[6/10] We built a proof-of-concept diagnostic report with ProtAIDe-Dx, visualizing diagnostic probabilities across conditions and highlighting proteins that contributed to the individual’s prediction and traits linked to those proteins, enabling a biologically interpretable diagnosis.[Figure 5]

[5/10] Could ProtAIDe-Dx aid real-world clinical application? In a memory clinic cohort, ProtAIDe-Dx significantly improved differential diagnosis (especially PD and CVD) when combined with accessible clinical biomarkers (MMSE, MRI, plasma p-tau217 & NfL). (Figure 4D).

[4/10] ProtAIDe-Dx identified a novel and compact set of top predictive proteins. Notably, OMG and Histone H1-2 (H1-2) distinguished controls from all neurodegenerative conditions, and ubiquitin (UBB) was a discriminative protein for FTD. [Figure 3A].

[3/10] The models’ diagnostic probabilities highlighted subgroups of patients with proteomic patterns more similar to a different condition, e.g., AD patients resembling stroke patients, indicating possible misdiagnosis, potential co-pathology, or may even indicate distinct etiologies. [Figure 2A].

[2/10] We developed a multi-task model, ProtAIDe-Dx, to provide simultaneous probabilistic diagnosis across six conditions associated with dementia in aging (Control, AD, PD, FTD, ALS, and Stroke/TIA). CVed balanced ranged from 69%-96% and AUCs > 79% across all conditions. [Figure 1B]