A Conformation Variant of p53 Combined with Machine Learning Identifies Alzheimer Disease in Preclinical and Prodromal Stages

Early diagnosis of Alzheimer’s disease (AD) is a crucial starting point in disease management. Blood-based biomarkers could represent a considerable advantage in providing AD-risk information in primary care settings. Here, we report new data for a relatively unknown blood-based biomarker that holds promise for AD diagnosis. We evaluate a p53-misfolding conformation recognized by the antibody 2D3A8, also named Unfolded p53 (U-p532D3A8+), in 375 plasma samples derived from InveCe.Ab and PharmaCog/E-ADNI longitudinal studies. A machine learning approach is used to combine U-p532D3A8+ plasma levels with Mini-Mental State Examination (MMSE) and apolipoprotein E epsilon-4 (APOEε4) and is able to predict AD likelihood risk in InveCe.Ab with an overall 86.67% agreement with clinical diagnosis. These algorithms also accurately classify (AUC = 0.92) Aβ+—amnestic Mild Cognitive Impairment (aMCI) patients who will develop AD in PharmaCog/E-ADNI, where subjects were stratified according to Cerebrospinal fluid (CSF) AD markers (Aβ42 and p-Tau). Results support U-p532D3A8+ plasma level as a promising additional candidate blood-based biomarker for AD.

Recombinant p53 protein produced in baculovirus (ActiveMotif) was exposed to EDTA, that subtracting Zn atom gets lost p53 wild type conformation towards a misfolding phenotype. The graph reports results of the immunoassay performed with 2D3A8 antibody on different amounts of recombinant p53 (0.3-0.5-1ng) before and after EDTA treatment. Data are expressed as Optical Density (O.D.).

Blocking-epitope peptide inhibits 2D3A8 binding to p53 recombinant protein
To demonstrate 2D3A8 is specific to the interaction with p53 protein we blocked 2D3A8 antibody through competition with a peptide that matches the sequence of 2D3A8-epitope. Thus, 1 µg of 2D3A8 antibody has been pre-incubated with increasing doses of the blocking peptide (ratio Ab: blocking peptide: 1:1; 1:2; 1:3 and 1:4) and then used in the in-house ELISA assay. 5 ng of p53 recombinant protein treated with EDTA has been tested. We found that increasing doses of blocking epitope peptide are able to inhibit the binding of 2D3A8 antibody with p53 protein thus supporting 2D3A8 specificity.
Supplementary figure 2. 5 ng of recombinant p53 + EDTA has been incubated with 2D3A8 antibody both in presence or absence of increasing doses of 2D3A8 blocking epitope-peptide. Epitope peptide has been pre incubated according with the following antibody/blocking peptide ratio: 1:1 (light grey) p<0.05; 1:2 (medium grey) p<0.01; 1:3 (medium dark grey) p<0.01 and 1:4 (dark grey) p<0.001. No blocking peptide pre-incubation is reported with the white bar. Data are expressed as mean ± s.e.m.

Linear Mixed Effects Model
Two separate Linear Mixed Effect (LME) models have been performed: one comparing stable CN with CN to AD, and the other comparing stable MCI with MCI to AD. Models were fit through restricted maximum likelihood (REML), also known as residual maximum likelihood. Standard likelihood method (ML) was not adopted since it is reported to be biased in small samples. Assumptions about normality of residuals and homoscedasticity were checked through visual inspection. The description of the model output is described here below.
Level 1 U − p53 2D3A8+ = 0 + 1 + Level 2 0 = 00 + 01 + 0 1 = 10 + 11 + 1 The level 1 submodel represents the individual change in plasma_U-p53 2D3A8+ occurring over time (years). The plasma_U-p53 2D3A8+ at occasion j, for person I, is a function of an intercept, which corresponds to participant's plasma_U-p53 2D3A8+ level at baseline 0 , and one slope parameter. A linear term is included to capture the rate of change over time 1 . Level-1 residuals represent the portion of subject i's value of plasma_U-p53 2D3A8+ levels at time j not predicted by the model.
The level 2 submodels capture systematic interindividual differences in trajectories. At level 2, the level 1 intercept and slope parameters become the outcomes, and they are predicted as a function of four parameters: 00 , 10 , 10 and 11 . These parameters are the average intercept ( 00 ), the hypothesized difference in the average true initial status between diagnosis groups ( 01 ), the average true annual rate of change ( 10 ), the effect of diagnostic faith in the average true monthly rate of change ( 11 ). Each submodel has its own residual ( 0 , 1 ) that permits the level-1 parameters of one person to differ stochastically from those of the others. Of note: time was treated as unstructured by using patient's actual age at each assessment. To have a better interpretability of 00 and 01 parameters, baseline was set at 70 years old.

Participants and clinical phenotyping: Supplementary Tables
Supplementary Table 2. InveCe.Ab population study: Description and conversion rate within the follow up.
The All participants from InveCe.Ab dataset have been run through Neuropsychological assessments addressing several cognitive areas using the applicable instruments, as listed below ( Table S3). The different subgroups have been then selected according with age, gender, comorbidity index*, severity index* and clinical category matched. Global cognition was assessed using MMSE 1 , corrected for age and years of education following the normative data published by Magni et al. 2 Verbal episodic memory was evaluated using the revised version of the Babcock Story Recall Test and the Rey Auditory-Verbal Learning Test 3 . Language was assessed using the Phonemic and Semantic Verbal Fluency Test 4 . Executive functions were gauged using Raven's Coloured Matrices 5 and Clock Drawing Test 6 . Simple and divided attention, and attention control were tested using the Attention Matrices and Trail Making Test 7 . Finally, visuospatial skills were evaluated using the Rey-Osterrieth Complex Figure (copy and recall) 8 . Each evaluation session was preceded by an informal interview to evaluate potential interfering factors and to help the participants feel at ease. The medical evaluation together with the neuropsychological assessment provided information useful to calculate the prevalence of dementia and cognitive impairment. In Table

Regression Trees Performance
Supplementary Figure 4. Detailed description of RTs applied to InveCe.Ab dataset where the rolling window procedure is used for evaluating Out-Of-Sample model performances of the models obtained. The algorithm has grown two trees using the variables in both cases (grey box). It was calibrated on two different training set: (i) Baseline (yellow box) and (ii) Baseline +T1 (orange box). At each step, the model is tested on fresh data: (i) T1 (red box) and (ii) T2 (white box) respectively. The two models obtained are very similar with respect to variables and thresholds selected, providing the stability of the predictors. Consequently, authors selected the more accurate RT2 while RT1 is available upon request. Table 7. Performances of RTs reported in figure 3 in-sample (ten-fold Cross-Validation) and Out-Of-Sample.