Fig. 5: Clinical utility of ProtAIDe-Dx.
From: A deep joint-learning proteomics model for diagnosis of six conditions associated with dementia
a, Multi-class classification of patients with AD, PD, FTD and stroke using various models inclusive of demographics, ProtAIDe-Dx embeddings and/or common clinical biomarkers. Left: one-versus-rest BCA computed by 1,000 bootstraps on testing patients. Right: confusion matrices on testing patients for model 2 (top) and model 3 (bottom). Error bars represent the s.d. across 1,000 bootstrap resamples. b, Left: model-predicted baseline diagnoses differentiated longitudinal MMSE trajectories of GNPC patients irrespective of true baseline clinical diagnosis. Trajectories were modeled using linear mixed-effects model of MMSE with Age, Sex, Site, BaselineDx, BaselinePredictedDx, Year and the BaselinePredictedDx × YearInteraction, with subject-specific random effects for SubjectID. Right: replication on BioFINDER-2 patients with MCI. Solid lines indicate linear mixed-effects model-predicted mean MMSE trajectories, and shaded regions represent 95% confidence intervals for the linear mixed-effects fixed-effects predictions. c, Biomarker distributions across predicted probability bins. CT, cortical thickness. d, Two-cutoff strategies for predicting biomarker positivity. Cutoffs were derived from non-SCD BioFINDER-2 participants (n = 1,524) to achieve 90% NPVs and PPVs, except for α-synuclein, where the PPV was set at 40% owing to sample validity constraints. These fit cutoffs were then applied to patients with SCD (n = 263) to estimate accuracy in this clinically relevant sample. Icons in d created in BioRender; An, L. https://biorender.com/q2by4y5 (2026).
