Adjusting for CSF Reference Proteins Improves Biomarker Accuracy

Abstract

AbstractBackgroundCSF biomarkers are key AD descriptors, but the CSF volume as well as the production and clearance rates of proteins varies in individuals. Adjusting for reference proteins, which reflect these factors, might improve the accuracy of CSF biomarkers, similar to when normalizing CSF Aβ42 levels to Aβ40 levels. For other biomarkers, individual CSF dynamics are seldom adjusted for, which may result in disagreement with other techniques (e.g., PET). Additionally, this may contribute to inaccurate conclusions about CSF protein correlations that mainly exist due to high or low CSF dilution levels. The aim of this work is therefore to establish the use of CSF reference proteins to improve CSF AD biomarker accuracy.MethodWe used measurements of 2944 CSF proteins in the Swedish BioFINDER‐2 cohort (n = 982, 80% train and 20% test). Reference protein candidates (reflecting CSF dilution levels without being related to any disease) were searched for in a data‐driven manner, using t‐distributed stochastic neighbor embedding (t‐SNE) and a semi‐supervised K‐means clustering algorithm. The final candidate proteins were evaluated in logistic regression models, using CSF P‐tau181 and CSF Aβ42 as main predictors of tau‐PET and amyloid‐PET, respectively.ResultMost proteins were associated with an individual dilution level, although at different strengths (Fig 1 and Fig 2e). From the different colorings of the t‐SNE space, proteins of similar expressions were placed nearby, creating areas of extra interest (yellow areas in subfigures 2c‐2h). The semi‐supervised K‐means algorithm highlighted an area (cluster 11) containing proteins with superior reference protein characteristics, from which single reference protein candidates were identified. All examined references (e.g., Aβ40, NTRK3, NTRK2, BLMH,CBLN4, PTPRN2 or PTPRS) outperformed using no reference for both the P‐tau181 and Aβ42 model (Figure 3). For P‐tau181, highest performance was achieved using CBLN4 (AUC = 0.944). For Aβ42, Aβ40 performed best as reference (AUC = 0.992).ConclusionThe individual CSF dilution level affects many CSF proteins. We describe several novel and previously suggested CSF reference proteins, which can significantly improve the accuracy of CSF biomarkers as key AD descriptors.