geniml

geniml: A Machine Learning Analysis Tool for Genomic Interval Data

Overview of Capabilities

geniml is a Python package specifically designed for machine learning on genomic interval data (BED files). It provides unsupervised embedding learning methods such as Region2Vec, BEDspace, and scEmbed, and supports genomic region similarity analysis, single-cell ATAC-seq clustering, and consensus peak set construction.

Use Cases

Single-Cell ATAC-seq Data Analysis

Suitable for dimensionality reduction and clustering analysis of single-cell chromatin accessibility sequencing data. scEmbed can be used to train cell-level embeddings and integrate seamlessly with scanpy workflows for cell type annotation and visualization.

Machine Learning on Large-Scale BED Files

Suitable for unsupervised embedding training on large-scale genomic datasets such as ChIP-seq and ATAC-seq. Region2Vec converts genomic regions into feature vectors, supporting region similarity searches, cross-experiment comparisons, and downstream supervised learning tasks.

Consensus Peak Set Construction and Standardization

Suitable for constructing reference peak sets (universes) from collections of BED files. It provides various statistical methods, including CC, CCF, ML, and HMM, which can be used to define standardized regions, create tokenization references, and integrate datasets.

Core Features

Region2Vec: Genomic Region Embedding Training

Based on an unsupervised word2vec-style learning method, Region2Vec trains genomic regions into low-dimensional embedding vectors. It supports learning region representations from collections of BED files for dimensionality reduction, similarity analysis, and feature construction for downstream machine learning tasks. A universe reference can be used for tokenization, and the embedding dimension and training parameters can be customized.

BEDspace: Joint Embedding of Regions and Metadata

BEDspace uses the StarSpace algorithm to train a shared embedding space for region sets and metadata labels. It supports cross-modal queries, such as querying labels by regions or regions by labels. It is suitable for scenarios with metadata labels, such as cell types, tissues, and experimental conditions, enabling metadata-aware searches and joint analyses.

scEmbed: Single-Cell Chromatin Accessibility Embedding

scEmbed is a Region2Vec training method specifically designed for single-cell ATAC-seq data and can generate cell-level embedding vectors. It supports pre-tokenization to accelerate training. The resulting embeddings can be directly stored in the obsm field of an AnnData object and used with scanpy for clustering, UMAP dimensionality reduction, and cell type annotation.

Frequently Asked Questions

What input data formats does geniml support?

geniml primarily supports genomic interval data files in BED format, the standard format for representing genomic region coordinates. For single-cell ATAC-seq, it supports input through AnnData objects (.h5ad files), which must contain peak coordinate information. The tool also supports processing collections of BED files and coverage data through the command line or Python API.

How should I choose between Region2Vec, BEDspace, and scEmbed?

The choice depends on your data type and analysis goals. If you are working with single-cell ATAC-seq data and need cell-level analysis, choose scEmbed. If you have metadata labels, such as cell types or experimental conditions, and need cross-modal queries, choose BEDspace. If you are working with bulk genomic data, such as ChIP-seq or ATAC-seq, and only need region-level embeddings, choose Region2Vec. The three methods can be used independently or in combination.

What are the computational resource requirements for geniml?

The resource requirements of geniml depend on the data scale and the algorithm selected. Basic tokenization and the CC consensus peak method have relatively low computational requirements. Region2Vec and scEmbed training require machine learning dependencies such as PyTorch, and GPU acceleration is recommended. BEDspace requires StarSpace. The ML and HMM consensus peak methods are computationally intensive and are better suited to smaller datasets or high-performance computing environments. For large-scale single-cell datasets, pre-tokenization and batch processing are recommended to manage memory usage.