Deep Learning-Based Identification of Pathogenicity Genes in Phytophthora infestans Using Time-Series Transcriptomics

TL;DR

This paper uses deep learning to identify genes in a pathogen that cause potato late blight, offering new targets for disease-resistant crops.

Contribution

A novel LSTM–Transformer hybrid model with temporal attention for identifying pathogenicity genes from time-series transcriptomics.

Findings

200 high-confidence pathogenicity-related genes were identified from potato infection time-series data.

These genes are enriched in 15 biologically meaningful pathways, including plant immunity and stress responses.

New candidates were found to participate in defense hormone pathways and cell wall modification.

Abstract

Potato (Solanum tuberosum L.) is the world’s fourth most important food crop, and despite China producing nearly one quarter of the global yield, its potato production is severely constrained by late blight. Identifying genes associated with pathogenicity is essential for breeding resistant cultivars and strengthening plant protection strategies. Traditional approaches based on differential expression and statistical modeling often fail to capture temporal dynamics or provide interpretable insights. Here, we introduce an LSTM–Transformer hybrid model designed for data-driven discovery of pathogenicity-related genes from gene expression time-series. The analysis was performed on a time-series expression dataset comprising 32,917 genes across 18 samples (three infection time points × six biological replicates per condition). In this study, we identified 200 high-confidence pathogenicity-related genes from potato infection time-series data. These genes are enriched in 15 biologically meaningful pathways, including plant immunity signaling, reactive oxygen species regulation, secondary metabolic processes, and stress-responsive transcriptional programs. Several newly uncovered candidates participate in defense hormone pathways and cell wall modification, suggesting previously unrecognized roles in late blight susceptibility and resistance. By revealing functional groups and regulatory signatures that characterize pathogenicity, this work provides valuable molecular targets for developing late blight-resistant cultivars. The framework integrates a biologically informed temporal–attention architecture, a gene time-series-specific data partitioning strategy, and an interpretable deep analysis module. A final methodological contribution is the use of a temporal attention-based analytical framework that enables reliable gene prioritization from time-series expression data.

Figures