Deep Learning of Nanopore Sensing Signals Using a Bi-Path Network

Abstract

Temporal changes in electrical resistance of a nanopore sensor caused by translocating target analytes are recorded as a sequence of pulses on current traces. Prevalent algorithms for feature extraction in pulse-like signals lack objectivity because empirical amplitude thresholds are user-defined to single out the pulses from the noisy background. Here, we use deep learning for feature extraction based on a bi-path network (B-Net). After training, the B-Net acquires the prototypical pulses and the ability of both pulse recognition and feature extraction without a priori assigned parameters. The B-Net is evaluated on simulated data sets and further applied to experimental data of DNA and protein translocation. The B-Net results are characterized by small relative errors and stable trends. The B-Net is further shown capable of processing data with a signal-to-noise ratio equal to 1, an impossibility for threshold-based algorithms. The B-Net presents a generic architecture applicable to pulse-like signals beyond nanopore currents.

A Guide to Signal Processing Algorithms for Nanopore Sensors

“…Based on the Residual Neural Network (ResNet), a bipath NN, named Bi-path Network (B-Net), has recently been established to extract spike features. 35 Since the task of counting the number of spikes is essentially different from that of measuring the amplitude and duration, the bipath design, composed of two ResNets, each one trained for one task, has been shown to be robust with compelling performance. During the training process, segments of time sequence traces are first sent to the NN.…”

“…The relative errors can be calculated for each data point/situation, and the average and standard deviation of these relative errors on the total output points m can be further derived to reflect the overall performance of the algorithm on a certain data set. 35 …”

“…The relative errors can be calculated for each data point/situation, and the average and standard deviation of these relative errors on the total output points m can be further derived to reflect the overall performance of the algorithm on a certain data set. 35 For discrete outputs from the classification tasks, such as identified classes of the analytes from step 4 and extracted spike features in qualitative, categorical, or attribute variables, error rate (ER), and accuracy (Acc) are commonly adopted to count the incorrectly and correctly classified data, respectively…”

“…Setting a large threshold increases the risk of omitting translocation spikes, misleadingly rendering a low translocation frequency. On the other hand, having a small threshold can mistakenly lead to assignment of noise fluctuations as translocation spikes, thereby incorrectly increasing the translocation counts . To reduce the subjectivity due to involvement of the user in the threshold selection, the background noise level can be used as a reference. − As an example, a certain multiple of the root-mean-square (RMS) value of the background noise can be taken as the threshold. , Nonetheless, two potential subjectivity risks persist.…”

“…Regarding the continuous output from the regression tasks, such as the denoised current trace from step 1, extracted spike segments from step 2, continuously varied spike features from step 3, and inferred properties of analytes with continuous values from step 4, relative error (err r ) and mean-squared error (err ms ) are usually employed as indexes to evaluate the performance, defined as where x is the measured value, x 0 the ground truth, and m the total number of output points. The relative errors can be calculated for each data point/situation, and the average and standard deviation of these relative errors on the total output points m can be further derived to reflect the overall performance of the algorithm on a certain data set …”