Marderstein*, Kundu* et al.

This repository contains all scripts for the manuscript "Mapping the regulatory effects of common and rare non-coding variants across cellular and developmental contexts in the brain and heart" by Marderstein, Kundu et al.

For any questions, please contact:

Andrew Marderstein & Soumya Kundu 📧 [email protected], [email protected]

---

Pre-processing

These sets of scripts in the preprocess/ directory generate all of the outputs used for the downstream analyses.

• 0_process_data/ – Scripts for processing the scATAC-seq data.
• 1_train_chrombpnet/ – Scripts for training the ChromBPNet models.
• 2_score_variants/ – Scripts for scoring the rare, common, and ASD variants.
• 3_shap_variants/ – Scripts for generating DeepLIFT / DeepSHAP contribution scores for both alleles of each variant.
• 4_shap_peaks/ - Scripts for generating DeepLIFT / DeepSHAP contribution scores for scATAC-seq peaks.
• 5_run_modisco/ - Scripts for running TF-MoDISco to identify the motif patterns learned by each model.
• 6_cluster_motifs/ - Scripts for running MotifCompendium to cluster the motif patterns from TF-MoDISco.
• 7_run_finemo/ - Scripts for running Fi-NeMo for identifying motif instances in the genome.

---

Analysis

These sets of scripts in the analysis/ directory generate all of the results presented in the manuscript.

---

1_ChromBPNet_PreprocessingAnalysis

These scripts process chromBPNet variant scoring outputs and compile annotation tables for downstream analyses.

Steps:

1. Extract outputs: Run 1_pull_scores.sh to extract relevant ChromBPNet outputs.
2. Analyze model performance:
   • 2a_model_performance.R evaluates performance metrics.
   • 2b_model_performance_plot.R identifies model outliers.
3. Annotate variants:
   • Use 3a_bed2vcf.Rare.CADD_VEP.R or 3b_bed2vcf.ASD.CADD_VEP.R to run CADD and VEP.
   • Process outputs with 3c_Process_CADD_VEP.R.
4. Merge results: Run 4_mergeData.R to integrate annotations, merge scores, and remove outliers.

---

2_Effects_Across_Contexts

These scripts correspond to the manuscript section "Variants effects are shaped by genomic context and TF binding".
They analyze:

• Genomic context – A variant’s proximity to transcribed regions.
• Cell-type specificity – How constrained or widespread variant effects are.
• Regulatory magnitude – The extent of chromatin accessibility and TF binding changes.

---

3_Application_to_GWAS_and_QTL_studies

These scripts support the manuscript sections:

• "Context-specific models reveal regulatory effects of fine-mapped eQTLs"
• "Pinpointing disease-relevant variants using cell-type-specific chromatin models"
• "Microglia-driven mechanisms of Alzheimer’s disease risk"

We use chromBPnet-predicted regulatory effects to identify candidate causal variants affecting gene regulation and disease risk.

---

4_Rare_vs_Common

These scripts correspond to:

• "Ultra-rare variants show larger and more shared regulatory effects than common variants"
• "Fetal neurons shape selective constraint in non-coding regions"

They compare rare and common variant effects to understand the selective pressures that influence allele frequency distributions across human populations.

---

5_FLARE

These scripts correspond to:

• "FLARE: a functional genomic model of constraint"
• "FLARE prioritizes de novo non-coding mutations in autism"

Since PhyloP scores are not context-specific, FLARE models the relationship between genomic context, regulatory effects, and evolutionary conservation within cell-type-specific contexts. FLARE:

1. Disentangles accessibility and regulatory effects from conservation.
2. Integrates multiple functional genomic features into a unified model.
3. Captures regulatory potential across multiple cell types.

We set up a FLARE repository for the FLARE method, which can be found by clicking here.

---

Cite

Marderstein^, Kundu^, et al. Mapping the regulatory effects of common and rare non-coding variants across cellular and developmental contexts in the brain and heart.