DRG Multiomic & Enhancer Activity Pipeline

This repository stores code for comprehensive single-cell multiomic processing for Mouse and Human Dorsal Root Ganglion (DRG) data, building in DRG based Gene Regulatory Network (GRN) inference via SCENIC+, and training Predicting Accessibility In Nociceptors-net (PAIN-net),.

---

📂 Repository Structure

1. PAIN-net (Predicting Accessibility In Nociceptors-net)

PAIN-net is a 1D Residual Neural Network (ResNet) designed to model the relationship between genomic sequence and cell-type-specific activity.

• Multi-Task Learning: The model predicts both Differential Expression (DE) $Log_2$ Fold Changes (primary task) and auxiliary ATAC-seq signal intensity (auxiliary task).

• Architecture: Features a stem Conv1D layer followed by 5 dilated residual blocks to capture long-range DNA interactions across 800bp input sequences.

• Weighted Loss Strategy: Implements inverse-frequency weighting plus explicit boosting (e.g., for C-PEP, Mrgprd, and Calca+Sstr2) to handle class imbalances.

• Custom Loss: Combines Mean Squared Error (MSE) and negative Pearson correlation to optimize for both magnitude and trend.

2. multiome/ (Species-Specific Processing)

This directory contains Level 1 processing pipelines for reconstructing and integrating single-cell data.

🐭 Mouse (Naive mDRG)

Processes 18 multiome libraries focusing on the RNA modality.

• Workflow: Performs Seurat v5 integration using CCA (Canonical Correlation Analysis) to correct batch effects.

👤 Human (hDRG)

Reprocesses raw fragment files for 32 human DRG libraries to generate a new Signac object.

• Workflow: Performs Seurat v5 integration using CCA (Canonical Correlation Analysis) to correct batch effects.

3. SCENIC+ (Regulatory Network Inference)

A stepwise workflow to generate region sets and topic models from Signac/Seurat objects for GRN inference.

Step	Script	Purpose
1	Step1_writepycisTopicInputs.R	Exports peak matrix, cells, and metadata from Signac for pycisTopic compatibility.
2	Step2_writeSignacDAPsBEDs.R	Exports cluster-specific Differential Accessibility Peaks (DARs) as BED files.
3	Step3_create_custom_cisTarget_database.sh	Builds a custom cisTarget motif database restricted to consensus peaks.
4	Step4_Run_pycisTopic.py	Runs MALLET topic modeling, binarizes topics (Otsu/Top-3k), and exports BEDs.
5	Step5_RunSCENICPlus.sh	Initializes and runs the SCENIC+ Snakemake pipeline.

---

🚀 Recommended Execution Plan

1. Parallel Group 1 (Export): Run Step1_writepycisTopicInputs.R and Step2_writeSignacDAPsBEDs.R to generate initial matrices and BED files.
2. Parallel Group 2 (Modeling): Build the custom motif database (Step 3) and run pycisTopic modeling (Step 4).
3. Final (Inference): Edit config.yaml to include species-specific resources (e.g., mm10 for mouse, hg38 for human) and run the SCENIC+ Snakemake pipeline.

---

⚠️ Tips & Pitfalls

• Genome Consistency: Ensure peaks, blacklist files, reference FASTAs, and chromsizes match your species (mm10 vs hg38).
• Barcode Reconciliation: In Step 4, ensure the ___cisTopic suffix is removed from cell names to match RNA AnnData barcodes for multiomic integration.
• Compute Intensity: Steps 3 and 4 are CPU-heavy and should be run as batch jobs on an HPC system.