Human Development Multiomic Atlas

This repository accompanies the preprint Dissecting regulatory syntax in human development with scalable multiomics and deep learning (Liu*, Jessa*, Kim*, Ng*, ..., Kundaje+, Farh+, Greenleaf+, bioRxiv, 2025).

* equal contribution
+ co-corresponding

• The repository is on GitHub here and you can view a rendered version here
• This repository contains primarily code, see the Data availability section for links to data, and our documentation here for download instructions and explanations of data formats
• Jump to the Code to reproduce figures section for links to code and rendered HTMLs for analysis presented in each figure

Contents

• Codebase
• Code to produce the figures
• Data availability
• Installation and system requirements
• Demo
  • Inputs and outputs
  • Vignettes
• Citation

Codebase

This repository is meant to enhance the Materials & Methods section by providing code for the custom analyses in the manuscript, in order to improve reproducibility for the main results. However, it is not a fully executable workflow.

• code --> pipelines, scripts, and analysis notebooks for data processing and analysis
  • utils --> contains .R files with custom functions and palettes used throughout the analysis
  • 01-preprocessing
    • 01-snakemake --> config files for processing raw bcl files into fragment files and count matrices
    • 02-archr_seurat_scripts --> per organ preprocessing scripts to create final Seurat objects and ArchR projects
    • 03-global --> creating global objects (e.g. global peak set, marker genes)
  • 02-global_analysis
    • 01 --> global QC and metadata visualizations per organ and per sample
    • 02, 03 --> construction of dendrogram on cell type similarity
    • 04 --> calculate TF expression levels
  • 03-chrombpnet
    • detailed README here
    • 00 --> prepare inputs for training ChromBPNet models
    • 01 --> train and interpret ChromBPNet models
    • 02 --> assembly of motif compendium/lexicon
    • 03 --> downstream analysis of ChromBPNet models and motif syntax/synergy
  • 04-enhancers
    • 01 --> export global accessible candidate cis-regulatory elements (acCREs)
    • 02 --> convert fragment files to tagalign for running Activity-By-Contact model (ABC)
    • 03 --> ABC workflow config files
    • 04 --> acCREs co-accessibility analysis
    • 05 --> acCREs peak-to-gene linkage analysis
    • 06 --> acCREs ABC enhancer-to-promoter linkage analysis
    • 07 --> overlap of HDMA acCREs with ENCODE v4 cCREs
    • 08, 09 --> overlap of HDMA acCREs with VISTA enhancers
  • 05-misc
    • 01 --> create global BPCells object
    • 02 --> examples for plotting tracks using BPCells
    • 04 --> examples for ChromBPNet use cases, including how to load models and make predictions
  • 06-variants
    • 00 to 03 --> analysis related to eQTLs
    • 04 to 05 --> causal variant analysis with gchromvar
    • 06 --> variant scoring using ChromBPNet models
    • 07a to 07c --> plot variant scoring results

Our pipeline for processing SHARE-seq data is available at https://github.com/GreenleafLab/shareseq-pipeline (v1.0.0).

Code to produce the figures

Code to reproduce analyses is saved in code. This table contains pointers to code for the key analyses associated with each figure. The links in the Analysis column lead to rendered HTMLs, where possible, and the links in the Path column lead to scripts or notebooks within the repository.

Figure	Analysis	Path
Fig 1b, Fig S2b,c	Global QC and metadata	code/02-global_analysis/01-global_QC.Rmd
Fig 1c	Dendrogram and dotplot	code/02-global_analysis/02-dendrogram.Rmd
Fig 1c	ChromVAR heatmap	code/02-global_analysis/03-dendrogram_chromvar.Rmd
Fig 2a-e, Fig S2f	ABC linking of acCREs	code/04-enhancers/06-abc.Rmd
Fig 2f-g, Fig S3a, Fig S4k	Analysis of VISTA-overlapping enhancers	code/04-enhancers/09-overlap_VISTA.Rmd
Fig S2d-e	Overlap of acCREs with ENCODE CREs	code/04-enhancers/07-overlap_ENCODE_cCREs.Rmd
Fig 3b, Fig 6a, Fig S5	Plotting tracks at select loci	code/03-chrombpnet/03-syntax/02-plot_tracks.Rmd
Fig 3c, Fig S4a,b,i,j	ChromBPNet QC and correlation plot	code/03-chrombpnet/01-train_models/03-model_QC.Rmd and code/03-chrombpnet/01-train_models/03b-plot_correlation.ipynb
Fig 3d-e, Fig 6b,d, Fig S4d-f, Fig S5b	Motif lexicon/compendium	code/03-chrombpnet/03-syntax/01-motif_compendium
Fig S4g-h	Visualize motif instances	code/03-chrombpnet/03-syntax/03-visualize_hits.ipynb
Fig 4, Fig 5a, Fig S6	Analysis of motif cooperativity/synergy and syntax	code/03-chrombpnet/03-syntax/04c-plot_cooperativity_results.Rmd
Fig 5b	Context-specific motif cooperativity	code/03-chrombpnet/03-syntax/05b-context_specific_cooperativity.Rmd
Fig 6f, Fig S7	eQTL enrichment analysis	code/06-variants/03-enrichment_test_collate_results.R
Fig 7b	g-chromVAR analysis	code/06-variants/04-gchromvar.R
Fig 7c-d	Plot tracks for variants of interest	code/06-variants/07b_rs12740374_muscle_endo_CAD.R and code/06-variants/07c_rs113892147_lung_macrophage_asthma.R
Fig S8	Plot tracks for all fetal-only variants	code/06-variants/07a_plot_fetal_only_hits_variant_scoring_results.R

