Computational Molecular Biology 2025, Vol.15, No.3, 151-159 http://bioscipublisher.com/index.php/cmb 152 approach is adopted, HPC remains the core pillar of large-scale genomic analysis and the underlying driving force for the true implementation of NGS variant detection (Ahmadet al., 2021). 2 Overview of the NGS Mutation Detection Workflow 2.1 Overall architecture of the mutation detection process In the matter of NGS mutation detection, each step is not actually independent; they are interlinked. Usually, researchers only obtain the original sequencing fragments. The first step is to conduct quality control. For example, use FastQC to check the data and remove those sequences of poor quality or contamination. Next, alignment software like BWA-MEM will be used to align these clean reads onto the reference genome and generate sorted BAM files. True mutation detection comes after this, using GATK HaplotypeCaller or DeepVariant (Poplin et al., 2018) to look for variations such as SNPS and indels. The detection is not over yet. The results still need to be screened, labeled, and finally the variations are displayed through visualization tools. The entire process seems linear, but in fact, each step may in turn affect the quality of the previous ones. 2.2 Common tools and algorithms When it comes to mutation detection software, GATK is a familiar face. It uses statistical models, such as logistic regression or hidden Markov models, to determine the differences in the sequence. However, AI has now stirred up this field. DeepVariant directly uses deep neural networks to view "stacked images" and identify variations, which is often more accurate than traditional algorithms. Tools such as FreeBayes, Strelka, and Platypus are also frequently used to detect different types of variations. In pursuit of speed, many commercial platforms have begun to implement multi-threading, GPU, and even FPGA acceleration. However, such optimizations are often costly, especially when dealing with big data. 2.3 Performance bottlenecks of traditional processes Performance issues are almost a persistent problem in all traditional processes. The computational burden is the heaviest during the alignment and mutation invocation phases. For example, when running GATK with a CPU, it may take several hours or even days for one sample. When there are more samples, the time doubles. Moreover, just the I/O operations are troublesome enough - frequent reads and writes during BAM file sorting and VCF generation will seriously slow down the speed (Costa et al., 2018). Although DeepVariant has a high accuracy rate, if it is not processed in parallel, the computing power consumption is also quite astonishing (Yang et al., 2020). To break the deadlock, relying solely on optimizing algorithms is not enough; support at the hardware level must also keep up. Only distributed computing based on Apache Spark, or directly on HPC clusters (Alganmi and Abusamra, 2023), is possible to truly alleviate these performance bottlenecks. 3 Fundamentals of High-Performance Computing in Bioinformatics 3.1 HPC architecture: clusters, grids, and cloud systems When it comes to high-performance computing, many people's first reaction is "clusters". But in fact, that's just one form. An HPC cluster is composed of many computing nodes connected through a high-speed network and is suitable for tasks that can be split into multiple parts and run simultaneously. Grid computing, on the other hand, takes a different path-it does not rely on a single institution or location, but connects computing resources scattered in different places for use. Then comes cloud computing, which is more flexible. Users don't have to worry about the underlying machines, and resources can be increased or decreased at any time. The overall goal is actually the same: to fully exploit the potential of computing power. For example, Munhoz et al. (2023) compared local HPC clusters and cloud clusters and found that CPU-intensive tasks performed similarly, but applications with frequent communication were still limited by network speed on the cloud. However, the idea of the cloud is quite different - it is more like providing a kind of "service", hiding the hardware behind the scenes and offering users an environment that can be extended and retracted at any time. In reality, these architectures are not clearly distinct. Many systems mix local clusters and cloud resources to strike a balance between cost and performance. 3.2 Parallel computing models: shared memory, distributed memory, and hybrid methods In parallel computing, no single model can solve all problems. The shared memory model is the most intuitive. Multiple threads share data on the same node, which has low overhead and high speed. However, once the task is
RkJQdWJsaXNoZXIy MjQ4ODYzNA==