region-level analyses.
• BEDspace: jointly embed regions and metadata labels for metadata-aware searches across regions and labels.
• scEmbed: generate cell embeddings from scATAC-seq data for clustering and annotation.
• Universe building: construct consensus peak universes to standardize tokenization references.
• Utilities: caching, randomization, evaluation, tokenization, and search backends for reproducible pipelines.

Use cases include clustering cells, performing similarity queries across datasets, and building tokenization universes for cross-project analyses.

Quick Start

Run Geniml with a prepared universe and tokenized BED files to train embeddings and evaluate them on a sample metadata file.