Application of RFdiffusion to predict interspecies protein-protein interactions between fungal pathogens and cereal crops
Plant pathogenic fungi secrete small proteins known as effectors which help overcome the plant defense response and cause disease. The concept of effector-triggered immunity in plants evolved from the "gene for gene hypothesis" which describes plant resistance or susceptibility to plant pathogens based on interspecies protein-protein interactions (PPIs) between plant-derived resistance (R) genes and pathogen-derived avirulence (Avr) effector genes. Understanding the molecular interactions mediating these host-pathogen interactions in effector-triggered immunity is thus essential to managing fungal disease. In silico methods of predicting interspecies PPIs have been heavily studied to identify target genes for crop resistance. But conventional sequence-based homology methods (i.e., interlog, domain-based inference) for predicting interspecies PPIs are not as powerful as methods that also incorporate structural homology. The objective of this study was to develop a computational workflow to predict PPIs between pathogenic fungi and their cereal hosts by leveraging recent advances in artificial intelligence and structural biology. This workflow proposes the use of a generative model, RFdiffusion, to predict the structure of truncated segments of proteins likely to bind to query effector proteins. The binder structures were filtered based on the number of contacts at the effectors known binding residues. Acceptable structures were then input into FoldSeek to search the host proteome for host proteins containing similar sub-structures. Experimentally-validated PPIs between rice (Oryzae sativa cv. Japonica) and rice blast fungus (Magnaporthe oryzae) were used for workflow validation. The effects of binder length and the binding residues mode of action (i.e., residues at active/substrate recognition sites) on the binder quality and presumptive host protein matches were explored. Ultimately, 11 out of 14 experimentally validated PPIs were recovered computationally, indicating a high recall (>78%) for the workflow. The shorter binders recovered most of the PPIs, but may have produced the most false positives, as functional analyses revealed that these host proteins displayed a wide variety of functions. These findings emphasize that subject matter expertise is still required to decipher the prediction results. Yet, this framework for elucidating interactions between fungal pathogens and host proteins could provide valuable insight into mechanisms of susceptibility or resistance at a scale friendly to limited computational resources, and facilitate the development of control strategies that reduce crop diseases.