The current investigation addresses the pressing need to integrate orphan or underutilized crops into mainstream agriculture, focusing on Perilla (Perilla frutescens L.) due to its superior nutritional profile. A major challenge is the lack of fast, cost-effective, and labor-efficient screening methods for germplasm. Near-Infrared Reflectance Spectroscopy (NIRS) addresses this by providing precise and rapid determination of crucial biochemical parameters. This study developed Modified Partial Least Squares (mPLS) regression-based NIRS prediction models using WinISI and 1D Convolutional Neural Networks (CNN) to enable high-throughput screening for moisture, ash, proteins, total soluble sugars (TSS), and phenols in Perilla germplasm. Calibration with WinISI involved mathematical treatments, optimizing for each trait: "2,6,6,1" for moisture, "3,4,4,1" for ash and TSS, "3,4,6,1" for protein, and "2,4,6,1" for phenols. The 1D CNN model, with lower mean absolute error (MAE), was further validated. External validation metrics, including RSQexternal, SEP(C), slope, bias, and RPD, assessed prediction accuracy. Comparative evaluation showed WinISI performed better for moisture prediction, while the 1D CNN model excelled in predicting ash, protein, TSS, and total phenol, highlighting the importance of model selection for specific traits. This rapid screening tool aids in identifying nutritionally dense Perilla genotypes, guiding targeted breeding efforts, and represents the first comparative mPLS and DL-based modeling using NIRS data for Perilla.
