HDMA

Overview of the steps involved in training ChromBPNet models on the atlas and analyzing outputs.

We show how to load and use trained ChromBPNet models in the tutorial at code/05-misc/04-ChromBPNet_use_cases.ipnyb (html).

Tools used

• ChromBPNet: https://github.com/kundajelab/chrombpnet
• TF-MoDISco (tfmodiscolite implementation): https://github.com/jmschrei/tfmodisco-lite
• gimmemotifs: https://gimmemotifs.readthedocs.io/en/master/
• Fi-NeMo: https://github.com/austintwang/finemo_gpu
• tangermeme: https://github.com/jmschrei/tangermeme

chrombpnet_utils

Python utilities for downstream ChromBPNet related analysis, plotting, etc.

tangermeme_utils

Python utilities for downstream ChromBPNet related analysis, particularly for interfacing with the tangermeme package. These are mainly used for the in silico experimentation.

0. Preparing inputs, 00-inputs

These scripts pertain to generating the necessary inputs for training ChromBPNet models.

• 01: get fragments per cluster per sample from ArchR projects
• 02: process fragments for calling peaks (generating pseudoreplicates)
• 03: group fragments across samples per cell type, and sort fragments
• 04: call peaks with macs2 on pseudoreplicates
• 04b: call peaks on the fragments of each cluster (no pseudoreps), to generate p-value and fold-change signal bigwig tracks
• 05: format peaks for ChromBPNet
• 06: make bigwigs
• 07: make chromosome splits for test/train/validation folds
• 08: get GC-matched negative peaks
• 09: tabix-index peaks for viewing in browser
• 10: randomly subset backgrounds to 100 regions per cell type to use for downstream analysis

1. Training models, 01-train_models

These scripts run the ChromBPNet workflow, which involves training a bias model, using the bias model to train chromatin accessibility models in each cell type, calculate contribution scores, generate model predictions, and perform motif discovery.

• 01: train bias models
• 02: train ChromBPNet models with a given bias model, initial interpretation with 30k subsample peaks
• 03: perform model QC and make figures with performance metrics
• 04: get contribution scores (aka DeepLIFT or DeepSHAP scores) as h5 files per fold per cell type
• 05: average contribs across folds, separately for profile and counts
• 06: deep TF-MoDISco run with 1M seqlets on averaged contribs
• 07: generate contribution score bigwigs, for viewing in browser
• 08: predict accessibility in peak regions across folds, and average the predictions (bias-corrected) to generate bw
• 09: prepare supplementary table with ChromBPNet performance and QC metrics

2. Assembly of motif compendium, 02-compendium

These scripts pertain to assembling a motif compendium from across the atlas, calling instances of each motif in each cell type, and then annotating those instances.

• 01: convert MoDISco CWMs to PFMs to prepare the input for gimme cluster: this is run per organ, and motifs from all cell types (models) within an organ are concatenated. Done separately for positive and negative patterns.
• 02: for each organ, run gimme cluster on PFMs of motifs from all cell types in that organ, to get broad clusterings. Done separately for positive and negative patterns.
• 02b: run gimme cluster again, using as input all the clusters from the per organ runs of gimme cluster in Step 02. Dine separately for positive and negative patterns. This produces “superclusters” of patterns from across cell types.
• 03: merge similar MoDISco patterns within each supercluster together, using the modiscolite.SimilarPatternCollapser functionality. This produces one MoDISco h5 object per supercluster, containing merged, non-redundant patterns.
• 04: combine MoDISco h5 objects from the merging step into one h5 object, match to known patterns, assign unique index, and prepare as MEME format
• 05: call hits in each cell type, using the unified motifs
• 06: generate HTML reports to assess the quality of hit calling with unified motif set, filter & label hits using manual annotation
• 06b: filter out low QC hits, reconcile overlapping hits, and get genomic annotations
• 07: run NucleoATAC per cell type

3. Analysis of TF binding site syntax, 03-syntax

• 00: compute the mean and summed importance score for all trimmed CWMs
• 01: global overview of de novo motifs and predictive instances
• 02: visualization of genomic tracks at example loci
• 03: visualization of motif instances (hits) for a few examples
• 04: in silico analysis of cooperativity
  • 04a: assemble the set of composites to test
  • 04b: run the in silico analysis
  • 04c: summarize the results and generate plots
  • 04d: make animations showing hard and soft synergy
• 05: in silico analysis of cell type-specific cooperativity
  • 05a: run the analysis
  • 05b: visualize results
• 06: in silico ablations of negative motifs
• 07: compute motifs with enriched co-occurrence
• 08: investigate overlaps in peaks and motif instance calls across clusters

NOTE: variant scoring using ChromBPNet models is done in the code/06-variants directory.

This site is open source. Improve this page.