PT  - JOURNAL ARTICLE
AU  - Ye Yuan
AU  - Ziv Bar-Joseph
TI  - CNNC: Convolutional Neural Networks for Co-Expression Analysis
AID  - 10.1101/365007
DP  - 2018 Jan 01
TA  - bioRxiv
PG  - 365007
4099  - http://biorxiv.org/content/early/2018/07/08/365007.short
4100  - http://biorxiv.org/content/early/2018/07/08/365007.full
AB  - Co-expression analysis has been extensively used in genomics studies and tools for over two decades. To date, most methods for such analysis are unsupervised and symmetric. Such methods cannot infer causality and are prone to both overfitting and false negatives resulting from differences between cells in bulk studies. Here we present a new, supervised method based on convolutional neural networks (CNNs) for co-expression analysis. We use a normalized histogram image of gene pair co-expression as the input to the CNN. Testing our method on several co-expression prediction tasks we show that it outperforms prior methods and that scRNA-Seq data leads to more accurate results when compared to bulk data. The method can be directly extended to integrate sequence and epigenetic data and to infer causal relationships.Supporting website with software and data: https://github.com/xiaoyeye/CNNC.