For the past three decades, we have studied the world’s largest known pedigree in which a presenilin-1 (PSEN1) mutation (p.Glu280Ala, E280A), often referred as the Paisa mutation, dominantly cosegregates with early-onset of Alzheimer’s disease (AD).1 This founder effect dates from the Spanish Conquistadors colonizing Colombia during the early sixteenth century.1, 2, 3, 4 Of the more than 5000 individuals descended from the original founder, we have enrolled 1784 in a comprehensive ongoing clinical monitoring study. Of 1181 genotyped participants, 459 are mutation carriers and 722 are non-carriers of the mutation.4

The importance of this pedigree is highlighted by the recent decision of the National Institutes of Health (NIH) to launch the first prevention trial for AD, with the Paisa PSEN1 pedigree as one of the principal focuses. Along with the presence of exhaustive and detailed comprehensive medical records of thousands of individuals, this pedigree originated from a founder effect that makes it a valuable resource for genetic research,2, 5 and for the development of biomarkers for predicting and following up the natural history of AD.6

Although the median age of AD age of onset (ADAOO) in patients with the E280A Paisa mutation is 49 years, the ADAOO extends widely from the early 30s to the late 70s.4 We considered that the high variability in ADAOO of this pedigree could include patients with an extreme phenotype (i.e., group of signs or symptoms departing from the disease’s natural history that would manifest in patients with an extreme early or late ADAOO (e.g., 30s vs 70s, respectively),7, 8, 9 supporting the hypothesis that the variance in ADAOO is influenced by major modifiers. Since members of this pedigree share a relatively homogeneous environment and culture, we hypothesized that these modifiers of ADAOO could be genes shaping the natural history of cognitive decline. Indeed, a previous analysis by our group using a genome-wide association study approach found that genome-wide intronic variants significantly associated as ADAOO modifiers.10 Other approaches using whole-genome sequencing have suggested some rare variants as potential modifiers of the ADAOO in this pedigree.5

To challenge the ADAOO variance, we scrutinized functional variants distributed through the whole exome in 71 PSEN1 E280A mutation carriers, all of them descendants from the original Paisa pedigree founder. Subsequently, we evaluated some of these functional variants in 93 PSEN1 E280A mutation carriers, and in a set of patients with sporadic AD (sAD) from the Paisa genetic isolate. Here, we disclose evidence that mutations harbored in APOE (apolipoprotein E), a gene implicated in the susceptibility and modification of AD risk, and additional new loci in GPR20, TRIM22, FCRL5, AOAH, PINLYP, IFI16, RC3H1 and DFNA5, might modify the ADAOO and therefore substantially change the natural history of this condition. Furthermore, this oligogenic model exhibits substantial sensitivity and specificity to predict the ADAOO.

Materials and methods


E280A pedigree

Detailed clinical assessment and ascertainment procedures of this pedigree have been presented elsewhere.4, 11, 12, 13 Briefly, we have collected data from participants including clinical evaluations, family history, comprehensive neurological and neuropsychological examinations, functional MRI during face–name associative memory encoding and novel viewing and control tasks, and structural magnetic resonance imaging. Clinical, neurological and neuropsychological assessments at the Group of Neurosciences AD Clinic used a Spanish version of the CERAD (Consortium to Establish a Registry for Alzheimer’s Disease) evaluation battery14 adapted for the cultural and linguistic characteristics specific to this population4, 11, 12, 13 (described in detail in the Supplementary Material). Patients were defined as affected by mild cognitive impairment based on the Petersen’s criteria and as AD if the DSM-IV (Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition) criteria were met.15, 16 The Ethics Committee of the University of Antioquia approved this study.

From the 459 E280A mutation carriers, we ascertained 71 patients from the extremes of the ADAOO distribution (44 women (62%) and 27 men (38%)) (Supplementary Figure 1a). The ADAOO (mean±s.d.) was 47.8±5.8 years in these patients (Supplementary Figure 1a). Mean ADAOO did not differ significantly by gender (female: 47.6±6.1; male: 48.4±5.5, P=0.55) (Supplementary Figure 1b). A total of 43 patients (26 women (60%) and 17 men (40%)) had an ADAOO below 48 years.10 As intended, mean ADAOO differed significantly between patients with ADAOO 48 and <48 years (44.2±2.48 vs 53.5±5.13, P=1.49 × 10−10). In those individuals with available information (n=57), years of education ranged from 0 to 19 years: four patients (7%) never attended school, 28 (49%) finished primary school (grades 1–5), 21 (37%) finished high school (grades 6–11, inclusive) and only 4 (7%) had tertiary education. Mean ADAOO did not differ significantly across education groups (F3,53=2.721, P=0.053) (Supplementary Figures 1c and d).

