Explainable Deep Learning-assisted Fluorescence Sensor Array Analysis

Inspired by human sensory system,the cross-reactive sensor arrays can make differentiated responses to different analytes by integrating multiple sensors.The modeling methods are used to analyze the differential information,so as to realize the recognition of multiple substrates.However,the response information of sensor arrays is usually high throughput,and it is difficult to establish the relationship between detection information and analytes by traditional modeling methods,which hinders the popularization and application of the sensor arrays in practice.Deep learning algorithms,an important branch of artificial intelligence,are good at mining nonlinear relations.More and more researches are based on deep learning to solve problems in chemical research.Deep learning assisted sensor array analysis shows great advantages for multi-substrate analysis of complex systems.Due to the complexity of network structure,deep learning is generally regarded as black box models,and it is difficult to explain the decision-making mechanism of the models.Studying the decision mechanism of deep learning models helps to understand the response mechanism of sensors and the rationality of data acquisition mechanism.In this paper,different kinds of antibiotics in complex systems are analyzed.Sufficient and effective sensor information is collected by fluorescence sensor array.Combined with explainable deep learning algorithm,an explainable deep learning-assisted fluorescence sensor array analysis method is established.The main research contents include:1.Based on the chemical structure of aminoglycoside antibiotics in the similarities and differences,we designed a fluorescence sensor array consisting of three commercial fluorescent indicator systems(Naphthalene-2,3-diformaldehyde,Alizarin red-S-Diphenylborinic acid 2-aminoethyl ester and Sulfonyl rhodamine B)to obtain the chemical structure information of six different aminoglycoside antibiotics.Different antibiotics and fluorescence indicator system have different effects,resulting in different fluorescence signals.Fluorescence fingerprint images of different antibiotics were obtained by visualization techniques.This convenient and readily available sensor array provides a new idea for complex systems such as drug detection and food analysis.2.In the data analysis section,different deep learning and machine learning algorithms are used to excavate the fingerprint information of antibiotic molecules in fluorescent fingerprint images.Finally,convolutional neural network algorithm with the highest image information mining efficiency was selected to construct and optimize the classification and regression models,realizing rapid and accurate qualitative and quantitative analysis of aminoglycoside antibiotics in different environmental water systems.3.The attention mechanism of the convolutional neural network model is further studied by using the class activation mapping method.Mining the chemical characteristic information which contributes most to the decision of convolutional neural network model.In this way,the design mechanism of sensor array is revealed,and based on this,feature extraction is carried out to promote the evolution of sensor array design.This “end-to-end” feedback mechanism provides a new way to guide the design and optimization of sensor arrays,environmental monitoring,disease diagnosis,and even new scientific discoveries.4.Taking β-lactam antibiotics as the research objects,the applicability of explainable deep learning assisted fluorescence sensor array method was verified.A fluorescence response sensor array based on three spirooxazine-metallic complexes was designed to obtain chemical structure information of antibiotics.Spirooxazine can take place in open and closed loop reaction in the absence of light,ultraviolet and visible light,accompanied by changes in fluorescence signal.When combined with metal ions,it can produce a different response to antibiotics,changing the balance of the open-loop reaction.Different antibiotics have different affinity with spirooxazine-metal complex,so the balance of open closed loop is different,leading to the difference of fluorescence signal.The rich sensing information of different antibiotics was obtained by visualization techniques.The explainable deep learning algorithm was used to mine the fingerprint information of antibiotics in fluorescent fingerprint images to achieve rapid qualitative and quantitative analysis of antibiotics in different systems,and class activation mapping method was used to explain the classification decision mechanism of the deep learning model.