Base-resolution deep learning framework for mapping transcription factor (TF) binding dynamics in the single-cell and bulk ATAC-seq atlas of the mouse and rat pituitary gland.
This repository contains the custom pipeline, downstream analysis scripts, and metadata developed for the MSc Applied Bioinformatics dissertation "Applying deep learning to map TF-binding patterns in the single-cell ATAC-seq atlas of the pituitary gland" (Georgia Goddard, King's College London, 2026), supervised by Cynthia Andoniadou, Grace Hui-Chun Lu, and Bence Kövér, Andoniadou Lab.
The pipeline applies ChromBPNet — a bias-factorised, base-resolution deep learning model of chromatin accessibility — to the Andoniadou Lab's Consensus Pituitary Atlas (CPA) and complementary bulk ATAC-seq as well as BPNet models on ChIP-seq and CUT&RUN datasets, in order to identify TF binding events de novo and characterise regulatory dynamics across pituitary lineage commitment.
- Overview
- Repository structure
- Pipeline workflow
- Data
- Environment & installation
- Reproducing the analysis
- Model availability
Two aims motivated this research and are outlined within this repository:
- Aim 1 — Identify TF binding events de novo and characterise the key TFs associated with regulatory dynamics across pituitary lineage transitions and cell types.
- Aim 2 — Establish a framework for identifying regulatory sites near genes of interest, using the prolactin (Prl) locus as a candidate gene.
To do this, ChromBPNet models were trained per cell type / experimental condition on:
- Single-cell ATAC-seq from the Consensus Pituitary Atlas (adult and neonatal mouse, ~350,000 cells, 59 datasets), including a stem cell → gonadotroph pseudotime trajectory.
- Bulk ATAC-seq from pituitary cell lines (GH3, AtT-20, LβT2, TaT1), including a GH3 Nfia/Nfib and Sox2 overexpression panel.
- ChIP-seq / CUT&RUN (POU1F1, TBX19, SOX2), modelled with the related BPNet architecture for orthogonal validation of ChromBPNet's de novo motifs. Model interpretation used DeepLIFT contribution scores, TF-MoDISco de novo motif discovery, and marginal footprinting, benchmarked against the JASPAR 2026 vertebrate motif database.
Pituitary_Chromatin_Atlas/
├── README.md
├── BPNet_environment.yml # Conda environment for the BPNet pipeline
├── ChromBPNet_environment.yml # Conda environment for the ChromBPNet pipeline
├── figures/ # Figures generated for the research project writeup (.svg)
│
├── BPNet_pipeline/ # BPNet models trained on ChIP-seq/CUT&RUN data
│ ├── 0_negatives_bpnet.sh # GC-matched background region generation
│ ├── 1_generate_JSON.sh # Pipeline configurations
│ └── 2_run_bpnet.sh # Training, evaluation, DeepLIFT attribution (contribution scores)
│
├── ChromBPNet_pipeline/ # Custom ChromBPNet pipeline
│ ├── 1_Preprocessing/ # Stage 1: fragment splitting, merging, peak/non-peak prep
│ ├── 2_Model_Training/ # Stage 2: Tn5 bias + bias-factorised ChromBPNet training
│ ├── 3_Data_Generation/ # Stage 3: predicted BigWigs, footprints, contributions, motifs
│ ├── functions/ # Shared/custom pipeline functions
│ └── PipelineDiagram.png # Schematic of the full workflow
│
├── Downstream_Analyses/
│ ├── contribution_score_analysis/ # IGV locus plots, correlations, MotifMatchr/FIMO scanning of high-contribution loci
│ ├── marginal_footprint_analysis/ # Plot footprints, PCA of footprint changes, pseudotime comparisons
│ ├── motif_analysis/ # TF-MoDISco output parsing, JASPAR comparison, novel motif identification
│ └── other/ # RNA-seq integration, expression filtering, miscellaneous notebooks, model evaluation plotting
│
├── Example_Model_Evaluations/
│ ├── Tn5_bias_model_fold_0_mm10_adult_0.5/ # Example trained bias model + logs
│ ├── ChromBPNet_model_mm10_adult_lactotrophs/ # Example trained ChromBPNet model + logs
│ ├── bias_model_metrics.csv # Pearson r / JSD metrics, all bias models
│ └── chrombpnet_metrics.csv # Pearson r / JSD metrics, all ChromBPNet models
│
└── metadata/
├── consensus_chromatin_landscape_mm10_adult.bed # Adult mm10 consensus peak set
├── consensus_chromatin_landscape_mm10_neonatal.bed # Neonatal mm10 consensus peak set
├── consensus_chromatin_landscape_hg38.bed # hg38 consensus peak set
├── JASPAR_CORE_2026_non-redundant.meme # JASPAR 2026 vertebrate motifs (2059 motifs)
├── JASPAR_CORE_2026_redundant.meme # JASPAR 2026 vertebrate reduced list of motifs (951 motifs)
├── JASPAR_CORE_2026_with_novel_POU1F1.meme # JASPAR 2026 vertebrate motifs + de novo POU1F1 motif (2060 motifs)
├── motif_sequences.tsv # Curated motif sequences from JASPAR_CORE_2026_non-redundant.meme
├── PCA_motif_sequences.tsv # Reduced list of motif sequences used for downstream PCA analysis
├── atac_grouping_lineage_markers.csv # Curated lineage-marker TF list (CPA)
├── median_RNAexpression_per_celltype.csv # Pseudobulked expression (CPA), log10CPM, used for TF filtering
└── pituitary_epigenomics.xlsx # Sample/dataset metadata sheet
The custom ChromBPNet_pipeline/ is organised into three sequential stages (see
PipelineDiagram.png), each implemented as a set of custom functions wrapping the
chrombpnet command-line pipeline. This adapts and expands the original
pipeline to accomodate the large scale Consensus Pituitary Atlas.
