Minisymposium

Application proxies in high-performance computing play an important role for software/hardware co-design. To broaden the types of computation available for co-design, we are developing a suite of proxy apps based on MetaHipMer2 (mhm2), a DOE-developed, scalable, de novo metagenome assembler. MetaHipMer2 is implemented in C++, and offloads several routines to GPU (K-mer Analysis, Alignment, Local Assembly). It has been used to assemble large (> 50 Terabase), complex metagenomes on exascale-class machines (e.g., Summit). Our first proxy, "mhm2-kmer-analysis," focuses on the expensive K-mer Analysis step, which we have implemented in the Kokkos performance portability framework. In this this talk, we give an overview of mhm2, its execution phases, and their correlation to common “big data” computational patterns. Further, we make the case for modernization of codes via vendor-independent portability frameworks, such as Kokkos, and discuss our porting experience, including vignettes on expressing common CUDA idioms in Kokkos. Finally, we give an outlook on future software and hardware trends that will likely impact computational biology. Sandia National Laboratories is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

Pangenomics is an emerging field that is allowing us to accurately and comprehensively study the within-species genetic diversity and its relationship to physical traits (phenotypes) by using a collection of genomes of a species instead of a single reference genome. Future pangenomics applications would require analyzing ultra-large and ever-growing collections of genomes. While existing pangenome data formats can represent the genetic variation in a collection of genomes, they do not store their shared evolutionary and mutational histories and are also unlikely to keep up with the speed and volume of genome sequencing data. In this talk, I will present ongoing work from my lab on a novel pangenomic data representation that achieves significant improvements in memory efficiency and the representative power of pangenomes. I will then discuss how we are leveraging GPUs and HPC systems to construct massive pangenomes consisting of millions of sequences. While the focus will be on microbial genomes, I will also discuss how these approaches can be extended to more complex genomes.

Analyzing genomic data provides critical insights for understanding and treating diseases, outbreak tracing, evolutionary studies, agriculture, and many other areas of the life sciences and personalized medicine. Modern genome sequencing devices can rapidly generate large amounts of genomic data at a low cost. However, genome analysis is bottlenecked by the computational and data movement overheads of existing systems and algorithms, causing significant limitations in terms of speed, accuracy, application scope, and energy efficiency of the analysis. In this talk, we will focus on substantially improving the speed and energy efficiency of a computationally costly machine learning (ML) technique used in many important genomics applications. We will introduce ApHMM, which resolves significant inefficiencies that make an expectation-maximization technique costly for profile Hidden Markov Models (pHMMs) on general-purpose processors. ApHMM achieves this by effectively co-designing both hardware and algorithm. As a result, ApHMM provides substantial improvements in performance (up to two orders of magnitude) and energy efficiency (up to three orders of magnitude) compared to CPUs and GPUs.

The enormous data growth continuously shifts the life sciences from model-driven towards data-driven science. The need for efficient processing has led to the adoption of massively parallel accelerators such as GPUs. As a consequence, genomics and proteomics method development nowadays often heavily depends on the effective use of these powerful technologies. Furthermore, progress in both computational techniques and architectures continues to be highly dynamic including novel deep neural network models and AI accelerators. For example, contemporary groundbreaking AI-tools like AlphaFold can generate highly accurate 3D protein structure predictions. In this talk, I present two novel tools for accelerating large-scale protein homology search on modern GPU systems: CUDASW++4.0 and MMseqs2-GPU, which advance the state-of-the-art in this area. For example, MMSeqs2-GPU can be used to significantly accelerate the computation of multiple sequence alignments in the ColabFold server for protein structure prediction, which is one of the most frequently used bioinformatics tools worldwide.