基于k-mer分布和机器学习的资源节约分类法
Resource saving taxonomy classification with k-mer distributions and machine learning
摘要：
现代高通量测序技术如宏基因组测序产生了数百万个序列，这些序列必须基于它们的分类等级进行分类。现代方法要么对现有数据集（如MMseqs 2）进行局部对齐和比较，要么使用深度神经网络，如DeepMicrobes和BERTax中所做的那样。基于对齐的方法在运行时方面是昂贵的，特别是当数据库变得越来越大时。对于基于深度学习的方法，计算需要专用硬件，这会消耗大量能量。在本文中，我们建议使用从DNA获得的$k$ -mer分布作为特征，使用子空间$k$ -最近邻算法、神经网络或袋装决策树等机器学习方法对其分类起源进行分类。此外，我们提出了一种特征空间数据集平衡方法，该方法允许减少用于训练的数据集并提高分类器的性能。通过比较性能，时间和内存消耗，我们的方法的国家的最先进的算法（BERTax和MMseqs 2）使用几个数据集，我们表明，我们的方法提高了属级的分类，并取得了可比的结果为超王国和门级。链接：https://es-cloud.cs.uni-tuebingen.de/d/8e2ab8c3fdd444e1a135/? p=%2FTaxonomyClassification&mode=list
英文摘要：
Modern high throughput sequencing technologies like metagenomic sequencing generate millions of sequences which have to be classified based on their taxonomic rank. Modern approaches either apply local alignment and comparison to existing data sets like MMseqs2 or use deep neural networks as it is done in DeepMicrobes and BERTax. Alignment-based approaches are costly in terms of runtime, especially since databases get larger and larger. For the deep learning-based approaches, specialized hardware is necessary for a computation, which consumes large amounts of energy. In this paper, we propose to use $k$-mer distributions obtained from DNA as features to classify its taxonomic origin using machine learning approaches like the subspace $k$-nearest neighbors algorithm, neural networks or bagged decision trees. In addition, we propose a feature space data set balancing approach, which allows reducing the data set for training and improves the performance of the classifiers. By comparing performance, time, and memory consumption of our approach to those of state-of-the-art algorithms (BERTax and MMseqs2) using several datasets, we show that our approach improves the classification on the genus level and achieves comparable results for the superkingdom and phylum level. Link: https://es-cloud.cs.uni-tuebingen.de/d/8e2ab8c3fdd444e1a135/?p=%2FTaxonomyClassification&mode=list
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
