Basic Science and Pathogenesis.

Area of Science:

• Artificial Intelligence
• Genomics
• Neuroscience

Background:

• Large-scale multi-omics data from consortiums like ROSMAP hold promise for understanding Alzheimer's disease (AD) mechanisms.
• Challenges in multi-omics integration include modality collapse and data incompleteness, limiting machine learning applications.
• Existing methods struggle to fully leverage rich, yet often incomplete, multi-omics datasets.

Purpose of the Study:

• To develop an Alzheimer's disease (AD) prediction model that effectively integrates incomplete multi-omics data.
• To identify key biomarkers for AD diagnosis and treatment through feature importance analysis.
• To overcome the limitations of modality collapse and data incompleteness in multi-omics research.

Main Methods:

• A novel deep learning model with Encoder, Aggregator, and Predictor modules was developed.
• The model encodes, aggregates, and predicts AD using multi-omics data, ensuring meaningful contribution from each modality.
• A collective loss function aligns heterogeneous embeddings, and missing modality information is compensated by amplifying available data.

Main Results:

• The model achieved 0.890 accuracy in classifying cognitive diagnoses (CogDX), outperforming state-of-the-art methods.
• Ablation studies confirmed all omics modalities contributed, preventing modality collapse and highlighting the value of incomplete data.
• Feature importance analysis identified known AD biomarkers, consistent with existing literature.

Conclusions:

• The proposed method successfully integrates incomplete multi-omics data for Alzheimer's disease (AD) prediction.
• The identification of relevant biomarkers demonstrates the potential of deep learning in multi-omics research for AD.
• This approach offers a robust framework for leveraging complex biological datasets in neurodegenerative disease research.