INTRODUCTION: Alzheimer's disease (AD) progresses over more than a decade, with brain damage occurring before noticeable behavioral symptoms. Early detection is crucial for slowing disease progression and enhancing patient quality of life. Existing biomarkers like amyloid β (Aβ40, Aβ42) are inadequate for early-stage detection, necessitating novel biomarkers. We developed the Proteonano™ platform for deep, untargeted plasma proteome profiling to detect AD early.

METHODS: Proteonano™ technology was developed as an affinity-selective mass spectrometry platform, including usage of nanoparticle-based affinity protein binders (nanobinders) to enrich low abundance proteins and employment of an automated workstation for parallel sample preparation. Patients were serially enrolled under approved ethical review. Plasma samples were collected at time of enrollment, and participants were stratified as normal (N), mild cognitive impairment (MCI), and dementia (D). Plasma samples were processed through the ProteonanoTM pipeline and analyzed by a ThermoFisher Orbitrap Astral mass spectrometer at data independent acquisition mode. Raw data were analyzed by using DIA-NN, normalized, further processed by using a customized biostastic and bioinformatic pipeline.

RESULTS: Plasma samples from 206 serially enrolled participants (N=142, MCI=35, D=29) were analyzed. 4347 protein groups were identified, with 2344 ± 37 (AVG±SE) proteins per sample. 2704 proteins matched the human plasma protein project published plasma protein database. Protein concentrations ranged over nine orders of magnitude, with the lowest at 1.6 pg/mL. Differential expression analysis showed 64, 91, and 159 proteins with different abundances between MCI vs. N, D vs. MCI, and D vs. N groups, respectively. MGP was the most upregulated and SPTBN1 was the most downregulated in MCI group relative to normal group. Eight feature selection methods were employed, and top features selected in each method were subjected to Akaike information criteria (AIC) based model selection. Identified features were then combined for another round of selection. Differentiating powers of these models were assessed by receiver operating curve (ROC) and precision and recall (PR) methods. The best model for differentiating MCI from N used nine proteins, achieving a ROC-AUC of 0.92 (95 % confidence interval: 0.89-0.98)and PR-AUC of 0.84, outperforming Aβ 40 (ROC-AUC=0.80). Models for differentiating D vs. N (eight proteins, ROC-AUC=0.99, PR-AUC=0.96) and D vs. MCI (nine proteins, ROC-AUC=0.98, PR-AUC=0.97) also performed well. This indicates that nanobinder-assisted untargeted proteomics can effectively identify protein features distinguishing different cognitive states, enhancing early AD detection.