K-mer list for an imaginary nucleotide sequence.

An important issue in biotechnology is predicting whether a piRNA and an mRNA will or will not bind. Research and treatment of diseases, drug discovery, and the silencing and regulation of genes, transposons, and genomic stability may all benefit from accurate binding predictions. The literature offers numerous deep-learning models for piRNA and mRNA binding prediction. However, a proper adjustment of the effective transformer model and the impact of important design alternatives has not been evaluated thoroughly. This paper summarizes the models available in the literature, briefly introduces transformers, then offers a novel deep learning model and evaluates various design alternatives, including k-mer size, number of core modules, choice of optimization algorithm, and whether to use self-attention. The results show that rbpTransformer can be a good candidate for building deep AI models to predict the binding of piRNA and mRNA sequences with an AUC value of 94.38%. The test results also reveal how the design affects the model’s accuracy.生物技术领域的一项核心课题，是预测piRNA（piwi-interacting RNA）与mRNA（messenger RNA）是否会发生结合。疾病的研究与诊疗、药物研发，以及基因、转座子的沉默与调控、基因组稳定性的维持等诸多领域，均可从精准的结合预测中获益。现有文献中已提出诸多用于piRNA与mRNA结合预测的深度学习模型，但针对高效Transformer（Transformer）模型的合理调优，以及关键设计选型的影响，尚未得到充分评估。本文首先梳理了现有文献中的相关模型，简要介绍Transformer模型，随后提出一种全新的深度学习模型，并对多种关键设计选型展开评估，包括k-mer（k聚体）长度、核心模块数量、优化算法的选择，以及是否采用自注意力机制。实验结果表明，rbpTransformer可作为构建piRNA与mRNA序列结合预测深度学习AI模型的优质候选方案，其AUC（受试者工作特征曲线下面积）值可达94.38%。本次测试结果还揭示了不同设计选型对模型精度的影响机制。