Cohort of sporadic cases

We assembled an independent sample of 128 sporadic sAD cases recruited from the metropolitan area of Medellin, Antioquia, Colombia, to test whether variants associated with ADAOO in the E280A pedigree were also associated with ADAOO in this sample. Although these individuals do not carry the PSEN1 E280A mutation, several population genetic analyses have shown that the community inhabiting this area has not been subject to microdifferentiation and shares the same genetic background and genealogy as the population with the E280A mutation.2, 3 Therefore, our rational is that given this similar genetic background, the genetic load modifying gene effects predisposing to AD is common. As in the E280A pedigree, neurological and neuropsychological assessments of these patients with sAD were conducted at the Group of Neurosciences AD Clinic using the modified CERAD evaluation battery (Supplementary Material).

Fifty-four patients, placed at the extremes of the ADAOO distribution (43 women (80%) and 11 men (20%)), were selected for this study from our collection of 128 patients with sAD (Supplementary Figure 2a). We chose them because the lower ADAOO in these patients had a similar distribution to that of patients with the E280A mutation. Similarly, the upper limit was defined to keep a similar distance between the average ADAOO and the upper limit of the E280A cohort. The average ADAOO was 63.26±6.94 years (Supplementary Figure 2a) in the 54 patients, with no statistically significant differences by gender (females: 63.83±6.29; males: 62.4±6.06, P=0.52) (Supplementary Figure 2b) or education group when including individuals with at least 1 year of education (F2,46=2.61, P=0.08) (Supplementary Figure 2c). The number of years of education ranged from 0 to 18 years: 1 patient had no information, 1 (2%) never attended school, 22 (42%) completed primary school, 23 (43%) completed high school and 7 (13%) attended tertiary education (Supplementary Figure 2d).


Whole-exome genotyping

One hundred and eleven individuals (57 with AD from the E280A pedigree and 54 individuals with sAD) were whole-exome genotyped using Illumina HumanExome BeadChip-12v1_A. This SNP-chip covers putative functional exonic variants selected from over 12 000 individual exome and whole-genome sequences, and consists of ~250 000 markers representing diverse populations (including European, African, Chinese and Hispanic individuals), and a range of common conditions such as type 2 diabetes, cancer, metabolic and psychiatric disorders. In addition to pure exonic variation, the HumanExome BeadChip-12v1_A (Illumina, San Diego, CA, USA) chip covers single-nucleotide polymorphisms (SNPs) in splice sites, in stop variants, in promoter regions and genome-wide association study tag markers, among other potentially functional variation. Samples with calls below Illumina’s expected 99% SNP call rates were excluded.

Whole-exome capture

We performed whole-exome capture on 14 individuals with AD from the E280A pedigree. Genomic DNA was extracted from peripheral blood from all patients and processed by the Australian Genome Facility (Melbourne, VIC, Australia). DNA libraries were constructed from 1 μg of genomic DNA using an Illumina TruSeq Genomic DNA Library Kit (Illumina, San Diego, CA, USA), and libraries were multiplexed with six samples pooled together (500 ng each). Exons were enriched from 3 μg of pooled library DNA using an Illumina TruSeq Exome Enrichment Kit (Illumina, San Diego, CA, USA), and ran on a 100 base pair paired-end run on an Illumina HiSeq 2000 sequencer (Illumina). A total of 201 071 genomic regions (sampled at ~50 × coverage) were surveyed using the whole-exome capture platform.

Sequencing image data were processed in real time using Illumina’s Real-Time Analysis Software and converted to FASTQ files using the CASAVA pipeline (Illumina). The entire workflow of data curation and analysis for variant calling was developed by the Genome Discovery Unit at The Australian National University, and consists of the following key components: (i) quality assessment; (ii) read alignment; (iii) local realignment around the known and novel indel regions to refine indel boundaries; (iv) recalibration of base qualities; (v) variant calling; and (vi) assigning quality scores to variants. The resulting FASTQ files were further processed for variant analysis using Golden Helix’s® SNP variation suite (SVS) 8.3.0 (Golden Helix, Bozeman, MT, USA).

