PPEPFinder: A deep learning framework integrating sequence embeddings and structural graph representations for...

Plant pathogens, such as fungi and oomycetes, primarily rely on the secretion of effector proteins to successfully colonize host plants and suppress immune responses, ultimately leading to disease symptoms. Therefore, accurately identifying these effector proteins is crucial for understanding the pathogenic mechanisms and developing strategies for disease resistance. However, the current effector protein identification mainly depends on experimental screening methods, which are labor-intensive and costly. In recent years, the rapid development of deep learning approaches and the emergence of protein language models (PLMs) have presented unprecedented opportunities for developing new bioinformatics methods to identify effector proteins. In this context, we propose Plant Pathogen Effector Protein Finder (PPEPFinder), an integrated deep learning framework designed to predict effector proteins of fungi and oomycetes. PPEPFinder consists of three individual predictive models: (i) a sequence-based transformer model that utilizes rich semantic embeddings generated by a pre-trained PLM, Evolutionary Scale Modeling (ESM), as input; (ii) two structure-based Graph Attention Network models that represent protein structures as residue contact graphs using the ESM-generated embeddings or the structure pre-trained model SaProt embeddings as node features. To maximize the predictive performance, PPEPFinder employs a logistic regression model to aggregate the three predictors' outputs into a final prediction score. Compared to existing state-of-the-art effector prediction tools, PPEPFinder demonstrates superior performance by jointly leveraging sequence and structural information. We have made PPEPFinder and its associated datasets freely accessible to the community through an online prediction server (http://zzdlab.com/PPEPFinder/) and GitHub (https://github.com/mdiyuan/PPEPFinder).