Explainable Deep Learning-Assisted Fluorescence Discrimination for Aminoglycoside Antibiotic Identification

The complexity and multivariate analysis of biological systems and environment are the drawbacks of the current high-throughput sensing method and multianalyte identification. Deep learning (DL) algorithms contribute a big advantage in analyzing the nonlinear and multidimensional data. However, most DL models are data-driven black boxes suffering from nontransparent inner workings. In this work, we developed an explainable DL-assisted visualized fluorometric array-based sensing method. Based on a data set of 8496 fluorometric images of various target molecule fingerprint patterns, two typical DL algorithms and eight machine learning algorithms were investigated for the efficient qualitative and quantitative analysis of six aminoglycoside antibiotics (AGs). The convolutional neural network (CNN) approached 100% prediction accuracy and 1.34 ppm limit of detection of six AG analysis in domestic, industrial, medical, consumption, or aquaculture water. The class activation mapping assessment explicates how the CNN model assesses the importance of sensor elements and makes the discrimination decision. The feedback mechanism guides the sensor array evolution for less material using a simplified operation or efficient data acquisition. The explainable DL-assisted analysis method establishes an end-to-end strategy to resolve the black box of the DL algorithm, promote hardware design or principle optimization, and contribute facile indicators for environment monitoring, disease diagnosis, and even new scientific discovery.

Automated summary

This study developed an explainable DL-assisted visualized fluorometric array-based sensing method for efficient qualitative and quantitative analysis of antibiotics. The convolutional neural network achieved high prediction accuracy and guided the improvement of the sensor array through a feedback mechanism.