Refer to the Wiki for the organisation and functions of the workflow.
BPNet_pipeline/ mirrors the ChromBPNet workflow for orthogonal validation datasets
(POU1F1 CUT&RUN, TBX19 and SOX2 ChIP-seq), using
bpnet-lite v1.0.0 (no bias-correction step, as
ChIP/CUT&RUN data does not use the Tn5 transposase enzyme).
contribution_score_analysis/— locus-level IGV inspection, FIMO/MotifMatchr scanning of high-contribution sequences (e.g. the Prl upstream regulatory regions).marginal_footprint_analysis/— PCA-based quantification of footprint shape/amplitude across pseudotime, cell types, and age.motif_analysis/— TF-MoDISco motif parsing, cross-referencing with JASPAR, curation of the novel POU1F1 motif.other/— RNA-seq/scRNA-seq integration for TF expression filtering and validation, as well as other miscellaneous scripts.
| Dataset | Description | Source |
|---|---|---|
| Consensus Pituitary Atlas (CPA) | scATAC-seq/scRNA-seq, ~1.3M cells, 59 datasets | Kövér et al., 2026 (Andoniadou Lab) |
| Bulk ATAC-seq, GH3 lines | CTRL, Nfia+b OE (AB), Sox2 OE (SOX2), Nfia+b+Sox2 OE (ABS) | Andoniadou Lab (unpublished) |
| Bulk ATAC-seq, AtT-20 (Neo/Pax7), LβT2, TaT1 | Corticotroph/melanotroph, gonadotroph, thyrotroph cell lines | Mayran et al. 2019; Ruf-Zamojski et al. 2018; Daly et al. 2021 |
| POU1F1 CUT&RUN | TaT1 cell line | Daly et al. 2021 |
| TBX19 ChIP-seq | AtT-20 cell line | Zhang et al. 2015 |
| SOX2 ChIP-seq | AtT-20 cell line | Drouin Laboratory (unpublished) |
| JASPAR CORE 2026 vertebrates | 2059 non-redundant motifs | JASPAR |
| Reference genomes | mm10 (mouse), rn6 (rat) + ENCODE blacklist | UCSC / ENCODE |
Raw sequencing data and full-size trained models are not hosted in this repository due to
size; see Model availability below. Consensus peak sets and curated
motif/metadata files needed to reproduce downstream analyses are provided in Metadata/. All
data described above that is not included within the Consensus Pituitary Atlas are retrievable from
the Metadata/pituitary_epigenomics.xlsx database.
Two conda environments are provided at the base of this repository, mirroring the exact
setups used on the HPC cluster for this project (see Supplementary Table 1 of the
dissertation for the full version-pinned package list underlying each):
BPNet_environment.yml— forBPNet_pipeline/(used for the ChIP-seq/CUT&RUN validation models).ChromBPNet_environment.yml— forChromBPNet_pipeline/andDownstream_Analyses/(chrombpnet,modisco-lite, and associated Python analysis packages). Clone the repository, then create both environments from the repo root:
git clone <repo-url> Pituitary_Chromatin_Atlas
cd Pituitary_Chromatin_Atlas
conda env create -f BPNet_environment.yml
conda env create -f ChromBPNet_environment.ymlActivate the relevant environment depending on which pipeline is being run:
# For BPNet_pipeline/ scripts
conda activate BPNet-lite
# For ChromBPNet_pipeline/ and Downstream_Analyses/ scripts
conda activate ChromBPNetR (v4.5.1) environment (for motif matching, plotting, GenomicRanges-based workflows):
install.packages(c("universalmotif", "ggplot2"))
BiocManager::install(c("motifmatchr", "TFBSTools", "GenomicRanges",
"SummarizedExperiment", "rhdf5", "BiocParallel",
"BSgenome.Mmusculus.UCSC.mm10",
"BSgenome.Rnorvegicus.UCSC.rn6"))All deep learning models were trained on NVIDIA A100 GPU nodes; GPU access (CUDA-enabled)
is required for model training steps (Section 2), though inference/downstream analysis
(Section 3, Downstream_Analyses/) can run on CPU.
Wiki with detailed step-by step instructions: https://github.com/ge8rgia/Pituitary_Chromatin_Atlas/wiki
This tutorial describes required inputs, expected outputs, and adjustable parameters (e.g., bias_threshold_factor, seqlet counts, PCA distance thresholds), all of which can be adapted based on the required analysis.
The ChromBPNet and BPNet pipelines are annotated numerically, and should be reproduced following this structure. All downstream analyses can be conducted in any order,
separately or in combination depending on the research expected. All Downstream_Analyses notebooks describe the inputs required, which link to one of the steps in the ChromBPNet_pipeline/3_Data_Generation workflow.
- Code & configuration: this repository.
- Trained models (all 40 Bias, ChromBPNet, and BPNet models): hosted on Hugging Face
- Example model + evaluation outputs:
Example_Model_Evaluations/(adult mm10 lactotroph ChromBPNet model and its corresponding Tn5 bias model, provided as worked examples of expected directory structure and QC metrics).