Machine Learning for Molecular Perceptron: A Perception-Based Sensing System

Abstract

Single wall carbon nanotube (SWCNT) based biosensors provide opportunities for building an ultra-sensitive biosensing system due to their unique optical properties and strong sensitivity to changes in the local environment. Consequently, much effort has been made to develop SWCNT-based sensors. However, the usual method is based on one-to-one recognition which is a difficult way to detect various molecules since it requires the same number of highly specific receptors as the number of molecules one wishes to detect. To detect a combination of various analytes simultaneously, an effective and automatic data processing system is essential. In this study, we propose a new perception-based sensing system using weakly-specific sensor arrays that can be analyzed by an artificial perception model, which we call the Molecular Perceptron. We show how machine learning algorithms along with choice of feature representation is designed to predict presence and concentration of biomarkers or direct prediction of disease states. For example, we demonstrate that the Molecular Perceptron can detect Human epididymis protein 4 (HE4) in the presence or absence of other analytes; HE4 is one of two FDA-approved serum biomarkers for ovarian cancer which provides noticeable sensitivity and specificity for ovarian cancer diagnosis. DNA/SWCNT hybrids were utilized to optically detect the analytes by observing changes in the fluorescence spectra of each SWCNT. Using the experimental data, machine learning models were trained using three different algorithms: Support Vector Machine, Random Forest, and Artificial Neural Network. The models were then validated using new experimental data for different analyte concentrations. Overall, the machine learning models successfully predict the presence of HE4 at the concentrations of 10 nM or higher by giving F1-scores of ~0.85. This is strongly suggestive of the idea that the perception mode of sensing can make accurate judgements in a noisy sensing environment.