JMIR Publications

ABSTRACT

Background:

Complex multiomics data requires interpretation on different levels of molecular biology processes. However, genomics and transcriptomics methods are more sensitive than proteomic methods because of PCR reaction and ability to multiply molecules to level enough for a successful detection. This forms the gap between thousands of genes with known RNA expression and hundreds of detected proteins.

Objective:

The solution was designed as an extrapolating tool for online public web-usage, because of that only light-weighted deep neural architectures with fast inference were used.

Methods:

To predict protein abundance by known RNA level in a sample we have used a multi-channels input deep convolutional neural network to store label-free LC-MS/MS data, amino acids sequence and database annotations in one tensor. Gene mappings databases with many-to-many relationships were used to construct input feature maps.

Results:

We have considered two datasets with normal human tissues and human tumor cell lines. Healthy human tissues dataset Tissue29 contained 12417 proteins for 29 samples which leads to 251349 experiments. Human tumor cell lines dataset NCI60 contained 251706 experiments on 46 samples for 7812 proteins. The validation set was once created by randomly selected 20% of proteins in both of used datasets. Validation results for the healthy human tissues dataset showed 0.58 R2 coefficient of determination and 0.71 Spearman correlation for all samples and up to 0.67 R2 per tissue. For the tumor cell lines dataset 0.54 R2 coefficient of determination and 0.71 Spearman correlation for all validation samples and up to 0.62 R2 per cell line.

Conclusions:

The solution allows a researcher to use any annotation collections and modifiable amino acids sequence to extrapolate protein abundances with known RNA quantity in the sample. A “one model – many datasets” pipeline with comparable metrics was implemented.