Genetic, statistical and bioinformatics analyses

Quality control, filtering and classification of functional variants

After importing the genetic data to Golden Helix’s SVS 8.3.0, a single genetic data file was constructed by merging common and uncommon exonic variants from both the whole-exome genotyping and whole-exome capture platforms. Genotypes for 71 individuals from the E280A pedigree were obtained and quality control subsequently performed using the following criteria: (i) deviations from Hardy–Weinberg equilibrium with P-values <0.05/m (where m is the number of markers included for analysis); (ii) a minimum genotype call rate of 90%; (iii) presence of two alleles (i.e., we excluded monoallelic markers, and markers that were present in more than two alleles). Markers not meeting any of these criteria were excluded from analyses. Genotype and allelic frequencies were estimated by maximum likelihood. Following previous recommendations,17 variants with a minor allele frequency (MAF)0.01 were classified as common and as rare otherwise.

Exonic variants with potential functional effect were determined using the functional prediction information available in the dbNSFP_NS_Functional_Predictions GRCh_37 annotation track.18 This filter uses SIFT, PolyPhen-2, Mutation Taster, Gerp++ and PhyloP19, 20, 21 and is fully implemented in the Golden Helix SVS 8.3.0 Variant Classification module. This module was also used to examine interactions between variants and gene transcripts to classify variants based on their potential effect on genes. Variants were classified according to their position in a gene transcript. In addition, variants in coding exons were further classified according to their effect on the gene’s protein sequence.

Genome-wide association study analysis of common and rare variants

We studied the association of common exonic functional variants (CEFVs) to ADAOO using single- and multilocus additive, dominant and recessive linear mixed-effect models (LMEMs)22 with up to 10 steps in the backward/forward optimization algorithm. The advantage of these models is the inclusion of both fixed (genotype markers, sex and years of education) and random effects (family or population structure), the later to account for potential inbreeding by including a kinship matrix (i.e., the identity-by-descent matrix, which in our case was estimated between all pairs of individuals using markers excluded from the final analysis after linkage disequilibrium pruning). A single-locus LMEM assumes that all loci have a small effect on the trait, whereas a multilocus LMEM assumes that several loci have a large effect on the trait.22 Both types of models are implemented in SVS 8.3.0. The optimal model was selected using a comprehensive exploration of multiple criteria including the Extended Bayes Information Criteria, the Modified Bayes Information Criteria and the Multiple Posterior Probability of Association. After the estimation process using the forward/backward algorithm is finished, the coefficients was performed for the ith CEFVs to obtain the corresponding P-value (i=1,2,…,m). Thus, the collection P1, P2,…,Pm of P-values were subsequently corrected for multiple testing using the false discovery rate (FDR),23 and a method based on extreme-values theory.24 Because the tests of hypothesis being performed are of the same type, this correction is to be performed on the resulting m P-values only.23 Using a type I error probability of 5%, any FDR-corrected P-value 3.62 × 10−2 is considered statistically significant. This threshold was derived after correcting all m P-values using the p.adjust function in R.25 Similarly, any raw P-value 1.26 × 10−6 is statistically significant after Bonferroni’s correction. For the exploratory analysis, only CEFVs located in genes modifying ADAOO in the E280A pedigree were included for association analysis in the cohort of sporadic cases using LMEMs.

For the analysis of rare exonic functional variants (REFVs), regression- and permutation-based kernel-based adaptive cluster methods were used.26 Kernel-based adaptive cluster, implemented in SVS 8.3.0, catalogs rare variant data within each of a number of regions into multimarker genotypes, and, as variants are rare, only a relatively few different multimarker genotypes are found in any given region. A special test is subsequently applied to determine their association with the (case–control) phenotype, weighting each multimarker genotype by how often that genotype was expected to occur according to both the data and the null hypothesis that there is no association between that genotype and the case–control status of the sample.26 Thus, genotypes with high sample risks are given higher weights that can potentially separate causal from non-causal genotypes. Further, a one-sided test was applied because of the weighting procedure and the P-values were estimated using 10 000 permutations. Individuals with ADAOO 48 years were defined as cases and as controls otherwise. This cutoff value was selected based on previous studies of the ADAOO in the E280A pedigree.4, 10

Genomic/clinical-based predictive framework with ARPA

We used advanced recursive partitioning approach (ARPA) to construct a predictive decision tree-based model of AD status (ADAOO <48 and 48 years) in our patients with PSEN1 E280A AD using functional genetic variants and other clinical factors.6, 27, 28 Gender, years of education and CEFVs identified as ADAOO modifiers were used as predictors. ARPA offers fast solutions to reveal hidden complex substructures and provides non-biased statistical analyses of high dimensional seemingly unrelated data, and is widely used in predictive analyses as it accounts for nonlinear hidden interactions better than alternative methods and is independent of the type of data and of the data distribution type.29 ARPA was applied using the Classification and Regression Tree (CART), Random Forest (RF) and TreeNet modules implemented in the Salford Predictive Modeller software suite (Salford Systems, San Diego, CA, USA).

