S5: Percellome database provides genome wide mRNA expression profiles

Percellome database provides genome wide mRNA expression profiles derived from various organs of rodents (mice and rats) that are exposed to various chemical compounds

PRESENTING AUTHOR: 

Takeshi Hase

INSTITUTION / COMPANY : 

Medical Data Sciences Office, TMDU

POSITION: 

Associate Professor

ABSTRACT CONTENT / DETAILS: 

For each gene, the database provides a 3D-surface image that visualizes dose- and time-dependent changes in gene-expression induced by these chemicals. Researchers have visually inspected the images to identify informative genes that are keys to understand mechanisms of toxicity of a chemical compound [Kanno et al. The Journal of Toxicological Sciences 38.4 (2013): 643-645]. However, large volume and complexity in the 3D-surface image datasets (e.g., ~40,000 images for each chemical) make such visual-inspection process time-consuming and labor-intensive and requires a powerful computational tool to analyze them.

Here we have developed DTOX, a Deep neural network based computational tool to analyze omics data in TOXicology, that provides efficient analysis of 3D-surface images to identify informative genes. The computational tool is based on a state of the art architecture of deep convolutional neural network, ResNet [He et al. arXiv:1512.03385 (2015)], to build a classifier model to classify informative genes by using 3D-surface images as input of the model. We implemented the tool on computational server with Nvidia Teska K80 GPU and benchmarked the tool on a test image dataset generated through visual inspection by experts. The tool can analyze ~1,300 3D-surface images in five minutes and has successfully identified the informative genes with ~80% accuracy.

Our computational tool, DTOX, would provide efficient analysis of large volume of omics data in Percellome database to infer reactions induced by a variety of chemicals and their mechanisms of toxicity.