Abstract: In order to address the issues of insufficient unimodal features, underutilization of multimodal data, and high data noise in drug-target interaction prediction (DTI), this paper proposes a drug target interaction prediction method based on multimodal stack hybrid autoencoder (SHADTI). SHADTI consists of three components: a data preprocessing module, a latent feature extraction module, and a prediction module. The data preprocessing module employs the random walk with restart (RWR) and positive pointwise mutual information (PPMI) algorithms to process the global topological structure of the multimodal data of drugs and targets. The latent feature extraction module utilizes a deep autoencoder that integrates denoise blocks, sparse blocks, and stacked blocks to effectively mine the information embedded in the multiple modalities, thereby generating latent drug-target feature vectors. The prediction module concatenates the latent features of drugs and targets and feeds them into a fully connected layer for prediction. The proposed method is compared with existing deep learning methods on five publicly available datasets, and the experimental results consistently outperform the compared methods. The results show that SHADTI effectively leverages complementary information from multimodal data and improves the accuracy of DTI prediction.

Key words: drug-target interactions, multimodal, autoencoder, deep learning

摘要： 针对药物-靶标相互作用（drug-target interactions，DTI）预测中存在的单模态特征不足、多模态数据利用不充分、数据噪声大等问题，提出了一种基于多模态栈式混合自编码器（stacked hybrid autoencoder, SHADTI）的药物靶标相互作用预测方法。SHADTI包括数据预处理模块、潜在特征提取模块和预测模块三部分。数据预处理模块利用随机游走和PPMI（positive pointwise mutual information）算法对药物和靶标的多模态数据进行全局拓扑结构处理。潜在特征提取模块利用深度自编码器混合了降噪块、稀疏块、堆栈块，充分挖掘多模态之间蕴含的信息，生成潜在药物靶标特征向量。预测模块将药物和靶标的潜在特征拼接后输入到全连接层进行预测。所提方法在5个公开数据集上与现有深度学习方法进行对比，实验结果均优于所对比的方法，表明SHADTI能够有效利用多模态数据间的互补信息，提高了DTI预测精度。

关键词: 药物-靶标相互作用, 多模态, 自编码器, 深度学习