Accelerating Genomics Pipelines Using Intel’s Open Omics Acceleration Framework on AWS

This post was contributed by Sanchit Misra, PhD, Vasimuddin Md., PhD,  Saurabh Kalikar, PhD, and Narendra Chaudhary, PhD, research scientists at Intel Labs.

Introduction

We are in the epoch of digital biology, that is fueled by the convergence of three revolutions:

• • the measurement of biological systems at high resolution,
  • novel data science (AI and data management) techniques that can be applied on this data, and
  • widespread use of the massive public data repositories, large collaborative projects, and consortia, which in turn promote the use of cloud due to easy data access.

Genomics is a primary example of this trend, where high-throughput next-generation sequencing (NGS) devices are being used to sequence DNA, mRNA, regulatory regions, the gut microbiome, etc. Computational workflows are also being developed, standardized rapidly, and scaled by running on the cloud. With the enormous quantities of genomic data being collected, processing times are often in the order of billions of core hours, and the cost of processing increase commensurately. As a result, customers are looking for optimized tools and systems that incur the shortest runtimes and lowest costs.

Intel’s Open Omics Acceleration Framework (in short, Open Omics) is an open-sourced high throughput framework for accelerating omics pipelines. Intel is developing this framework with the following characteristics:

• Community driven: Open Omics framework is being built based on extensive discussions with thought leaders in digital biology to understand the requirements of the user community. Moreover, Intel is building the framework with a modular design. This enables the developer community to use efficient modules to achieve faster performance for existing and new software tools in a productive manner.
• Open-sourced: so that anyone can customize it for variations in use-cases.
• Hardware accelerated: uses the underlying hardware efficiently to reduce cloud costs.
• Supports full application stack: The application layer supports a wide range of applications in genomics, single cell analysis, and drug discovery. The middleware layer has scalable and efficient implementations of key building blocks, such as data management and key compute motifs. All of this is optimized for the processor, memory, storage, and networking.

In this blog, we showcase the first version of Open Omics and benchmark three applications that are used in processing NGS data – sequence alignment tools BWA-MEM, minimap2, and single cell ATAC-Seq on Xeon-based Amazon Elastic Compute Cloud (Amazon EC2) Instances.

Applications benchmarked for this blog

BWA-MEM and Minimap2 are popular software tools for aligning short reads and long reads to a reference sequence. The Open Omics version of BWA-MEM is called BWA-MEM2 and that of minimap2 is called mm2-fast. These are efficient architecture-aware implementations of original tools that were built in collaboration with Prof. Heng Li. They are both drop-in replacements that significantly reduce runtime and cloud costs while maintaining command line interface and output identical to original tools [2,3] and have been open sourced. Open Omics BWA-MEM has been used by more than 40 peer-reviewed genomics studies already, including research on Covid-19 [5,6], gut microbiome [7], and cancer [8].

ATAC-seq assays are used for identifying accessible chromatin regions in the DNA. ATACWorks [1] is a toolkit that is used to de-noise and identify accessible chromatin regions, and it uses deep learning on 1D data. The Open Omics version of ATACWorks builds an efficient 1D dilated convolution layer and demonstrates reduced precision (BFloat16) training to achieve significant performance gain without any loss of accuracy [4].

Benchmarking the Open Omics Acceleration framework on AWS

Amazon EC2 Instances used in this benchmarking

The four types of Amazon EC2 Instances used in this benchmarking study are detailed in the following table.

Prerequisites

An AWS account with permissions to provision Amazon S3 buckets for input and output data storage, as well as sufficient permissions/limits to provision Amazon EC2 C5, M5, C6i, and M6i Instances.

How to benchmark Open Omics Acceleration Framework on AWS

The configuration details and steps used for benchmarking baseline and Open Omics versions of all three applications on EC2 Instances are detailed at IntelLab’s GitHub page. Typical process involves launching the corresponding EC2 Instances, connecting to the instances, installing the software, downloading the datasets, and executing the baseline and Open Omics versions. In the following subsections, we report results for the three applications on on-demand instances with dedicated tenancy. Compared to on-demand costs shown, the EC2 Spot Instances can provide nearly 75% cost savings.

Benchmarking Results: BWA-MEM

We used m5.12xlarge and m6i.16xlarge instances, with 48 and 64 threads (one thread per vCPU), respectively. The m-instance types were used because they provide 4 GB memory per vCPU that is required to run Open Omics BWA-MEM.

Figure 1 shows that on the same instance type (m5), Open Omics BWA-MEM achieves 1.8-2.3x speedup over the baseline BWA-MEM. Using the m6i instance type gives further performance gain, achieving 2.6-3.5x over baseline BWA-MEM on m5. The performance reported here of Open Omics BWA-MEM on m6i instance is ~1.7x faster than best performance on latest GPU. Please refer to this blog post and this video for a comparison.

The speedups are lower for ERR194147 dataset because it has reads of length ~100, providing less scope of parallelization. In comparison, the other two datasets that have reads of length ~150. A majority of the modern short read sequencers have read lengths ≥ 150 and they are expected to grow further. Therefore, we can expect higher speedups in the future.

Figure 1: Comparison of execution time of baseline BWA-MEM and Open Omics BWA-MEM on m5 and m6i instances for two different use cases – paired end and single end – for the three datasets used. The vertical bars show the execution time, while the line graph shows the speedup compared to baseline BWA-MEM on m5.