CART is a nonparametric approach whereby a series of recursive subdivisions separate the data by dichotomization.30 The aim is to identify, at each partition step, the best predictive variable and its best corresponding splitting value while optimizing a splitting criterion. As a result, the data set is successfully split into increasingly homogeneous subgroups.30 We used a battery of different statistical criteria as splitting rules (including the Gini index, Entropy and Twoing) to determine the splitting rule mostly decreasing the relative cost of the tree while increasing the prediction accuracy of target variable categories. The best split at each dichotomous node was chosen by either a measure of between-node dissimilarity or an iterative hypothesis testing of all possible splits to find the most homogeneous split (lowest impurity).30 Similarly, we used a wide range of empirical prior probabilities to model numerous scenarios recreating the distribution of the targeted variable categories in the population.30 Subsequently, each terminal node was assigned to a class outcome. To avoid overfitting in the CART predictive model, and to ensure that the final splits were well substantiated, tree pruning was applied. During this procedure, predictor variables that were close competitors (surrogate predictors with comparable overall classification error to the optimal predictors) were pruned to eliminate redundant commonalities among variables, thus the most parsimonious tree had the lowest misclassification rate for an individual not included in the original data.30 The final CART predictive model was selected based on the performance measures presented in Supplementary Table 4.

RF was conjointly applied with a bagging strategy to identify exactly the most important set of variables predicting AD status.31 Unlike CART, RF uses a limited number of variables to derive each node while creating hundreds to thousands of trees, and has proven to be immune to overfitting.31 In the RF strategy, variables that appeared repeatedly in trees as predictors were identified, and the misclassification rate was recorded. Finally, TreeNet was used as a complement to CART and RF strategies because it reaches levels of accuracy that are usually not attainable by either of the other two.32 This algorithm generates thousands of small decision trees built in a sequential error-correcting process that converge to an accurate model.32 Cross-validation training with all data and then indirectly testing with all the data was performed to derive honest assessments of the derived models and have a better view of their performance on future unseen data (i.e., review the stability of results across multiple subsets).30 To do so, the data were randomly divided, with replacement, into 10 separate partitions (folds). The used TreeNet (Boosting) decreases both the bias and the variance of the learning process, resulting in statistics that are immune to either inflation or overfitting. As described in the Results section, TreeNet corroborated the results of CART in both the learning and test data sets with slightly higher performance measures. This indicates a substantial predictive power of our ARPA-based CART predictive framework for identifying E280A mutation carriers with early and late ADAOO.

Pathway and network analyses

To identify key physiological pathways and networks involving genes harboring those variants disclosed by the common variant analysis, and to evaluate overrepresentation of common either ontogenetic or cellular processes, we performed network and pathway enrichment analyses with MetaCore® version 6.20 build 66481 (Thomson Reuters, New York, NY, USA). Genes with potential functional effect were examined with the ‘Analyse Network’, ‘Process Networks’, ‘Shortest Paths’ and ‘Direct Interactions’ algorithms. This collection of analyses provides a heuristic interpretation of maps and networks and rich ontologies for diseases based on the biological role of candidate genes. The presence of artifacts in statistical analyses (which can arise from genes in the database that may be in the same network but have no functional connection or interaction with any gene from our filtered list) was minimized by only including nodes with direct physical interactions between the encoded proteins in the database (known as the high trust set).


Quality control and genetic population structure

After quality control, assembling and filtering process, a total of 49 191 common and rare variants with potential functional effects remained for genetic analyses (Figure 1a). We estimated genetic stratification (population subdivision) using the Fst statistic of S Wright. The estimated Fst value for these cohorts was of 0.0187 (there is formal consensus that Fst values >0.17 are correlated with microdifferentiation). Further, we estimated the kinship coefficient of relatedness. Pairs of fAD and sAD cases with Fst values >0.025 were discarded.

