multiHIVE 0.1.0

multiHIVE, a hierarchical multimodal deep generative model for integrating multiomics data

multiHIVE

Overview

multiHIVE is a hierarchical multimodal deep generative model designed to infer cellular embeddings by integrating multi-omics data with different modalities from the same cell. It uses:

• Hierarchically stacked latent variables to capture shared biological signals
• Modality-specific latent variables to model private (modality-unique) variation

This enables multiHIVE to perform:

• Joint integration of multi-modal data
• Denoising
• Protein imputation
• Integration of multi-modal and uni-modal datasets

Additionally, multiHIVE enables factorization of denoised gene expression into interpretable gene expression programs, facilitating the identification of biological processes at multiple levels of cellular hierarchy.

Basic Installation

we recommend users to directly clone our stable main branch and set multiHIVE as the working directory and install following dependencies in a new conda environment python>=3.11

git clone https://github.com/Zafar-Lab/multiHIVE.git
pip install scvi-tools==1.3.0
pip install scanpy==1.11.0
pip install scikit-misc

Or install directly via pip

pip install multiHIVE

Tutorials

Explore the following tutorials to get started:

• CITE-seq integration: Tutorials/CITE_seq_Integration.ipynb
• Protein imputation: Tutorials/Protein_Imputatoin.ipynb
• TEA-seq integration: Tutorials/TEA-seq_integration.ipynb

1. Main Script:

# features should genes followed by regions [genes, regions]
multiHIVE.setup_anndata(adata, batch_key="batch", protein_expression_obsm_key = "protein_expression")

vae = multiHIVE(adata,  
            n_genes=(adata.var["modality"] == "Gene Expression").sum(), # number of genes 
            n_regions=(adata.var["modality"] == "Peaks").sum(), # number of regions
            n_proteins=46, # number of proteins 
            latent_distribution="normal", kl_dot_product=True, deep_network=True)
vae.train()
vae.get_latent_representation()

2. Model Parameters:

Parameter	Description
latent_distribution	Distribution for latent variables (e.g., "normal")
kl_dot_product	Enables regularization using dot-product of modality-specific latents
deep_network	Uses deeper neural networks; recommended for datasets > 100,000 cells

3. Results:

• vae.get_latent_representation() gives zs1, zs2, zr and zp or/and za
• zs1 is the joint latent variable.
• zs2 is the hierarchical joint latent variable.
• zr is the gene modality specific latent variable.
• zp is the protein modality specific latent variable.
• za is the chromatin accessibility specific latent variable

Documentation

For more advanced settings, preprocessing tips, and API references, refer to the multiHIVE Documentation (link coming soon)

Citation

multiHIVE: Hierarchical Multimodal Deep Generative Model for Single-cell Multiomics Integration
Anirudh Nanduri*, Musale Krushna Pavan*, Kushagra Pandey, Hamim Zafar
bioRxiv 2025.01.28.635222; doi: https://doi.org/10.1101/2025.01.28.635222
*Equal contribution

Contact

For questions, issues, or contributions, please open an issue on the GitHub repository