IMPACT (Inference and Modeling of Phenotype-related ACtive Transcription)

A data aggregation strategy to model specific transcriptional processes based on transcription factor and cell-state specificity.

Publication in American Journal of Human Genetics: IMPACT: Genomic Annotation of Cell-State-Specific Regulatory Elements Inferred from the Epigenome of Bound Transcription Factors

• AJHG

Publication in Nature Genetics: Improving the trans-ancestry portability of polygenic risk scores by prioritizing variants in predicted cell-type-specific regulatory elements

• Nature Genetics

Using IMPACT

Dependencies

1. A motif finding algorithm, such as HOMER as used in our implementation of IMPACT.
2. R version 3.2.2 or higher
3. R package GenomicRanges

Data

Transcription factor binding motif PWMs (position weight matrices) and genome-wide positions of motifs

Motifs/ 

Epigenomic and sequence features used in the IMPACT framework to predict cell-state-specific regulatory elements.

Features/ 

Implementing IMPACT

1. Create positive (regulatory) and negative (non-regulatory) training sets:

Training/ Pipeline.sh "TF" "cell type" "number input (ChIP-seq) files"

2. Run model on a per-TF basis, to get betas on epigenomic and sequence features:

GenomeWide/IMPACT_modelfit/ENet_fit7_TFArgument.R "TF"

3. Predict cell-state-specific regulatory elements genomewide based on model fits:

GenomeWide/GenomeTracks/Predictvalues_TFarg.R "TF"

4. Then compute minimum and maximum values per chromosome:

GenomeWide/GenomeTracks/MinMaxValues_TFarg.R "TF"

5. Finally, scale logistic regression output to be between 0 and 1:

GenomeWide/GenomeTracks/Scale_TFarg.R "TF"

6. Create IMPACT annotations formatted to be run through S-LDSC to compute heritability estimates:

GenomeWide/ldsc/CreateAnnotations_TFarg.R "TF"

7. Run S-LDSC in European and East Asian populations:

sLDSC/ 

Assessing IMPACT performance:

Compute AUC and sensitivity of IMPACT to predict TF binding of a motif:

Performance/ 

Compute AUC and sensitivity after parameter variation and categorical feature knock outs:

KnockOuts/ 

Find loci with predicted IMPACT regulatory elements that do not have evidence of TF binding:

CandidateLoci/ 

Compare IMPACT predictions at TF target genes, defined by ChIP-seq, to TF non-target genes.

TargetvNonTargets/