Imputing missing RNA-seq data from DNA methylation by using transfer learning based neural network

Abstract

Multi-omics integrative analysis can capture the associations of different omics and thus provides a comprehensive view of the complex mechanisms in cancers. However, it is common that one portion of samples miss one type of omics data due to various limitations in experiments, which can be an obstacle for downstream analysis where complete dataset is needed. Current imputation methods mainly focus on single cancer dataset, which are limited by their ability to capture information from large pan-cancer dataset. We present a novel transfer learning-based deep neural network to impute missing gene expression data from DNA methylation data, namely TDimpute. The pan-cancer dataset was utilized to train a general model for all cancers, which was then fine-tuned on each cancer dataset for the specific cancer. We compared our method to other state-of-the-art methods on 16 cancer datasets, and found that our method consistently outperforms other methods in terms of imputation error, methylation-expression correlations recovery, and downstream analysis including the identification of DNA methylation-driving genes and prognosis-related genes, clustering analysis, and survival analysis. The improvements are especially pronounced at high missing rates. Author summaryAs an epigenetic modification, DNA methylation plays an important role in regulating gene expression. However, due to limitations of sample availability and cost, some samples aren't measured with gene expression, which results in a reduced sample size for integrative analysis of DNA methylation and gene expression. The accuracy of traditional imputation methods are limited since they cannot effectively utilize the information from DNA methylation data and other relevant datasets. With the power of modeling nonlinear relationship, we used deep neural network to impute missing gene expression data using the nonlinear transformation from DNA methylation data to gene expression data. We also employed transfer learning to alleviate the data insufficiency in training the deep leaning model. In 16 cancer datasets from The Cancer Genome Atlas (TCGA), our method yields higher accuracy compared to other methods. More importantly, better performance of the downstream analysis on imputed gene expression datasets are achieved, which indicates the missing data imputed by our method are more biologically meaningful.

A Review of Integrative Imputation for Multi-Omics Datasets

“…It reliably imputes RNA-seq data by making use of external data from DNA methylation probe datasets. TDimpute (Zhou X. et al, 2019) is a deep neural network (DNN)-based transfer learning approach that imputes missing gene expression data using DNA methylation datasets. It employs a DNN model to recover missing gene expression data by constructing a non-linear mapping between DNA methylation data and gene expression data.…”

“…Since gene-trait associations are mostly detected in strongly relevant tissues, it is recommended to use trait-relevant tissues in order to boost the correlation between GReX of related tissues (Zhang et al, 2019). For the TDimpute model, it can be further improved by integrating prior biological knowledge regarding the gene-gene interaction factors in order to reduce the parameters of the DNN model (Zhou X. et al, 2019).…”

Imputing missing RNA-sequencing data from DNA methylation by using a transfer learning–based neural network

“…The datasets and pretrained pan-cancer models supporting the results of this article are available in the Synapse with ID: syn21438134 [ 44 ]. Snapshots of our code and data further supporting this work are openly available in the GigaScience repository, GigaDB [ 45 ].…”

Cell-specific imputation of drug connectivity mapping with incomplete data

“…Furthermore, to impute entirely missing expression profiles, one must incorporate additional domain information, such as data from nearby time points or functional relationships between genes' expression patterns. In another example, transfer learning has been used to impute entire bulk RNA-sequencing profiles when methylation profiles for the same samples are available (Zhou et al, 2020). Here, we use expression profiles of related drugs and cells.…”