Biomarkers.

Area of Science:

• Neuroscience
• Proteomics
• Artificial Intelligence

Background:

• Plasma proteomic biomarkers show promise for dementia diagnosis.
• Validation in large neurodegenerative cohorts for multi-disease diagnosis is lacking.
• AI models can predict major dementia forms and underlying pathologies.

Purpose of the Study:

• Develop AI models to predict major dementia forms using plasma proteomics.
• Account for multiple underlying pathologies and output probabilistic information.
• Validate AI model performance in large neurodegenerative cohorts.

Main Methods:

• Utilized SomaLogic 7K plasma proteomics data from 17,170 Global Neurodegeneration Proteomics Consortium (GNPC) participants.
• Developed 'ProtAIDe', a deep network for classifying six clinical diagnostic categories.
• Employed 10-fold cross-validation, leave-one-site-out generalization, and feature permutation for analysis.

Main Results:

• ProtAIDe achieved balanced classification accuracy >0.7 and AUC >0.8 using ~200 proteins.
• Baseline embeddings predicted longitudinal cognitive decline (AUC 0.76±0.09).
• Probabilities correlated with cognitive scores and APOE genotype, revealing disease clustering and comorbidities.

Conclusions:

• ~200 key plasma proteins can differentiate major dementia forms at the patient level.
• ProtAIDe provides probabilistic information for identifying co-pathologies.
• The model aids in determining individual-level protein drivers of symptoms.