BPS2026 – Single microRNA identification in low-concentration samples with nanopore measurement and transfer learning

Nanopore sensing is a widely used technique for detecting nucleic acids due to its ability to analyze molecules at single-molecule resolution. The process is used in various applications, including DNA sequencing, RNA analysis, and protein detection. In our study, we measured microRNAs (miRNAs), which are important biomarkers for early cancer detection varying expression patterns according to the type of cancer. The measurement involves passing molecules through a nanoscale pore in the lipid membrane. Common target nucleic acids are double-stranded, which slows down the translocation and results in a variety of signal properties. However, identifying short and single-stranded miRNAs is challenging due to the similarity in signal shapes among various molecules. In this study, we generated images for each signal combining results of multiple analyses to enable more advanced interpretations of complex information. Additionally, we utilized convolutional neural network (CNN) to precisely detect small differences in images. In neural network training, data set size directly affects classification performance, while acquiring a large number of signals for unquantified, low-concentration nucleic acids via nanopore is time-consuming. In high-concentration samples, significant proportions of miRNAs form different structures, such as dimers, which can alter the signal shapes. To address this, we implemented transfer learning, which involves two training steps. Initially, knowledge is acquired by training on a large data set from samples with concentrations 1,000 times higher. By learning lower concentration samples based on this knowledge, high performance can be achieved even with small amounts. The results suggested that combining multiple analyses to generate images enhanced the accuracy of miRNA classification. The comparison with models trained exclusively on high concentration data sets demonstrated that applying transfer learning significantly improves the model's accuracy in identifying low-concentration signals and recognizing the miRNA ratio in test samples.