DeepChrome: Deep-learning for predicting gene expression from histone modifications about deep-review

[gwaybio]

Well written analysis with clear methodology to predict binary gene expression (low vs. high expression) using 5 chromatin marks in 46 different cell types.

Biology

Gene expression across ~20,000 genes binarized to -1 vs. 1 for low vs. high expression based on chromatin marks +/- 5,000 base pairs away from transcription start site (TSS). The five histone marks are: H3K4me3, H3K4me1, H3K36me3, H3K9me3, and H3K27me3.

Computation

Very well described methods showing exactly what the input data is (Figure 1), what the architecture of the deep model is, and what the outputs are (Figure 2). The model is a single layer convolution of per gene chromatin marks for 100 100bp bins with max pooling followed by a fully connected multilayer perceptron with a softmax discretized output layer. Very well done comparison between SVM and Random Forest methods that have been used previously for this task.

Why it should be included in our review

• While there are definitely concerns over the binary output (gene expression is way more complicated than that!) the authors demonstrate how a deep learning model outperforms state of the art simple machine learning algorithms on the task.
  • Proof of concept - next step is predicting actual expression (not just median...)
• The authors provide an interesting solution to the black box problem
  • Plot histograms of internal activation energies within the convolutional step to show how bins near the TSS are more activated.
    • Mirrors the distribution of chromatin marks observed in wet-lab studies
  • Visualize combinatorial interactions between marks at different bins that coordinate differently to produce high vs. low expression
    • Show coordination between specific histone marks that indicate gene repression
• Performed at the per gene level can give more specific predictions when there are altered histone peaks