MOF Nanozyme Sensor Array Encodes Wavelength-Resolved Catalytic Fingerprints for Interpretable Pesticide Screening

Rapid pesticide screening is limited by low-dimensional optical readouts that may not preserve sufficient information to distinguish structurally diverse residues and estimate their concentrations. Here, three poly(vinylpyrrolidone) (PVP)-modulated MOF-199 copper nanozymes were combined with laccase-like, peroxidase-like, and superoxide dismutase-like chromogenic reactions to construct a 3 × 3 cross-reactive material-reaction sensor array. Quantitative XPS and N2 sorption revealed PVP-dependent differences in surface composition, external surface area, and pore volume. Across eight pesticides and twelve concentrations from 0.1 to 100 μg mL-1, full UV-vis spectra from 400 to 700 nm were assembled into 9 × 301 matrices. A multitask one-dimensional convolutional neural network (1D-CNN) decoded these matrices to jointly perform pesticide classification and concentration regression, achieving 94.4% accuracy, macro-F1 = 94.4%, R2 = 0.894, and RMSE = 10.46 μg mL-1 in standard-solution tests. Spectral-window ablation showed improved prediction with broader wavelength coverage, while an endpoint comparison further supported the advantage of the full spectral representation. Input-gradient and Grad-CAM analyses mapped model reliance to material-reaction channels and chromogenic wavelength regions, supporting interpretable identity-concentration decoding. In apple, cucumber, and pepper matrices, the trained nine-class classifier assigned representative pesticides and blank controls with 97.2% overall accuracy. These results demonstrate that wavelength-resolved catalytic fingerprints support interpretable classification and concentration estimation of single-pesticide samples in standard solutions and retain class-discriminating information in the tested produce matrices.