DeepGreenGO: a graph neural network-based deep learning model for plant-specific protein function prediction

Abstract

Automated protein function prediction remains challenging in plants because experimentally supported annotations are limited, particularly for crop and non-model plant species. Integrating deep learning techniques in plant molecular biology can offer a transformative opportunity to innovate research and support sustainable agriculture. But experimentally annotated plant proteins are not well represented in most multi-taxon datasets used to train function prediction models limiting the transferability of their performances for plant specific purposes. This study introduces DeepGreenGO, a deep learning multilabel classifier that integrates ProtBERT-BFD residue embeddings with contact maps derived from protein structures. Residue representations are processed sequentially using GCN and GATv2 layers, followed by attention pooling to predict Gene Ontology terms specific to the model’s training ontology. This model was trained on proteins from Viridiplantae . This method was subsequently deployed to predict proteins involved in rice ( Oryza sativa ) seed development, to demonstrate its functionality. The curated Viridiplantae dataset comprised 7,534 experimentally annotated protein structures and was partitioned into training, validation, and test sets, containing 6,026, 754, and 754 PDB chains, respectively, using sequence-similarity-aware clustering. The integration of ProtBERT-BFD embeddings provided most of the predictive signal. DeepGreenGO was evaluated against other sequence homology-based tools, and other recent sequence- and structure-based deep learning methods. DeepGreenGO showed strong performance, particularly in the biological process (BP) ontology, achieving the highest protein-centric \(\:{F}_{max}\) of 0.227 and the lowest \(\:{S}_{min}\) of 19.73. Particularly, it scored the highest information-content weighted area under precision recall (IC-weighted AUPRC) in comparison to all other methods for BP, indicating improved prediction of rare and informative functions. Application of the model to 43,649 rice proteins identified 372 candidate proteins associated with seed development. Functional enrichment and transcriptomic analyses showed that many predicted proteins were linked to seed-development-related biological processes and displayed elevated expression in ovary, embryo, and endosperm tissues, supporting the biological relevance of the predictions. This study introduces DeepGreenGO, a taxonomically focused framework for predicting plant protein functions. Its strongest advantages were observed for BP ontology, informative annotations, and proteins with limited detectable similarity to the training set. The ablation results further show that pretrained sequence representations are central to its performance, while identifying opportunities to improve the integration of structural information. Its application in predicting rice seed development proteins highlights the potential to support plant biology research and sustainable agriculture.