Automated Diagnostic Classifiers Using Imaging, Genotyping, and Gene Expression Data

S46

IC-P-093

Alzheimer’s Imaging Consortium: IC-P-Poster Imaging

PREDICTING INITIAL CHANGES IN BRAIN VOLUME WITH EFFECTIVE ANTI-AMYLOID AGENTS: JUST ASK ALICE

Jesse Cedarbaum, Cytokinetics, Inc., South San Francisco, California, United States. Background: Changes in the rate of brain atrophy have been proposed as biomarker surrogate measures for Alzheimer’s disease (AD) clinical trials. The prevailing hypothesis is that a successful therapeutic agent will slow, arrest or reverse the decline in the inexorable progression of brain atrophy associates with the disease. Numerous recent reports have put forward “power” and “sample size” calculations for hypothetical clinical trials based on hypothesized slowing of brain shrinkage in treated, compared with control subjects. On the other hand, in the clinical trials of the experimental anti-amyloid vaccine AN1792, there appeared to be an increase in the apparent rate of brain atrophy in subjects who developed significant titers of anti-amyloid antibodies in the blood, and the “acceleration” of brain atrophy appeared proportional to titer. The few subjects studied 4.5 years after the conclusion of the trial showed no greater brain tissue loss than they had at the end of the 1-year study, and autopsied brains of several persons dying after receiving AN1792 contained no amyloid deposits. Can the AN1792 observations and the expectation of decreased atrophy with successful disease-modifying AD treatment be reconciled? Methods: We identified several competing factors which might affect rates of brain volume change with anti-amyloid treatment, including 1) normal rates of age-associated brain volume loss; 2) the additional disease-related contribution to brain volume loss; 3) estimated volume of the brain occupied by amyloid deposits and its regional distribution; 4) rate of accumulation of amyloid in the brain; 5) the contribution of inflammation to brain volume and 6) the rate of brain amyloid removal or decreased rate of brain amyloid accumulation. A mathematical model was constructed, using literature-derived values for as many of the variables as possible. Results: The model was used to predict changes in brain volume under typical clinical trial conditions. Results will be presented. Conclusions: Several factors contribute to the steady-state levels of amyoid deposits in the brains of AD patients. These factors, as well as the mechanism of action of the drug under study, may affect observed changes in brain volume during clinical trials. IC-P-094

IIIA (MYO3A), a gene encoding the subtype of myosin that interacts with the GRINLIA glutamatergic receptor, while showing regionally similar associations with cortical thinning, also showed an association with the left lateralparietal cortex (left hemisphere pcorrected ¼ 0.049, right hemispherep ¼ 0.12). Higher expression of growth hormone 1 (GH1), a regulator of neuronal survival via insulin-like growth factor, showed a pattern of association similar to that of MYO3A with trend-level significance on the left (pcorrected ¼ 0.059). Finally higher expression of Amyotrophic lateral sclerosis (ALS) 2 chromosome region11(ALS2CR11), a gene encoding a protein with calcium binding properties that has been implicated in the pathogenesis of ALS, correlated with bilateral cortical thinning of the sensorimotor strip, supplementary motor area and precuneus/posterior cingulate cortex as well as left inferior frontal and anterior cingulate cortex (left hemisphere pcorrected ¼ 0.044, right hemisphere p ¼ 0.3). Conclusions: Altered peripheral blood gene expression is associated with cortical atrophy in the pre-dementia cognitive spectrum.

PERIPHERAL BLOOD GENE EXPRESSION CORRELATES OF CORTICAL ATROPHY ACROSS IN COGNITIVELY NORMAL ELDERLY AND MCI

Liana Apostolova1, Kristy Hwang1, Giovanni Copolla1, Jessica Lane1, Fying Gao1, Jeffrey Cummings2, Paul Thompson1, 1UCLA, Los Angeles, California, United States; 2Lou Rouvo, Las Vegas, Nevada, United States. Background: Human genome-wide gene expression studies have generated important knowledge about the unique influences of multiple genes in health and disease. The complex interactions between gene expression and imaging biomarkers in Alzheimer’s disease (AD) and the at-risk state of mild cognitive impairment (MCI) may yield important insights about the genetic influences and pathogenesis of AD. Methods: We collected peripheral blood and 1.5T 3D MPRAGE T1-weighted brain MRI data from 38 cognitively normal (NC), 18 nonamnestic and 19 amnestic MCI subjects. Peripheral blood RNA was extracted using Paxgene tubes, amplified, labeled, and hybridized onto Illumina Human Ref Seq-8 Bead Chip arrays, querying the expression of w24,000 RefSeq curated transcripts followed by quality control, quantile normalization with R and Bioconductor packages.The imaging data were analyzed with the cortical pattern matching and cortical thickness mapping techniques. Linear regression was used to examine the relationship between log2-transformed absolute gene expression levels and cortical thickness while adjusting for age and gender. For multiple comparison correction we used permutation tests with a threshold of p < 0.01. Results: Higher expression of Sialic acid binding lg-like-lectin 10 (SIGLEC10), a gene associated with tissue damage-induced immune responses, showed associations with bilateral temporo-occipital, precuneal, posterior cingulate and inferior temporal cortical thinning (left hemisphere pcorrected ¼ 0.032;right hemisphere pcorrected ¼ 0.043). Higher expression of Myosin

