The Institute of Scientific and Industrial Research, Osaka University, Mihogaoka 8-1, Ibaraki, Osaka, 567-0047, Japan.
National Institute of Advanced Industrial Science and Technology, Takamatsu, Kagawa, 761-0395, Japan.
Small Methods. 2021 Jul;5(7):e2100191. doi: 10.1002/smtd.202100191. Epub 2021 May 14.
Noise is ubiquitous in real space that hinders detection of minute yet important signals in electrical sensors. Here, the authors report on a deep learning approach for denoising ionic current in resistive pulse sensing. Electrophoretically-driven translocation motions of single-nanoparticles in a nano-corrugated nanopore are detected. The noise is reduced by a convolutional auto-encoding neural network, designed to iteratively compare and minimize differences between a pair of waveforms via a gradient descent optimization. This denoising in a high-dimensional feature space is demonstrated to allow detection of the corrugation-derived wavy signals that cannot be identified in the raw curves nor after digital processing in frequency domains under the given noise floor, thereby enabled in-situ tracking to electrokinetic analysis of fast-moving single- and double-nanoparticles. The ability of the unlabeled learning to remove noise without compromising temporal resolution may be useful in solid-state nanopore sensing of protein structure and polynucleotide sequence.
噪声在真实空间中无处不在，会干扰电传感器中微小但重要信号的检测。在本文中，作者报道了一种用于电阻脉冲传感中去噪离子电流的深度学习方法。通过电泳驱动单个纳米颗粒在纳米波纹纳米孔中的迁移运动来检测。通过卷积自动编码神经网络来降低噪声，该网络通过梯度下降优化来迭代比较和最小化一对波形之间的差异。这种在高维特征空间中的去噪被证明可以检测到波纹衍生的波浪信号，这些信号在原始曲线中无法识别，也无法在给定噪声底线下在频域中进行数字处理后识别，从而能够原位跟踪快速移动的单和双纳米颗粒的电泳分析。无标签学习去除噪声而不影响时间分辨率的能力可能对固态纳米孔中蛋白质结构和多核苷酸序列的传感有用。
