Automatic Detection and Quantification of Chemicals of Interest Utilizing Deep Learning Algorithms in Electrochemical Biosensors

Abstract

Nanomaterial-based aptasensors are crucial for detecting small biological entities. Effective signal processing methods are essential for improving biosensor performance by enhancing the identification and quantification of target analytes. This thesis investigates signals from three electrochemical aptamer-based sensors, each featuring distinct receptors, analytes, and signal lengths. The ultimate objective of this study is the automated detection and quantification of target analytes within the sensor-recorded signals, resembling a classification task. In order to achieve this objective, the thesis explores evolutionary pathways. The initial classification task involves identifying and quantifying specific analyte concentrations across six distinct classes, ranging from zero presence to 10 µM. Subsequently, a second classification task was designed to differentiate abnormal from normal data segments, detecting the presence or absence of analytes in the sample. When detected, this task further aims to identify the specific analyte and quantify its concentration. Data scarcity presented a significant challenge, necessitating the use of data augmentation techniques. An initial approach involved scaling data augmentation to address limited original data availability. Long short-term memory (LSTM) networks were utilized for analyte concentration prediction, with systematic variations for fine-tuning in network configuration and data augmentation quantity showing notable impact on model performance. Consequently, building upon the findings from the previous step, an automatic anomaly detection method, utilizing autoencoder-based prediction models, was introduced in a semi-supervised learning approach. Autoencoder networks and kernel density estimation (KDE) helped detect anomalies, although uncertainties persist regarding segment length's impact on model performance. In the concluding phase, a data augmentation technique using conditional variational autoencoders was introduced to address data scarcity within deep learning algorithms. Recurrent-based networks were developed for signal extrapolation to ensure consistent signal lengths, and short-term Fourier transform (STFT) preprocessing was explored. Seven deep learning classification models (GRU, unidirectional LSTM (ULSTM), bidirectional LSTM (BLSTM), ConvGRU, ConvULSTM, ConvBLSTM, and CNN) were developed, demonstrating the effectiveness of preprocessing methods in improving neural network performance for analyte identification and quantification. Overall, this study contributes to automated analyte detection and quantification using aptasensors, offering practical data augmentation methods, signal reconstruction, and insights into unknown datasets. The introduced classification model provides inspiration for novel methodologies.