Torch-eCpG: A fast and scalable eQTM mapper for thousands of molecular phenotypes with graphical processing units

BACKGROUND

Gene expression is regulated, in part, by epigenetic mechanisms. A major unanswered question in genomics research is the functional contribution of epigenetic variation on gene expression.[1, 2] One method to evaluate for the potential functional effect of a methylation variation is to test for an association between levels of methylation and gene expression from the same samples. These expression-associated quantitative trait methylation (eQTM) loci may contribute to the regulation of gene expression (also called expression associated CpG loci, eCpG). These associations may be local (e.g., methylation located in the promoter region of a gene) or remote (e.g. methylation loci in a distant enhancer regions of a gene or on a different chromosome). There is growing interest in the integration of these data modalities and evaluating for eQTMs. For example, in terms of clinical research, recent studies have identified eQTMs from a variety of tissue types and outcomes.[3–7]

Recent advances in high throughput molecular methods allow for the collection of complementary methylation and gene expression data from the same sample in large numbers. Although an increasing number of recent studies have provided eQTM datasets,[3, 8] there are no open-source tools currently available to investigators to implement these analyses on their own.

Given the increase in the availability of complementary datasets and the biological utility of identifying eQTMs, analytic tools must be made available, free to use, and able to scale to handle thousands to millions of samples (e.g., patients or single cells). Current methylation array datasets provide hundreds of thousands of loci (e.g., Infinium MethylationEPIC, Illumina, San Diego, CA) and RNA-sequencing and microarray methods provide expression levels for tens of thousands of genes. An exhaustive evaluation of these datasets would result in tens of billions of tests for hundreds or thousands of samples, quickly outreaching the computing capacity of a desktop computer and requiring larger workstations, clusters, or cloud computing.[9] Future datasets will likely include more loci for evaluation and larger sample sizes. This need for computational resources will also require improvements in efficiency. Graphical processing units (GPUs) have provided major improvements in computation efficiency (i.e., runtime) for many bioinformatic software tools.[10] Readily available open source libraries implement numerous general-purpose methods and mathematical primitives that allow for major improvements in computational efficiency at relatively lower costs as compared to CPUs.[11]

Given the analytic utility of evaluating for eQTMs to identify relationships between gene expression and epigenetic changes, a lack of an available open-source tool to implement an eQTM analysis, and the performance benefits of utilizing a GPU, the objectives of this project were to develop an open source, general-use tool for eQTM mapping and evaluate for performance increases of a GPU implementation. Here, we present the Torch-eCpG tool (tecpg).

IMPLEMENTATION

RESULTS AND DISCUSSION

To provide an open-source tool for eQTM mapping, we developed the Torch-eCpG software package. To evaluate the reproducibility of the linear regression analysis, we compared our results with those implemented in the lm() function in the stats package in R. As shown in Figure 1, our implementation of the linear regression demonstrates high reproducibility.

For tecpg benchmarking, we evaluated eQTM mapping using the CPU and GPU implementation in tecgp for 300 patients from the GTP patient dataset (422,442 methylation loci and 17,653 genes). Across the mapping modalities, the GPU outperformed the CPU analysis by up to 18x. Our implementation of the cis-eCpG mapping was 1.4x faster on the GPU than that of the CPU (Figure 2A). For distal-eCpG mapping, our implementation was 5x faster on the GPU. Finally, for trans-eCpG mapping, our implementation was 18x faster on the GPU. In terms of tecpg using additional cores, we found that major incremental improvements were realized by increasing the CPU core count up to 8, after which the gains were minimal (Figure 3). Although the CPU performance did improve with additional cores, the GPU implementation was still 2x faster than the 24-core CPU implementation.

We found tecpg scales linearly across a wide range of methylation loci for a reasonably large sample size (n=1000 patients) (Figure 4). In addition, the total time to evaluate 1000 patients for whole transcriptome (2×10⁴ genes) and whole methylome array data (8.5×10⁵) was <15 hours. The short time needed to evaluate a dataset sized to the largest currently available methylation array (i.e., Infinium MethylationEPIC) highlights the utility of this tool to evaluate eQTM mapping of dataset of realistic size. The linear scaling suggests the tool will be useful for larger datasets in the future (e.g., >10,000 patients).

CONCLUSIONS

Torch-eCpG is the first freely available open-source tool to perform eQTM mapping. It provides a scalable and high-performance implementation that supports GPU enabled systems. By reducing computing time the tool offers cost-savings on shared systems (e.g., clusters) or cloud-based computing resources that charge by units of time. This tool allows for individual research labs with limited computational resources to perform analyses on affordable computer equipment or cloud-based virtual machines.

Availability of data and materials:

The datasets analyzed in this study are publicly available in the Gene Expression Omnibus repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE72680 and https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc= GSE58137)