GraphPI: Efficient\nProtein Inference with Graph Neural\nNetworks

The integration of deep learning approaches in biomedical\nresearch\nhas been transformative, enabling breakthroughs in various applications.\nDespite these strides, its application in protein inference is impeded\nby the scarcity of extensively labeled data sets, a challenge compounded\nby the high costs and complexities of accurate protein annotation.\nIn this study, we introduce GraphPI, a novel framework that treats\nprotein inference as a node classification problem. We treat proteins\nas interconnected nodes within a protein–peptide–PSM\ngraph, utilizing a graph neural network-based architecture to elucidate\ntheir interrelations. To address label scarcity, we train the model\non a set of unlabeled public protein data sets with pseudolabels derived\nfrom an existing protein inference algorithm, enhanced by self-training\nto iteratively refine labels based on confidence scores. Contrary\nto prevalent methodologies necessitating data set-specific training,\nour research illustrates that GraphPI, due to the well-normalized\nnature of Percolator features, exhibits universal applicability without\ndata set-specific fine-tuning, a feature that not only mitigates the\nrisk of overfitting but also enhances computational efficiency. Our\nempirical experiments reveal notable performance on various test data\nsets and deliver significantly reduced computation times compared\nto common protein inference algorithms.