Tomato yellow leaf curl virus (TYLCV) is among the most devastating pathogens affecting tomato production worldwide, with emerging virulent strains increasingly overcoming genetic resistance and triggering severe outbreaks. Traditional field diagnosis, reliant on visual inspection or image-based AI models, remains constrained by symptom dependence, environmental variability, and poor strain-level interpretability. To address these challenges, we introduce DeepTYLCV, a novel deep learning framework for accurate virulence prediction directly from viral genome-derived open reading frame (ORF) sequences. We first constructed a comprehensive dataset of globally sourced TYLCV sequences, curated by virulence annotations. DeepTYLCV integrates protein language model (PLM)-based embeddings with optimal concatenated conventional descriptors (optCCDs) using a hybrid architecture composed of a Transformer encoder and a multi-scale convolutional neural network (CNN), enabling effective extraction of both global and local sequence features. Benchmark analyses demonstrate that DeepTYLCV significantly outperforms our previously developed IML-TYLCV model, which was trained on Korean isolates and lacked global generalizability. Importantly, blind predictions on 15 uncharacterized or representative TYLCV isolates were experimentally validated in tomato plants, achieving 100% concordance between model predictions and observed symptom severity. Furthermore, 1D-Grad-CAM++-based interpretability analyses revealed that the model consistently focused on relevant sequence motifs associated with severe strains, offering mechanistic insights into symptom severity. DeepTYLCV is publicly available at https://balalab-skku.org/DeepTYLCV/, represents a powerful, interpretable, and globally scalable platform for early TYLCV surveillance, resistance monitoring, and strategic disease management in tomato cultivation.
Bupi et al. (Wed,) studied this question.