A deep joint-learning proteomics model for diagnosis of six conditions associated with dementia

Co-pathology is a common feature of neurodegenerative diseases that complicates diagnosis, treatment and clinical management. However, sensitive, specific and scalable biomarkers for in vivo pathological diagnosis are not available for most neurodegenerative neuropathologies. Here we present Proteomics-based Artificial Intelligence for Dementia Diagnosis (ProtAIDe-Dx), a deep joint-learning model on 17,187 patients and controls (age of 70.3 ± 11.5 years, 53.2% female), that uses plasma proteomics to provide simultaneous probabilistic diagnosis across 6 conditions associated with dementia in aging. ProtAIDe-Dx achieves cross-validated balanced classification accuracy of 70-95% and area under the curve of >78% across all conditions. The model's diagnostic probabilities highlighted subgroups of patients with co-pathologies and were associated with pathology-specific biomarkers in an external memory clinic sample, even among individuals without cognitive impairment. Model interpretation revealed a suite of protein networks marking shared and specific biological processes across diseases and identified novel and previously described proteins discriminating each diagnosis. ProtAIDe-Dx significantly improved biomarker-based differential diagnosis in a memory clinic sample, pinpointing proteins leading to diagnostic decisions at an individual level. Together, this work highlights the promise of plasma proteomics to improve patient-level diagnostic workup with a single blood draw.