Fig. 1: A biophysical interaction between APP and Lf supports machine learning analysis of a lead hit in classifying disease and predicting Aβ pathology.

A, B The top 10 genes identified by z-score using the feature selection algorithm, Boruta for classifying the ROSMAP cohort into neuropathologically positive AD cases and controls (A). Of these genes a small fraction were not identified by differential expression analysis using EdgeR (B). C Top 10 genes ranked by z-score predicting total amount of amyloid pathology. D Top 10 genes ranked by z-score predicting p-tau (AT8) immunoreactivity across all seven cortical regions and the hippocampus. E Sedimentation coefficient distributions of holo-Lf alone (5 μM; orange) and APP (2.5 μM) in the absence (black) and presence of Lf (green: 2.5 μM; blue: 3.75 μM; red: 5 μM). The c(S) distributions indicate two different complexes that form in a concentration dependent manner. See Supplementary Fig. S2 for example data sets showing data quality and the fit of the c(s) model to the data. F The weight average sedimentation coefficient obtained by integrating the c(S) distribution (as shown in E) calculated as a function of Lf concentration (Eq 1 in Supplementary data) with the assumption that APP contains two binding sites for Lf with two different dissociation constants. G Sedimentation velocity analysis of apo-Lf alone (2.5 μM; orange) and APP (2.5 μM) in the absence (black) and presence of apo-Lf (purple: 2.5 μM; maroon: 12.5 μM). H Confirmation of an interaction of APP with Lf in vivo using anti-APP (1:1000; 22C11) for detection and anti-Lf (1:200) for immunoprecipitation of brain homogenates from APP–/– mice and littermate controls either on a normal or high-iron diet. Specificity of interaction of APP to Lf was confirmed using anti-β-actin as the capture antibody (data not shown). Data are mean ± SEM of three experiments performed in triplicate.