Global transfer learning pipeline for protein disease association in Alzheimer's disease.

Researchers have prioritized the study of protein-disease associations to decode triggers of clinical pathology and isolate high-value targets for drug development. Comprehensive modeling of genetic network dynamics is equally vital for advancing our functional understanding of these disorders. In this study, we applied a hierarchical, transfer-learning framework to address the challenge of protein-disease association prediction, specifically targeting scenarios with limited labeled data for specific diseases. We used Alzheimer's disease as a case study to demonstrate the efficacy of our proposed Global Transfer Learning Pipeline model. We combined the embeddings generated from protein-protein interactions along with protein-cluster association and protein sequences to train the proposed model. We addressed the scarcity of reliable negatives by employing PU learning strategies with deep fusion architecture to ensure robustness of the model. The ordinal regression was integrated to the learning pipeline to learn granular confidence levels for protein-disease associations which was later fed into the stacked meta model. The model also uses techniques of hyperparameter optimization to enhance the prediction performance. Our model achieved a weighted F1 score of 96% with average AUC 0.852 and AUPRC 0.967 which outperforms the individual gradient boosting, tree models and deep neural network models. These results demonstrate the effectiveness of the proposed model in predicting protein-disease associations for Alzheimer's disease.