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

Area of Science:

• Plant pathology
• Bioinformatics
• Computational biology

Background:

• Plant pathogens secrete effector proteins to colonize hosts and suppress immunity, causing disease.
• Experimental identification of effector proteins is labor-intensive and costly.
• Deep learning and protein language models offer new bioinformatics solutions.

Purpose of the Study:

• To develop an advanced deep learning framework, PPEPFinder, for predicting effector proteins in fungi and oomycetes.
• To improve the accuracy and efficiency of effector protein identification compared to existing methods.

Main Methods:

• Developed PPEPFinder, an integrated deep learning framework using sequence and structure information.
• Employed a sequence-based transformer model with Evolutionary Scale Modeling (ESM) embeddings.
• Utilized two structure-based Graph Attention Network models with ESM or SaProt embeddings.
• Integrated predictions using a logistic regression model for a final score.

Main Results:

• PPEPFinder demonstrated superior performance in predicting effector proteins.
• The framework effectively leverages both sequence and structural information.
• Outperformed existing state-of-the-art effector prediction tools.

Conclusions:

• PPEPFinder offers a powerful and accurate bioinformatics tool for effector protein prediction.
• The framework enhances understanding of plant pathogen mechanisms.
• Facilitates the development of novel disease resistance strategies.