Qualitative detection of pesticide residues using mass spectral data based on convolutional neural network

Abstract Excessive pesticide residues in crops directly threaten human life and health, so rapid screening and effective measurements of agricultural pesticides residues have important application significance in the field of food safety. It is imperative to detect different pesticide residue types in actual complex crop samples cause mixture analysis can provide more information than individual components. However, the accuracy of mixture analysis can be obviously affected by the impurities and noise disturbances. Purification and denoising will cost a lot of algorithm time. In this work, we used the problem transformation method to convert pesticide residues prediction into multi-label classification problem. In addition, a new convolutional neural network structure Pesticide Residues Neural Network (PRNet) was proposed to solve the problem of multi-label organophosphate pesticide residue prediction. The method of binary correlation and label energy set was used to adapt 35 pesticide residues labels. The Cross Entropy were used as loss functions for PRNet. The comprehensive comparison performances (e.g. 97% optimal accuracy rate) of PRNet is better than the other four models. By comparing the ROC curves of the five models, PRNet performs the best. The PRNet can separate the independent mass spectrometry data by different collision energy applied to phosphorus pesticide compounds through a three-channel structure. No complicated data preprocessing is required, the PRNet can extract the characteristics of different compounds more efficiently and presents high detecting accuracy and good model performance of multi-label mass spectrometry data classification. By inputting MS data of different instruments and adding more offset MS data, the model will be more transplantable and could lay the foundation for the wide application of PRNet model in rapid, on-site, accurate and broad-spectrum screening of pesticide residues in the future