Hybrid Quantum–Machine Learning Models For High-Dimensional Gene Expression Profiling And Early Risk Prediction Of Parkinson’s And Alzheimer’s Disease

Abstract

Molecular changes in the early stage of Parkinson’s disease (PD) and Alzheimer disease (AD) occur before disease onset, and thus, transcriptomic profiling could be an effective approach to risk prediction in the early phases of the disease. Nonetheless, the generated datasets of gene expression are very high-dimensional usually having thousands of genes with few samples, that are not readily accommodated by traditional machine learning methods, which face the risk of overfitting and low nonlinear representational properties. Strong computational models are thus needed to obtain discriminative molecular signatures and generalise at the same time. This work suggests a quantum-machine learning (QML) framework to profile high-dimensional gene expression to enhance prediction of the early risk of both PD and AD and maintain biological insights. Most public transcriptomic datasets were preprocessed with log transformation, Z-score normalisation, and then they were subjected to the analysis of differential expression and mutually informative feature selection. The encoding of the genes of interest into a variational quantum circuit was through angle encoding, which allowed nonlinear representation to a higher-dimensional feature space. Regularised cross-entropy loss was used to train a hybrid quantum -classical architecture that used parameterized quantum layers followed by a classical classifier. Accuracy, precision, recall, F1-score and ROC-AUC were used to measure performance that was benchmarked against support vector machines, random forests and classical neural networks. This model proposed was more successful in higher ROC-AUC, F1-scores in single datasets of PD, and AD, and showed greater ability to perform in high-dimensional and small samples. The process of neuroinflammation via genes associated with neuroinflammatory pathways, mitochondrial dysfunction, and synaptic signaling were determined by feature importance analyses and pathway enrichment as per known neurodegenerative mechanisms.