Prior knowledge informs graph neural networks to improve phenotype prediction from proteomics.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• High-throughput proteomics data offers rich molecular insights for predicting patient health phenotypes.
• Complex, nonlinear protein interactions challenge traditional linear models and simple machine learning.
• Developing effective deep neural networks for biological data remains a significant hurdle.

Purpose of the Study:

• To develop an advanced deep learning framework for predicting disease-related traits from protein expression data.
• To design an innovative model architecture that integrates structured biological knowledge.
• To enhance the predictive performance of machine learning models in proteomics.

Main Methods:

• Developed a deep learning framework centered on a graph neural network (GNN) operating on bipartite graphs.
• Utilized gene ontology libraries to construct protein sets, forming nodes in the graph structure.
• Trained the model on UK Biobank plasma proteomics and phenotype data, employing a multi-head architecture.

Main Results:

• The best-performing architecture combined two GNNs with independent protein set libraries and a global tabular data head.
• The model demonstrated strong predictive performance for glycated hemoglobin (HbA1c) and other phenotypes, surpassing existing deep learning and linear models.
• Control models with permuted labels showed significantly worse performance, validating the model's benefit from biological knowledge.

Conclusions:

• The developed deep learning framework effectively predicts complex traits from large-scale proteomic data by integrating biological domain knowledge.
• The innovative GNN architecture exploits structured biological information, improving predictive accuracy, especially with limited data.
• This approach represents a significant advancement in applying deep learning to biological data for health-related predictions.