Quantifying tolerances or maximum residue limits of pesticide in food commodities via deep neural networks

Abstract BACKGROUND Quantifying tolerances or legal maximum residue limits (MRLs) of pesticide in/on food commodities is of significance to enforce the surveillance of food safety/quality and good agricultural practices (GAP). Current in silico models mostly focus on estimating field residue levels, for example via in situ or field maximum residue levels ( in situ MRLs), retention time and dissipation half‐life. In silico modelling of residue tolerances involves more complicated processes (e.g. in situ GAP MRLs estimation, daily dietary assessment and toxicological test), and receives little attention to the best of our knowledge. RESULTS In this work, based on the major considerations that tolerances settings use the maximum permissible intake (MPI) as toxicological risk index to accommodate the in situ MRLs estimated via supervised field trials under GAP, we conduct machine learning modelling of residue tolerances from the perspective of quantitative structure–activity relationships (QSAR). Through hierarchical clustering, we find that (1) structurally similar residues exhibit similar or transferable tolerances/MRLs profiles (termed as between‐residues MRLs transferability); and (2) close food species also exhibit similar or transferable tolerance/MRLs profiles (termed as between‐foods MRLs transferability). These findings provide a modelling basis for us to quantify the legal MRLs of a pesticide on untested food species from a mechanistic point of view. The available residue tolerance measurements of 438 pesticides in 128 food commodities are aggregated to train a global QSAR‐based deep neural network (DNN) regression model. The feature space is spanned by a chemical subspace represented with Morgan similarity spectra and a food subspace encoded with one‐hot vector. The results of 5‐fold stratified cross‐validation show that DNNs achieve overall 0.81 R 2 , significantly outperforming the conventional machine learning models, such as extreme gradient boosting (XGBoost), random forest (RF) and support vector regression (SVR). Computational results also show that the strategy of one model for one food species is not feasible for legal MRLs quantification, in which the majority of food commodities achieve negative R 2 . As a global model, the proposed DNN regression model encouragingly achieves ≥0.6 R 2 for >72.66% of food commodities. CONCLUSION Between‐residues and between‐foods MRLs transferability are proved to support the QSAR‐based modelling of residue tolerances in/on food commodities. The DNN, feature engineered via Morgan similarity spectra and food one‐hot vector, proves to be effective in setting residue tolerances. When considering GAP and field conditions, the proposed modelling strategy can be extended to or incorporate in situ field MRLs estimation. © 2025 Society of Chemical Industry.