Figure. Significace and correlation maps of the association between peripheral blood gene expression levels SIGLE10, MYO3A, GH1 and ALS2CR11 and cortical thickness IC-P-095

AUTOMATED DIAGNOSTIC CLASSIFIERS USING IMAGING, GENOTYPING, AND GENE EXPRESSION DATA

Liana Apostolova1, Kristy Hwang1, Omid Kohannim1, Giovanni Coppola1, Eric Klein1, Fuying Gao1, Jeffrey Cummings2, Paul Thompson1, 1UCLA, Los Angeles, California, United States; 2Cleveland Clinic Lou Ruvo Center for Brain Health, Las Vegas, Nevada, United States. Background: Genome-wide association and gene expression studies have revealed how multiple genes influence human health. Imaging genetics

Alzheimer’s Imaging Consortium: IC-P-Poster Imaging may yield important insights into genetic influences on the biology of Alzheimer’s disease (AD). We investigated the accuracy of a novel unsupervised multimodal biomarker classifier for differentiating cognitively normal elderly (NC) from subjects with amnestic mild cognitive impairment (aMCI). By combining imaging and genetic biomarker data, we hypothesized that we would achieve greater accuracy in differentiating the diagnostic groups. Methods: Using automated segmentation techniques, we derived hippocampal and lateral ventricle volumes from the T1-weighted MRI data of 46 NC and 35 aMCI subjects. We collected gene expression (GE) data from all subjects and single nucleotide polymorphism (SNP) data on common variants in ApoE, TOMM40, PICALM, CLU, CR1, MAPT and PCDH11X from 44 NC and 28 aMCI subjects. Using a novel automated support vector machine algorithm , we developed unimodal and multimodal imaging and genetic diagnostic classifiers. All classifiers included age and sex. Results: In the N ¼ 72 imaging/SNP dataset, a classifier that used hippocampal volume only achieved 76.4% diagnostic accuracy (area under the curve, AUC ¼ 0.67) compared to the classifier based on ventricular volume only - accuracy 69.4% (AUC ¼ 0.56) and the combined hippocampalventricular classifier - accuracy 74% (AUC ¼ 0.7). The addition of SNP variables led to a hippocampal-SNP classifier accuracy of 76% (AUC ¼ 0.74) and a hippocampal-ventricular-SNP classifier accuracy of 72% (AUC ¼ 0.69). Of the 7 SNPs entered, PICALM was selected by the hippocampalSNP classifier, ApoE by the hippocampal-ventricular-SNP classifier while TOMM40 was selected by both classifiers. The remaining SNPs were not included in the optimal classification algorithm. In the N ¼ 81 imaging/ gene expression dataset the hippocampal-only classifier achieved 69% diagnostic accuracy (AUC ¼ 0.63) and ventricular-only classifier achieved 69% accuracy (AUC ¼ 0.57) compared to the hippocampal-GE classifier - accuracy 78% (AUC ¼ 0.79), and the combined hippocampal-ventricular-GE classifier - accuracy 84% (AUC ¼ 0.82). 12 expressed genes and 8 expressed genes were selected as being useful for improving classification, by the final hippocampal-GE and hippocampal-ventricular-GE combined classifiers, respectively. Conclusions: As hypothesized, NC vs. aMCI classifier performance improved when combining imaging and genetic biomarkers. Automated classifiers show great promise for diagnostic analyses and potentially for predicting future conversion to AD.

