DeepEffector: A Protein Language Model-Driven Deep Learning Framework for Effector Prediction in Fungi and Oomycetes

Abstract Crop diseases caused by fungal and oomycete pathogens threaten sustainable agricultural production, and the efficient identification of pathogen effectors is essential for understanding virulence mechanisms and developing disease-resistant crops. However, the rapid evolutionary diversification of effectors often obscures conserved motifs, limiting the effectiveness of conventional computational approaches and making experimental prioritization challenging. To address this, we developed DeepEffector, a high-precision predictor for fungal and oomycete effector identification. DeepEffector leverages protein language models (ProtT5 for fungi and esm1b for oomycetes) to generate functional semantic embeddings, integrates G-SMOTE to mitigate class imbalance in the latent feature space, and employs a hybrid multi-head self-attention and convolutional neural network architecture to capture both global contextual dependencies and local functional patterns. Benchmarking on independent test sets showed that DeepEffector outperformed state-of-the-art tools, providing a reliable high-confidence candidate list to reduce downstream experimental screening. Its practical utility was further evaluated in Fusarium oxysporum f. sp. cubense TR4, the causal agent of destructive banana Fusarium wilt. Four prioritized candidates were experimentally confirmed to be secreted proteins, localized to host-relevant subcellular compartments, and significantly suppressed Bax-induced programmed cell death, supporting their potential roles as immune-suppressing effector candidates. DeepEffector is available as a user-friendly web server (https://deepeffector.hainanu.edu.cn/), offering an accessible pipeline to accelerate effector discovery and support crop protection research.