Graph-based learning and read-out of nanopore translocation event signals

Nanopores enable single-molecule analysis by measuring current signals through nanoscale pores in either biological or solid-state membranes. Accurate detection of analyte fingerprints within the pore environment is essential for reading-out the analyte type. We develop a framework for robust and label-free detection of the molecular nanopore events using a graph representation of the measured signals. To this end, we build a graph-based two-stage workflow based on a convolutional and graph neural networks that first perform a fast screening of the nanopore events, followed by a deep validation of these. The learned model can thus efficiently and in an unsupervised manner select possible molecular signatures (the current blockades) in the full signal, denoise, validate, reconstruct these, and predict the morphology of unseen molecular events. We could show that the learned model can efficiently predict the correct event morphology for the same analyte within a 2.4-fold range of transmembrane voltage values not included in the training. The developed graph-based workflow is modular, generalizable, and provided that it is trained on a huge amount of different nanopore experiments has the potential to become a blueprint model for nanopore read-out. Such a read-out model would be able to identify subtle differences in molecules like proteins, as well as their conformational or folding states. The proposed framework is developed using experimental signals from DNA translocation through an aerolysin pore and demonstrates a unified approach linking unsupervised feature learning to raw-signal inference for single-molecule sensing.