Figure 1
figure 1

(a) Filtering process applied to exonic variants. Filter 1 includes common and uncommon variants between the Illumina’s HumanExome-12V1_A BeadChip and the whole-exome capture (see Patients and methods). Filter 2 excludes variants with a genotype call rate <90%, in Hardy–Weinberg disequilibrium and with one or more than two alleles. Filter 3 excludes variants with minor allele frequency (MAF) <1% and Filter 4 excludes nonfunctional variants. A total of 49 191 variants (39 753 common and 9438 rare) with potential functional effects remained for genetic analyses. (b) Average Alzheimer’s disease (AD) age of onset (ADAOO) by APOE allele combination. (c) Average ADAOO (blue dot) as a function of the APOE alleles. The red and gray lines represent the ADAOO of 48 years and ±1.5 s.d., respectively, the latter calculated using nonparametric bootstrap with B=10 000 replicates. Analysis of variance (ANOVA) showed that the average ADAOO differs among allele groups (F4,88=4.74, P=0.00163). (d) Effect of the presence/absence of the APOE*E2 allele on ADAOO. A two-sample t-test indicates that the presence of this allele increases ADAOO by ~8.2 years (95% confidence interval (CI): 4.45–12.01, P=3.84 × 10−5) in presenilin-1 (PSEN1) E280A mutation carriers.

PowerPoint slide

CEFVs modifying ADAOO

From the 39 753 CEFVs (Figure 1a), on average, ~2.83 CEFVs located within each gene (Supplementary Figure 4), we identified the rs7412, APOE*E2 allele as a whole-exome-wide ADAOO modifier using a single-locus LMEM. This marker delays ADAOO by ~12 years (β=11.74, 95% CI: 8.07–15.41, P=6.31 × 10−8, PFDR=2.48 × 10−3) (Table 1a).

Table 1a Results of the association analysis for ADAOO in 71 patients with PSEN1 E280A Alzheimer’s disease

Effect of APOE*E2/E4 alleles on ADAOO

Given that rs7412 (APOE) is an ADAOO modifier in our patients with PSEN1 E280A AD, we genotyped the rs429358 (APOE) marker to subsequently determine the effect of the APOE haplotype variants, defined by the E1/E2/E3/E4 alleles (see Figure 1b), on ADAOO. A total of 93 individuals were genotyped. On average, the presence of the APOE*E2 delayed the ADAOO by 8.24 years (95% CI: 4.45–12.01, P=3.84 × 10−5) when compared with its absence (Figures 1c and d).

A multilocus LMEM with nine steps in the backward/forward optimization algorithm was selected as the optimal model best explaining the ADAOO variance (>90% in total; Supplementary Figure 3a). Nine mutations harbored in APOE (rs7412, P=5.44 × 10−35), GPR20 (rs36092215, P=3.36 × 10−26), TRIM22 (rs12364019, P=8.78 × 10−19), FCRL5 (rs16838748, P=8.79 × 10−14), AOAH (rs12701506, P=7.26 × 10−12), PINLYP (rs2682585, P=2.55 × 10−10), IFI16 (rs62621173, P=1.54 × 10−9), RC3H1 (rs10798302, P=3.80 × 10−8) and DFNA5 (rs754554, P=8.32 × 10−6) were significantly associated as ADAOO modifiers (Table 1a and Supplementary Figure 3b and c). Variant rs12701506, located in the AOAH gene, is an intronic SNP anchored in a site encoding a strong enhancer (State-5) according to the chromatin state segmentation from ChiP-seq data. On the other hand, variant rs10798302 is located in an intergenic region close to the RC3H1 gene, and is anchored in a CpG Island, DNaseI Hypersensitivity Uniform Peak according to the ENCODE/Analysis.

Exploratory analysis in a cohort of patients with sAD

From the 247 874 exonic variants available for genetic analysis in the sAD cohort, 17 variants were located within the genes significantly associated with ADAOO in the E280A cohort. A multilocus LMEM minimizing the Modified Bayes Information Criteria and Extended Bayes Information Criteria criteria and maximizing the Multiple Posterior Probability of Association criterion was selected. A CEFV harbored in GPR20 (rs34591516, P=2.34 × 10−3) was found to modify ADAOO in our cohort of patients with sAD (Table 1b). This modifier effect was subsequently confirmed using a single-locus LMEM (β=−21.68, s.e.β=6.96, P=3.1 × 10–3, PFDR=0.058).

