DeepDiffChrome

DeepDiff: Deep-learning for predicting Differential gene expression from histone modifications

@article{ArDeepDiff18,
author = {Sekhon, Arshdeep and Singh, Ritambhara and Qi, Yanjun},
title = {DeepDiff: DEEP-learning for predicting DIFFerential gene expression from histone modifications},
journal = {Bioinformatics},
volume = {34},
number = {17},
pages = {i891-i900},
year = {2018},
doi = {10.1093/bioinformatics/bty612},
URL = {http://dx.doi.org/10.1093/bioinformatics/bty612},
eprint = {/oup/backfile/content_public/journal/bioinformatics/34/17/10.1093_bioinformatics_bty612/2/bty612.pdf}
}

Feature Generation

We used the five core histone modification (listed in the paper) read counts from REMC database as input matrix. We downloaded the we used processed data files from REMC database and used bedtools to put it in the format that has been shared. We converted 'tagalign.gz' format to 'bam' by using the command:

gunzip .tagAlign.gz

bedtools bedtobam -i .tagAlign -g hg19chrom.sizes > .bam

Next, we used "bedtools multicov" to get the read counts. Bins of length 100 base-pairs (bp) are selected from regions (+/- 20000 bp) flanking the transcription start site (TSS) of each gene. The signal value of all five selected histone modifications from REMC in bins forms input matrix X, while log fold change in gene expression is the output y.

For gene expression, we used the read count files available in REMC database and added 1 to all counts.

We divided the genes into 3 separate sets for training(10,000 genes), validation(2360 genes) and testing(6100 genes).

We performed training and validation on the first 2 sets and then reported Pearson Correlation Coefficient(PCC) scores of best performing epoch model for the third test data set.

Sample dataset has been provided inside "data/" folder and all datasets used in DeepDiffChrome are provided in "data/ProcessedData". For two cell types "Cell1" and "Cell2" under consideration, the expression value is in Cell1.expr.csv and Cell2.expr.csv for all genes. The first column is geneID, and the second column is expression value. The train, valid and test set inputs are in Cell*.train.csv, Cell*.valid.csv, and Cell*.test.csv. The columns represent: geneID_window,H3K4me1 count,H3K4me3 count,H3K9me3 count,H3K27me3 count,H3K36me3 count.

Training Model

To train, validate and test the model for celltypes "Cell1" and "Cell2":

      python train.py --cell_1=Cell1 --cell_2=Cell2 --model_name=raw_d --epochs=120 --lr=0.0001 --data_root=data/ --save_root=Results/

Other Options

1. To specify DeepDiff variation:
   --model_name=
         raw_d: difference of HMs
         raw_c: concatenation of HMs
         raw: raw features- difference and concatenation of HMs
         raw_aux: raw features and auxiliary Cell type specific prediction features
         aux: auxiliary Cell type specific prediction features
         aux_siamese: auxiliary Cell type specific prediction features with siamese auxiliary
         raw_aux_siamese: raw features and auxiliary Cell type specific prediction features with siamese auxiliary

2. To save attention maps:
         use option --save_attention_maps : saves Level II attention values in .txt file

3. To change rnn size:
         --bin_rnn_size=32

Testing

To only test on a saved model:
python train.py --test_on_saved_model --model_name=raw_d --data_root=data/ --save_root=Results/

DeepDiffChrome was adapted from our previous work AttentiveChrome:

https://github.com/QData/AttentiveChrome

AttentiveChrome is a unified architecture to model and to interpret dependencies among chromatin factors for controlling gene regulation. AttentiveChrome uses a hierarchy of multiple Long short-term memory (LSTM) modules to encode the input signals and to model how various chromatin marks cooperate automatically. AttentiveChrome trains two levels of attention jointly with the target prediction, enabling it to attend differentially to relevant marks and to locate important positions per mark. We evaluate the model across 56 different cell types (tasks) in human. Not only is the proposed architecture more accurate, but its attention scores also provide a better interpretation than state-of-the-art feature visualization methods such as saliency map.

Meanwhile, here are some links for general data processing tools/guidance on ChIP-seq data:

https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1003326 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5389943/ https://bedtools.readthedocs.io/en/latest/