CSF proteome profiling reveals highly specific biomarkers for dementia with Lewy bodies

At Amsterdam UMC we train thousands of young people to become doctors, specialists or nurses. Amsterdam UMC has eight research institutes. The researchers affiliated with these are internationally among the top in medical scientific research. Through the cooperation of Amsterdam UMC with other hospitals in the region of Noord-Holland and Flevoland, the patient receives the highest quality of care in the right place. At Amsterdam UMC we offer excellent academic patient care, state-of-the-art scientific research and top-level education and training.

The data preprocessing and analyses were performed using R version 3.5.3 and SPSS version 25. Between-group analyses for demographic variables were conducted using two-sided one-way analysis of variance for normally distributed continuous data or Pearson's chi-square test for categorical variables. Analysis of covariance was performed when an association between classical AD CSF biomarker and age and/or sex was detected. Adjustment for multiple testing was done using the Bonferroni method. Non-Gaussian distributed data were analyzed using the Kruskal-Wallis Test. For the CSF proteome data, differences in protein abundance between pairs of clinical groups were evaluated using nested linear models. Each individual protein feature was assessed to determine if its addition to a base model containing age and gender contributed to model fit. Multiplicity was taken into account by controlling the False Discovery Rate (FDR) at q ≤ 0.05 based on the number of features analyzed. To identify the best CSF protein combination (CSF panels) that could discriminate the groups of interest while keeping the number of markers to a minimum, binary classification signatures (DLB vs. CON and DLB vs. AD) were constructed using penalized generalized linear modeling (GLM) with an elastic net penalty. Age and sex were included as covariates. The elastic net penalty, which is a linear combination of lasso and ridge penalties, enabled estimation in settings with a high feature to sample ratio. It also performed automatic feature decorrelation and feature selection. Multiple models were compared, considering different values for the elastic-net mixing parameter and the maximum number of proteins that could be selected under each model (up to 21 markers maximum). The optimal penalty parameters were determined based on balanced 10-fold cross-validation of the model likelihood. The predictive performance of all models was assessed using Receiver Operating Characteristic (ROC) curves and Area Under the ROC Curves (AUCs). The model with the highest AUC and lowest number of markers for each classification signature was selected. The fold-based selection proportions for each marker were assessed to identify and select the most promising markers within each model. To reflect the manual selection pressure for the final marker sets, each final logistic signature was subjected to ridge-regularization with a penalty parameter of 0.1. The performance (AUC) was evaluated by internal validation using repeated 5-fold cross-validation with 1000 repeats. External validations were also performed to assess the performance of the final models with the markers of interest in validation cohorts using ROC analysis. Non-parametric correlation analysis was conducted to understand the associations between the proteins within the CSF panels and the classical AD CSF biomarkers or cognitive function (MMSE score). This analysis was performed using the complete discovery cohort without stratifying per diagnostic category and conditioning on age and sex as covariates. Functional enrichment analysis was performed using Metascape, selecting GO Biological Processes as the ontology source. All the CSF proteins optimally analyzed with Olink arrays (a total of 645 protein gene products) were included as the enrichment background. Default parameters were used for the analysis, and terms with a p-value < 0.01, a minimum count of 3, and an enrichment factor > 1.5 were collected and grouped into clusters based on their membership similarities.