Table 1b Findings in 54 patients with sporadic Alzheimer’s disease

Rare exonic functional variants modifying ADAOO

A total of 9438 functional rare variants were available after quality control, assembly and filtering process. Regression-based kernel-based adaptive cluster analysis disclosed nominal associations between transcripts in the PDZD2 (P=1.80 × 10−2) and ATM (P=2.4 × 10−2) genes, as well as a borderline nominal association with transcripts in C10orf12 (P=5.0 × 10−2) (Supplementary Table 1a). Permutation-based kernel-based adaptive cluster confirmed the associations in PDZD2 (P=1.40 × 10−2) and ATM (P=1.40 × 10−2), and revealed a borderline nominal association between rare variants within SDK2 (P=5.29 × 10−2) and ADAOO (Supplementary Table 1b). These results were corroborated by using optimized Sequence Kernel Association Test (Supplementary Table 1c).

ARPA-based clinical diagnostic tool

A three-level tree with six terminal nodes was derived by CART, to identify E280A mutation carriers with late ADAOO (48 years) and early ADAOO (<48 years). Validation of this predictive model via RF and TreeNet produces comparable results (see below). Splitting nodes involved years of education and variants rs2682585 (PINLYP), rs7412 (APOE), rs16838748 (FCRL5) and rs62621173 (IFI16) (Figure 2a).

Figure 2
figure 2

(a) Classification tree for predicting late- (LO) and early-onset (EO) Alzheimer’s disease in E280A mutation carriers. Numbers in gray represent the split number, and N the sample size within each node. (b) Variable importance (left) and receiver operating characteristic (ROC) curve (right) for the Classification and Regression Tree (CART), Random Forest and TreeNet strategies. (c) Performance measures for the learning (blue) and test (pink) data sets for each model (b, right panel). AUC, area under the curve; CI, confidence interval; CR, classification rate.

PowerPoint slide

The presence of two copies of the A allele in rs2682585 identifies 50% of the E280A carriers with late ADAOO (node 4, n=24). In the second split, E280A carriers with the A/A genotype in rs2682585 and six or fewer years of education were mostly identified as having late ADAOO (node 5, n=20, 64.5%), whereas attending 6 or more years of education classified them as having early ADAOO (node 6, n=13, 76.5%). In the third split, the presence of the C/C genotype in rs62621173 classified most of the E280A carriers with late ADAOO (terminal node 5, n=20, 74.1%), whereas the presence of the C/T genotype in rs62621173 classified all of the E280A carriers with early ADAOO (terminal node 4, n=4, 100%) (Figure 2a).

Subsequently, individuals having one or two copies of the G allele in rs2682585 were mostly identified as having early ADAOO (node 2, n=23, 82.6%) (Figure 2a, left). In split 4, these number of copies of the G allele in rs2682585 and the C/C genotype in rs7412 (APOE) correctly classified all E280A carriers with early ADAOO (node 3, n=20), whereas the T/C genotype discriminated 75% of those individuals with late ADAOO, suggesting a potential gene × gene interactions between PINLYP and APOE to modify the delaying effect of the APOE*E2 allele on the ADAOO in PSEN1 E280A mutation carriers (Figure 1d). Finally, split 5 identified the remaining individual with late ADAOO based on the G/T genotype in rs16838748 (FCRL5) (Figure 2a, bottom).

The variable importance and receiver operating characteristic (ROC) curves for the CART, RF and TreeNet strategies are shown in Figure 2b. Although similar results were obtained with all strategies, CART included fewer variables than RF and TreeNet. Overall, these results show that (i) the top 5 variables included in the final model are comparable among strategies; and (ii) years of education and the genotype in rs2682585 (PINLYP) and rs7412 (APOE) provide the most consistent set of variables for differentiating patients by ADAOO.

Figure 2c displays the performance measures for the testing and learning data sets (see Supplementary Tables 4a and b for more details). For the learning data set, CART estimated an area under the curve of 84.9 (95% CI=75.7–92.6), classification rate of 84.5 (95% CI=76.1–91.5), sensitivity of 85.8 (95% CI=71.4–96.8), specificity of 83.8 (95% CI=72.1–93.9) and precision of 90.1 (95% CI=80.0–97.7), with overlapping 95% CI for the learning data set based on 10-fold cross-validation (Figure 2c). TreeNet corroborated these results in both the learning and test data sets with slightly higher performance measures than those produced by CART. On the other hand, RF in the testing data set produced similar point estimates for sensitivity and precision, and overlapping CIs to those from CART and TreeNet for other performance measures. Altogether, these measures indicate substantial predictive power of these genetic variants along with years of school attendance for identifying E280A mutation carriers with early and late ADAOO when combined in an ARPA-based CART predictive framework (Figure 2a).

