Direct Gene Detection in Raw Nanopore Signals Using Transformer Neural Networks

Abstract

Nanopore sequencing has transformed genomics by enabling real-time analysis of DNA and RNA in a compact, cost-effective device. However, conventional workflows require a separate basecalling step to convert raw electrical signals into nucleotide sequences, which can introduce errors and delay downstream analyses such as gene detection. Here, we present a novel approach that bypasses basecalling by directly analyzing raw nanopore signals using a transformer-based neural network. By adapting a model originally designed for ECG classification, we developed a system capable of detecting specific antibiotic resistance genes in Klebsiella pneumoniae samples. Raw signals were preprocessed through downsampling, z-normalization, and segmentation into 5,000-sample windows, yielding a dataset of 13,080 labeled segments. Experimental results demonstrate that our model effectively distinguishes gene-containing segments from non-target signals, achieving up to 80% accuracy in the “no target gene” category. In contrast, accuracy for other gene categories was lower, indicating that further optimization of the model is required. This direct-signal approach not only reduces the computational burden associated with basecalling but also streamlines the workflow, promising faster diagnostic turnaround times. These findings provide a significant step toward integrating advanced deep learning methods with nanopore sequencing for rapid, on-site genomic analysis and have potential applications in clinical diagnostics and epidemiological surveillance.