Accurate plant pathogen effector protein classification ab initio with deepredeff: an ensemble of convolutional neural networks

• Main Authors: Kristianingsih, R. (Author), MacLean, D. (Author)
• Format: Article
• Language: English
• Published: BioMed Central Ltd 2021
• Subjects: Activation analysis; AI; Biological insight; Chemical activation; Classification (of information); Convolution; Convolutional neural networks; Deep learning; Detection software; Effector protein; End effectors; Food supply; Global food security; Learning classifiers; Learning systems; Long short-term memory; Model architecture; Neural Networks, Computer; Personnel training; plant protein; Plant Proteins; Protein Classification; Proteins; Secreted protein; Sequence features; software; Software; Software testing
• Online Access: View Fulltext in Publisher

 Background: Plant pathogens cause billions of dollars of crop loss every year and are a major threat to global food security. Effector proteins are the tools such pathogens use to infect the cell, predicting effectors de novo from sequence is difficult because of the heterogeneity of the sequences. We hypothesised that deep learning classifiers based on Convolutional Neural Networks would be able to identify effectors and deliver new insights. Results: We created a training set of manually curated effector sequences from PHI-Base and used these to train a range of model architectures for classifying bacteria, fungal and oomycete sequences. The best performing classifiers had accuracies from 93 to 84%. The models were tested against popular effector detection software on our own test data and data provided with those models. We observed better performance from our models. Specifically our models showed greater accuracy and lower tendencies to call false positives on a secreted protein negative test set and a greater generalisability. We used GRAD-CAM activation map analysis to identify the sequences that activated our CNN-LSTM models and found short but distinct N-terminal regions in each taxon that was indicative of effector sequences. No motifs could be observed in these regions but an analysis of amino acid types indicated differing patterns of enrichment and depletion that varied between taxa. Conclusions: Small training sets can be used effectively to train highly accurate and sensitive deep learning models without need for the operator to know anything other than sequence and without arbitrary decisions made about what sequence features or physico-chemical properties are important. Biological insight on subsequences important for classification can be achieved by examining the activations in the model. © 2021, The Author(s).