We also have used the age of onset as an interval variable and prediction for the best fitting generalized boosting regression model was attempted. We found very similar patterns of variant inclusion and prediction when using the ADAOO as a continuous variable instead of one dichotomized into binary classes (see Supplementary Figure 7).

Pathway enrichment analysis

The pathway, network and enrichment analysis (using the 'Shortest Paths’ algorithm) disclosed statistically significant involvement of APOE, TRIM22, IFI16, RC3H1 and DFAN genes, interacting with PSEN1 in important physiological pathways, and gene ontology (GO) processes of apoptosis and immunological response, as well as in diseases such as AD (P=3.34 × 10−5), early-onset AD (P=1.68 × 10−2), delirium, dementia, amnestic and cognitive disorders (P=3.13 × 10−5), frontotemporal lobar dementia and degeneration (P=4.3 × 10−4), and neurodegenerative diseases (P=5.85 × 10−4) (Figure 3 and Supplementary Table 2).

Figure 3
figure 3

Resulting network involving genes harboring Alzheimer’s disease age of onset (ADAOO) modifier mutations (red dot) in presenilin-1 (PSEN1) E280A Alzheimer’s disease. Here, the 'Shortest Path’ algorithm was used. B, binding; C, cleavage; gray, unspecified; green, positive/activation; red, negative/inhibition; TR, transcription/regulation.

PowerPoint slide


Given the modest outcomes of the common disease-common allele hypothesis,33 new genomic approaches are needed in complex genetic disorders. Recently, a comprehensive alternative approach has emerged, which assesses genetic risk in terms of nonlinear interactions among genetic variants of major effect (i.e., functional mutations).27, 34, 35, 36, 37, 38 For this approach, the identification of extreme phenotypes in patients ascertained either from extended and multigenerational pedigrees or homogeneous cohorts from genetic isolates is recommended.2, 10, 39, 40

In this manuscript, we demonstrate the effectiveness of this strategy by identifying an oligogenic model comprised of functional variants of major effect harbored in the APOE, GPR20, TRIM22, FCRL5, AOAH, PINLYP, IFI16, RC3H1 and DFNA5 genes. Despite the relatively small sample size of our cohort, we have demonstrated the predictive efficiency of this model (Figure 2c) to delineate the ADAOO in E280A mutation carriers when clinical and demographic data are used in an ARPA-based classification tree. Thus, it is very unlikely that this tree might be the result of overfitting as using other techniques such as RF and TreeNet, which are immune to unbalanced group sizes, resulting in the same qualitative solution. Given the excellent performance of this tree in terms of sensitivity, specificity and precision (Figure 2c), we believe that this predictive framework can be used not only for predicting the ADAOO in the E280A pedigree but also for monitoring patients with AD at follow-up visits in future clinical trials. Having previously shown that imaging approaches such as 1HMRI could predict the onset of symptoms in the same pedigree,6 the next step would be to combine genomic, imaging and clinical data in an integrative predictive framework.

The characterization of APOE variants as one of the main drivers of the ADAOO model deserves further consideration. Indeed, preliminary analyses of this cohort were in apparent conflict with our current results. The first study using data from 31 E280A mutation carriers of the Paisa pedigree found nonsignificant effects of APOE variants on the ADAOO.41 The second study expanded the number of patients with the E280A mutation to 52 and found that carriers of the APOE ɛ4 allele were more likely to develop AD at an earlier age than non-carriers (hazard ratio=2.07; 95% CI=1.07–3.99; P<0.03), and that the APOE*E2 allele had a modest statistically nonsignificant ADAOO decelerator effect.42 By increasing the sample size to 93 patients with AD carrying the E280A mutation, we now show that individuals carrying the APOE*E2 allele develop AD at a later age. We also found that individuals carrying the APOE*E4 allele displayed a nonsignificant trend to develop AD at an early age. Our calculations show that this sample size is sufficient to detect a small-to-large effect size with >90% power for a type I error probability of 5% (Supplementary Figures 5 and 6).