Data availability

All data and analysis products (including fragment files, counts matrices, cell annotations, global acCRE annotations, ChromBPNet models, motif lexicon, motif instances, and genomic tracks) are deposited at https://zenodo.org/communities/hdma. A list of all data types and the corresponding URL and DOI is provided in Table S14 of the manuscript.

We provide a detailed description of the main data types deposited on Zenodo here, along with a demonstration of how to programmatically download files of interest.

All genomic tracks are also hosted online for interactive visualization with the WashU Genome Browser here at this link: https://epigenomegateway.wustl.edu/browser2022/?genome=hg38&hub=https://human-dev-multiome-atlas.s3.amazonaws.com/tracks/HDMA_trackhub.json. We demonstrate how to load tracks here.

Installation and system requirements

Installation

This repository can be cloned locally (expected time: several minutes due to large HTMLs included in the repository). Scripts are meant to be run on a Linux OS inside a Slurm-enabled HPC system. We performed our analysis on CentOS 7.9.2009. To run the code, one would need to set up the appropriate R and python environments, for example using renv and conda, respectively.

R

Analysis conducted in R was performed using R 4.1.2, and the package version for each analysis is listed in the "Session info" section at the end of each rendered HTML (e.g. here)

Python

Analysis conducted in python depended on key packages and their own dependencies. Package environments were managed using conda, and the conda environment used for a particular analysis is typically specified in the associated Slurm submission script. A full list of package versions in each environment is located at code/envs.

Demo

Inputs and outputs

Our repository is designed to enable reproducibility for the results by providing exact code and software/package versions, although it is not a fully executable workflow.

Below, we describe the major inputs and outputs for each section of the code. To demo the code on a small dataset, we link here to inputs and outputs from the Adrenal organ (total ~3k cells). Note that some analyses are, by definition, integrative, and won't be possible to execute with only one organ. All files are hosted on Zenodo (https://zenodo.org/communities/hdma/records?q=&l=list&p=1&s=10). When outputs are not specified, the major outputs are typically plots and data visualizations, wich can be viewed above in the section "Code to produce the figures". Code is intended to be run on a Slurm-enabled high-performance compute machine allowing for parallel jobs. On a small dataset such as the Adrenal organ, expected execution time on an HPC is on the order of several days.

• code/01-preprocessing:
  • Inputs: Gene expression counts and chromatin accessibility fragment files
  • Outputs:
    • Seurat object
    • ArchR project
    • BPCells object
• code/02-global_analysis:
  • Inputs: Seurat object
• code/03-chrombpnet:
  • Inputs: ArchR project
  • Outputs:
    • Trained ChromBPNet models in h5 format, five folds per cell type
    • Basepair-resolution contribution scores in h5 format, one per cell type
    • Bigwig tracks viewable in a genome browser
    • De novo motifs in h5 format, one per cell type
• code/04-enhancers:
  • Inputs:
    • Chromatin accessibility fragment files
    • ArchR project
  • Outputs:
    • ABC loops in TSV format, one per cell type
• code/06-variants:
  • Inputs:
    • ArchR project
    • ChromBPNet model
    • Motif instances

Vignettes

We provide a few notebooks with examples of how to interact with HDMA data, analysis outputs, and trained models:

• How to download specific files or data for specific cell types from across the Zenodo records: DATA.md (html)
• Plotting genomic tracks using BPCells: code/05-misc/02-bp_cells_plotting_examples.Rmd (html)
• Use cases for ChromBPNet models and outputs, including visualizing predicted accessibility and contribution scores at a region of interest, loading models, making new predictions, and predicting variant effect: code/05-misc/04-ChromBPNet_use_cases.ipynb (html)

Citation

If you use this data or code, please cite:

Dissecting regulatory syntax in human development with scalable multiomics and deep learning. Betty B. Liu, Selin Jessa, Samuel H. Kim, Yan Ting Ng, Soon il Higashino, Georgi K. Marinov, Derek C. Chen, Benjamin E. Parks, Li Li, Tri C. Nguyen, Sean K. Wang, Austin T. Wang, Serena Y. Tan, Michael Kosicki, Len A. Pennacchio, Eyal Ben-David, Anca M. Pasca, Anshul Kundaje, Kyle K.H. Farh, William J. Greenleaf, bioRxiv 2025.04.30.651381; doi: https://doi.org/10.1101/2025.04.30.651381