Biomarkers.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomarker Discovery

Background:

• Alzheimer's disease (AD) exhibits pathological heterogeneity, causing diagnostic and prognostic challenges.
• Normative modeling quantifies individual deviations from healthy norms to understand disease heterogeneity.
• This study applies a GAN-based approach to measure personalized disease effects in AD.

Purpose of the Study:

• To develop and validate a GAN-based normative modeling technique for quantifying individual-level neuroanatomical and proteomic abnormalities in Alzheimer's disease.
• To assess the model's ability to detect disease-related heterogeneity at both structural and molecular levels.
• To establish a personalized biomarker for disease effects and severity in AD patients.

Main Methods:

• Adapted a pix2pix Generative Adversarial Network (GAN) to translate diseased subjects to disease-free counterparts.
• Trained neuroanatomical models using MRI ROI volumes (n=6000) and proteomic models using CSF protein expression data (n=96) from healthy controls.
• Simulated patients by perturbing the healthy reference sample, with deviations serving as personalized biomarkers.

Main Results:

• The GAN-based models detected significant abnormalities in both CSF proteomic and structural ROI measures.
• Larger deviations were observed in individuals with Mild Cognitive Impairment who converted to AD (cMCI) and AD patients compared to controls and stable MCI patients.
• These findings indicate disease-related abnormalities detectable through both structural and molecular data.

Conclusions:

• The GAN-based normative modeling technique is effective for parsing heterogeneity at an individual level in Alzheimer's disease.
• The model demonstrates the capability to detect pathological effects in both structural and molecular data.
• This approach holds potential for personalized deviation detection and improved AD management.