ABSTRACT
Background The distribution and composition of cis-regulatory modules (e.g. transcription factor binding site (TFBS) clusters) in promoters substantially determine gene expression patterns and TF targets, whose expression levels are significantly regulated by TF binding. TF knockdown experiments have revealed correlations between TF binding profiles and gene expression levels. We present a general framework capable of predicting genes with similar tissue-wide expression patterns from activated or repressed TF targets using machine learning to combine TF binding and epigenetic features.
Methods Genes with correlated expression patterns across 53 tissues were identified according to their Bray-Curtis similarity. DNase I HyperSensitive region (DHS) -accessible promoter intervals of direct TF target genes were scanned with previously derived information theory-based position weight matrices (iPWMs) of 82 TFs. Features from information density-based TFBS clusters were used to predict target genes with machine learning classifiers. The accuracy, specificity and sensitivity of the classifiers were determined for different feature sets. Mutations in TFBSs were also introduced to examine their impact on cluster densities and the regulatory states of predicted target genes.
Results We initially chose the glucocorticoid receptor gene (NR3C1), whose regulation has been extensively studied, to test this approach. SLC25A32 and TANK were found to exhibit the most similar expression patterns to this gene across 53 tissues. Prediction of other genes with similar expression profiles was significantly improved by eliminating inaccessible promoter intervals based on DHSs. A Random Forest classifier exhibited the best performance in detecting such coordinately regulated genes (accuracy was 0.972 for training, 0.976 for testing). Target gene prediction was confirmed using CRISPR knockdown data of TFs, which was more accurate than siRNA inactivation. Mutation analyses of TFBSs also revealed that one or more information-dense TFBS clusters in promoters are required for accurate target gene prediction.
Conclusions Machine learning based on TFBS information density, organization, and chromatin accessibility accurately identifies gene targets with comparable tissue-wide expression patterns. Multiple, information-dense TFBS clusters in promoters appear to protect promoters from the effects of deleterious binding site mutations in a single TFBS that would effectively alter the expression state of these genes.
LIST OF ABBREVIATIONS
- TF
- transcription factor
- TFBS
- transcription factor binding site
- CRM
- cis-regulatory modules
- iPWM
- information theory-based position weight matrix
- IDBC
- information density-based clustering
- ChIP-seq
- chromatin immunoprecipitation with massively parallel DNA sequencing
- HM
- histone modification
- mRNA
- messenger RNA
- siRNA
- small interfering RNA
- CRISPR
- clustered regularly interspaced short palindromic repeats
- DHS
- deoxyribonuclease I hypersensitive region
- TP
- true positive
- TN
- true negative
- RPKM
- reads per kilobase of transcript per million mapped reads
- GTEx
- genotype-tissue expression
- ENCODE
- encyclopedia of DNA elements
- TSS
- transcription start site
- SVM
- support vector machine
- RBF
- radial basis function
- PC
- absolute Pearson correlation coefficient
- SC
- the absolute Spearman correlation coefficient
- CARS
- the absolute combined angle ratio statistic
- SNP
- single nucleotide polymorphism.