A machine learning framework for noise-driven DNA sequencing in solid-state nanopores.

Abstract

Solid-state nanopore devices have emerged as a promising platform for electrical signal based single molecule DNA sequencing with atomic-scale precision. However, stochastic noise in solid-state nanopores severely obscures signal quality, impeding nucleotide identification accuracy. Here, we have systematically examined the impact of noise on transverse tunnelling signals and their machine learning (ML) classification for nucleotide identification. The ML classification accuracy exhibited a non-monotonic relationship with an increase in noise intensity, where moderate noise levels can paradoxically yield higher accuracy than noise-free signals, reflecting the complex interplay between noise-induced variability and model learning behaviour. Denoising preprocessing significantly improved the signal-to-noise ratio (SNR >12) along with ML classification accuracy as high as 98% across all noise levels. Furthermore, graph neural networks (GNNs) applied directly to raw noisy signals achieved similarly high accuracy without any signal preprocessing, demonstrating an inherent robustness gained by learning the graph-structured feature representations. These results highlight that the GNN architectures can offer a noise resilient alternative for accurate nucleotide identification in solid-state nanopore sequencing, paving the way for robust single-molecule analysis under realistic experimental conditions.