<p>Ablation experiments in Dataset 1 and Dataset 2.</p>

Phages (or bacteriophages) play a critical role in microbial communities, and accurately predicting the hosts of phages is essential for understanding the dynamics of these viruses and their impact on bacterial populations. In the prediction of classification of phage hosts, feature extraction is a critical step that directly affects the accuracy of the predictions. Among the techniques used for feature extraction, k-mers are the most commonly employed method. Although many methods based on k-mers have been proposed, these methods typically use only the frequency information of k-mers as features. However, when frequencies are identical, the frequency information of these k-mers becomes less useful. To address this limitation, we propose a novel method called PhageCGRNet, which not only utilizes the frequency information of k-mers but also incorporates the positional information of k-mers. In our method, we represent each genome sequence as a three-dimensional matrix containing k-mers frequency features and positional features, and then utilize the Convolutional Neural Network model to predict the host category. Specifically, we combine the frequency information of k-mers directly extracted from the sequences with the positional information of k-mers obtained using the Chaos Game Representation method to construct the feature matrix, which serves as the input to the Convolutional Neural Network. We conducted experiments on two benchmark datasets, and compared PhageCGRNet with existing advanced methods for phage host classification. The experimental results demonstrate that PhageCGRNet achieves higher accuracy at both taxonomy levels of species and genus on these two datasets compared to other state-of-the-art methods.噬菌体（bacteriophages，简称phages）在微生物群落中发挥关键作用，精准预测噬菌体的宿主，对于理解这类病毒的动态及其对细菌种群的影响至关重要。在噬菌体宿主分类预测任务中，特征提取是直接影响预测准确率的关键步骤。在特征提取的常用技术中，k聚体（k-mers）是应用最为广泛的方法。尽管已有诸多基于k聚体的方法被提出，但这类方法通常仅将k聚体的频率信息作为特征。然而，当不同k聚体的频率相同时，其频率信息的区分度会大幅下降。为解决这一局限，本文提出一种名为PhageCGRNet的新型方法，该方法不仅利用k聚体的频率信息，还融入了k聚体的位置信息。在该方法中，本文将每条基因组序列表示为包含k聚体频率特征与位置特征的三维矩阵，随后利用卷积神经网络（Convolutional Neural Network）模型完成宿主类别预测。具体而言，本文将直接从序列中提取的k聚体频率信息，与通过混沌游戏表示法（Chaos Game Representation）获取的k聚体位置信息相结合，构建特征矩阵并将其作为卷积神经网络的输入。本文在两个基准数据集上开展了实验，并将PhageCGRNet与现有的先进噬菌体宿主分类方法进行了对比。实验结果表明，相较于其他当前最优方法，PhageCGRNet在这两个数据集的物种和属两个分类学层级上均取得了更高的预测准确率。