Figure 2 shows the price-performance chart for BWA-MEM. It demonstrates that Open Omics BWA-MEM achieves significant cost cuttings compared to baseline BWA-MEM. Moreover, the m6i instances not only provide faster performance compared to m5 instances, they also incur lower costs.

Figure 2: Comparison of On-Demand Instance costs per sample processed of baseline BWA-MEM and Open Omics BWA-MEM on m5 and m6i instances for two different use cases – paired end and single end – for the three datasets used.

Benchmarking Results: Minimap2

For this experiment, we used c5.12xlarge and c6i.16xlarge instances, using 48 and 64 threads (one thread per vCPU), respectively. Figure 3 shows that on the same instance type (c5), Open Omics minimap2 achieves 1.5-1.9x speedup over the baseline minimap2. Using the c6i instance type, gives further performance gain achieving 2-2.4x over baseline minimap2 on c5.

Figure 3: Comparison of execution time of baseline minimap2 and Open Omics minimap2 on c5 and c6i instances for mapping reads obtained from different sequencing technologies – Oxford Nanopore Technologies (ONT), Pacific Biosciences CLR (CLR), Pacific Biosciences HiFi (HiFi) – to the reference human genome. The vertical bars show the execution time, while the line graph shows the speedup compared to baseline minimap2 on c5.

The price-performance chart shown in Figure 4 clearly demonstrates that Open Omics minimap2 costs nearly the same on the c6i and c5 instances, while achieving significant cost savings over the baseline minimap2 running on c5.

Figure 4: Comparison of On-Demand Instance costs per sample of baseline minimap2 and Open Omics minimap2 on c5 and c6i instances for the three datasets used.

Benchmarking Results: ATAC-Seq data analysis

Figure 5 compares the execution time of the baseline and Open Omics versions of ATACWorks on c5 and c6i instances. The baseline version of ATACWorks is created by replacing the CUDA based deep learning modules with Intel^® oneDNN library. Open Omics version uses Intel’s new optimized implementation of the 1D convolutions. The chart shows that on the same instance type (c5), Open Omics achieves 9.3x speedup over the baseline. Using the single c6i instances, gives further performance gain achieving ~12x over baseline minimap2 on c5. Using multiple c6i instances, Open Omics demonstrates excellent scaling to achieve up to ~145x speedup using 16 instances. The performance reported here of Open Omics on 16 c6i instances is ~1.9x faster than best published GPU performance [1] as demonstrated in [4].

Figure 5: Comparison of execution time of baseline and Open Omics on c5 and c6i instances. The execution time per epoch is shown above the corresponding bar. The line graph shows the speedup compared to the baseline on c5.

Figure 6 shows the price-performance comparison of baseline and Open Omics version of ATACWorks on c5 and c6i instances. Single c6i instance with Open Omics is nearly 9-fold cheaper than single c5 instance with baseline and is the cheapest option. However, Open Omics on 16 c6i instances provides dramatically higher performance at just fractionally higher cost.

Figure 6: Comparison of on-demand instance costs per epoch of baseline and Open Omics on c5 and c6i instances.

Cleanup

Terminate all EC2 instances used to run benchmarks to avoid incurring additional charges.

Conclusion

To support the digital biology revolution by accelerating the data science experiments, Intel has envisioned the Open Omics framework. The first version of Open Omics provides significant performance gains on key genomics applications, resulting in not only reduction in computation time, but also in costs. We showed that the optimized Open Omics tools running on m6i instances showed speedups of up to 3.5x, 2.4x, and 145x for BWA-MEM, minimap2, and ATAC-seq, respectively.

To learn more about Open Omics, please refer to the following research papers – IPDPS’19, Nature’22, BioRxiv’22, and the following GitHub repository – Open Omics Acceleration Framework.

The content and opinions in this blog are those of the third-party author and AWS is not responsible for the content or accuracy of this blog.

References

1. Avantika Lal et. al. “AtacWorks: A deep convolutional neural network toolkit for epigenomics”. bioRxiv 829481.
2. Vasimuddin Md, Sanchit Misra, Heng Li, Srinivas Aluru. “Efficient Architecture-Aware Acceleration of BWA-MEM for Multicore Systems”.IEEE Parallel and Distributed Processing Symposium (IPDPS), 2019.
3. Saurabh Kalikar, Chirag Jain, Vasimuddin Md, Sanchit Misra. “Accelerating long-read analysis on modern CPUs“. Nature Computational Science volume 2, pages78–83 (2022).
4. Narendra Chaudhary, Sanchit Misra, Dhiraj Kalamkar, Alexander Heinecke, Evangelos Georganas, Barukh Ziv, Menachem Adelman, Bharat Kaul. “Accelerating Deep Learning based Identification of Chromatin Accessibility from noisy ATAC-seq Data”. To appear at 21st IEEE International Workshop on High Performance Computational Biology (HiCOMB), 2022.
5. Trevor Bedford et. al. “Cryptic transmission of SARS-CoV-2 in Washington state”. Science Translational Medicine, 13, 595, (2021).
6. Haogao Gu et. al. “Genomic epidemiology of SARS-CoV-2 under an elimination strategy in Hong Kong”. Nature Communications volume 13, Article number: 736 (2022).
7. Yuqing Feng et. al. “Metagenome-assembled genomes and gene catalog from the chicken gut microbiome aid in deciphering antibiotic resistomes”. Nature Communications Biology volume 4, Article number: 1305 (2021).
8. Chao Wang, Chun Liang. “The insertion and dysregulation of transposable elements in osteosarcoma and their association with patient event-free survival”. Nature Scientific Reports volume 12, Article number: 377 (2022).