Figure. Receiver Operating Characteristic (ROC) curvers for the unsupervised NC vs amnestic MCI classifier using structural imaging and SNP (top row) or gene expression data (bottom row).AUC stands for area under the curve. IC-P-096

TOMM40 RS2075650, TOMM40 RS157580 AND TOMM40 POLYT POLYMORPHISM EFFECTS ON VENTRICULAR ENLARGEMENT IN INDIVIDUALS WITH AND WITHOUT MILD COGNITIVE IMPAIRMENT

Sona Babakchanian1, Kristy Hwang1, Giovanni Coppola1, Erick Klein1, Jessica Lane1, Sterling Johnson2, Paul Thompson1, Jason Lee1, Jeffrey Cummings3, Liana Apostolova1, 1UCLA, Los Angeles, California,

S47

United States; 2University of Wisconsin, Madison, Wisconsin, United States; Cleveland Clinic Lou Ruvo Center for Brain Health, Las Vegas, Nevada, United States. 3

Background: Apolipoprotein E4 (APOE4) is the most established sporadic Alzheimer’s disease (AD) susceptibility gene. TOMM40, a gene adjacent to APOE4, has been postulated to increase one’s risk of AD. TOMM40 polyT polymorphism has been implicated in modulating AD age of onset among APOE4-negative subjects. Methods: Genotyping for APOE4, TOMM40 rs2075650 and rs157580, and TOMM40 polyT polymorphism analyses were performed on 44 cognitively normal elderly (NC) and 48 mild cognitive impairment (MCI) subjects. A novel automated ventricular segmentation technique and the radial distance mapping approach were applied to the subjects’ T1-weighted magnetic resonance imaging data. Multiple linear regression with a permutations threshold of p < 0.01 was used to measure the effect of TOMM40 rs2075650 and TOMM40 rs157580 on ventricular radial distance while correcting for APOE4 genotype. We also analyzed associations between the short (S ¼ 20) and very long (VL ¼ 30) polyT repeat length, and ventricular enlargement in our 59 APOE4-negative individuals: 10 were S/S, 28 were S/VL and 21 were VL/VL carriers. The S/S, S/VL and VL/VL groups were compared. Results: TOMM40 rs2075650 and TOMM40 rs157580 failed to show significant associations with ventricular radial distance. Presence of VL showed significant association with the right temporal (pcorrected ¼ 0.039) and left occipital horns (pcorrected ¼ 0.014) in ApoE4-negative subjects. Trend level effects were detected in the right occipital (pcorrected ¼ 0.085) and frontal (pcorrected ¼ 0.094) horns. In between-group comparisons, S/VL carriers showed significantly smaller occipital horns than S/S carriers (left pcorrected ¼ 0.014; right pcorrected ¼ 0.015). Conclusions: Among APOE4-negative subjects, presence of VL repeats associates with smaller lateral ventricles. These data are in agreement with our previous report of VL repeat being associated with larger hippocampi.

IC-P-097

OVER-TIME RELIABILITY OF MTA GRADING IN A CLINICAL SETTING 1 € ard3, Lena Cavallin , Kirsti L€oken2, Rimma Axelsson1, Anne-Rita Okseng Lena Bronge4, Lars-Olof Wahlund3, Knut Engedal5, 1CLINTEC Karolinska Institute, Stockholm, Sweden; 2Curato, Oslo, Norway; 3NVS Karolinska Institute, Stockholm, Sweden; 4CLINTEC Karolinska Hospital, Stockholm, Sweden; 5Aldring og Helse, Oslo, Norway. Background: Atrophy inmedial temporal lobe, (MTA) is one of the first MR signs for patients with Alzheimer’s disease. Visual assessment of MTA developed by Scheltens et al(1992, 1995), is a quick method suitable for a clinical setting. It is easy to repeat and used as a tool in monitoring Alzheimer’s disease progression. The reliability of visual assessment of MTA over time is not studied. The aim of this study is to investigate what happens to the MTA scores if two radiologists rate the same MRI-scans 6 times over a period of a year. Methods: One hundredout-door patients were included in this method study. Fifty patients with memory problems from Minnesmottagningen Huddinge Karolinska University Hospital,Stockholm, Sweden and 50 patients from Hukommelseklinikken, Ullev al University Hospital, Oslo, Norway. All patients underwent an MRI with protocol for geriatric purposes according to the local clinical standard, including a coronal T1 sequence overhippocampus head and body. Two radiologists, one neuro radiologist (LC) and one general radiologist(KL) performed the investigation. At time zero, the two radiologists rated all 100 patients together in consensus. One week, 1 month, 3 months, 6 months and 1 year after consensus, they rated the same 100 patients by themselves without information about the previous ratings. Results: The over all intra rating kappa values for LC was 0.78on the right side, and 0.74 on the left side. For KL the over all kappa value was 0.60 at theright side, and 0.56 on the left side. Weighted kappa values was excellent for both radiologists. Inter Inter rating reliability was fair to moderate. Kappa value varied from 0.29 to 0.48 and weighted kappa values varied from 0.72 to 0.84. There was a statistical significant difference in rating between the two radiologists but not between the Swedishand Norwegian populations Conclusions: Intrarating reliability show high agreement, if visual assessment of MTA is performed many times daily. Inter rating reliability is low, probably