The resulting network from the pathway and enrichment analysis shown in Figure 3 also includes the ESR1, SP1, CASP1 and UBC genes, and the p53 tumor suppressor protein, all of which interact with some of our ADAOO modifier genes. In particular, p53 is activated by IFI16, and positively regulates both DFAN5 and TRIM22. A mutation of either of these genes may affect this pathway and result in the upregulation of p53 in AD.43

CASP1, an essential component of NLPR3 inflammasomes, encodes a cysteine protease that is largely overexpressed in brains of patients with AD.44 One of the key biological functions of CASP1 is the cleavage of prointerleukin-1β (IL-1β) into the mature biological active cytokine, which is largely overexpressed in brains of patients with AD.44 Preclinical studies support the concept that CASP1-directed increase of IL-1β can augment the deposition of β-amyloid within the brain to worsen Alzheimer’s symptoms.44 Thus, it has been recently hypothesized that the NLRP3 inflammasome could be a potential target to treat AD.44 Interestingly, IFI16, the product of the ADAOO accelerator gene IFI16 identified in our study, binds and positively coactivates CASP1, which could trigger the inflammatory cascade to worsen Alzheimer’s symptoms.45 On the other hand, IFI16 interacts with and inhibits SP1 interaction to its cognate binding sites on DNA.46 It is noteworthy that SP1, a transcription factor that regulates the expression of several amyloid and tau-related genes,47 has been found to be abnormally expressed in the frontal cortex and hippocampus of patients with AD.48 Hence, mutations in IFI16 could result in the exacerbation of the activation of the inflammatory cascade orchestrated by CASP1, and/or the increased transcriptional activity of SP1 controlling the expression of key genes that are involved in the pathogenesis of AD. Our pathway and enrichment analysis also suggests that IFI16 activates ESR1, which in turn leads to the positive regulation of APOE whilst SP1 positively regulates APOE and PSEN1 (although these mechanisms are not yet well understood). In search for therapeutic targets, our data support that ESR1, SP1 and CASP1 could be important alternatives. Additional relevant GO processes involving at least two of the ADAOO modifier genes reported here are presented in Supplementary Table 3.

TRIM22 is an E3 ubiquitin ligase that is a member of the tripartite motif (TRIM) family, which includes three zinc-binding domains, a RING finger, a B-box type 1 and a B-box type 2, and a coiled-coil region. Although there are no reports relating TRIM22 with AD, TRIM11, another member of this family, was shown to bind to and destabilize humanin, a neuroprotective peptide that suppresses AD-related neurotoxicity.49 Interestingly, RC3H1 also encodes a protein that has an amino-terminal RING-1 zinc-finger, which is characteristic to that of the E3 ubiquitin ligase family.50 In the pathway enrichment analysis (Figure 3), it is predicted that both TRIM22 and RC3H1 bind to UBC, which in turn bind to PSEN1 to cause inhibitory effects. In previous studies it has been shown that presenilin proteins can be ubiquitinated.51, 52 Moreover, there is evidence suggesting that presenilin ubiquitination can cause reduction of endoproteolisis, which in turn reduces the formation of two fragments (presenilin N- and C-terminal fragments) that are essential for the γ-secretase activity.51 Thus, mutations in these genes may affect PSEN1 ubiquitination, thereby potentially leading to protein degradation and favoring β-amyloid and hyperphosphorylated Tau deposition.

GPR20 was not modeled in the pathway enrichment analysis shown in Figure 3. GPR20 belongs to the family of the G-protein-coupled receptors, which participate in intracellular second messenger systems triggered by different hormones and neurotransmitters, and amyloid β.53, 54, 55 GPR20 is expressed in AD brain.56

Three novel variants modifying the ADAOO are located in FCRL5, AOAH and PINLYP. Despite not being modeled in the pathway enrichment analysis and the lack of reported evidence of their possible relationship with AD, these genes encode key products having relevant functions within the immune system, which seems to influence several neurodegenerative and mental diseases.57, 58 FCRL5 encodes Fc receptor-like 5 that regulates B-cell antigen receptor signaling.59 Even though the expression of FCRL5 within the brain has not yet been reported, other Fc receptors were found to be expressed in microglia, the brain-resident macrophages.60 AOAH encodes acyloxycyl hydrolase, which is an enzyme expressed in antigen-presenting cells that deacylates and inactivates endotoxin.61 Finally, PINLYP encodes phospholipase A2 inhibitor and LY6/PLAUR domain containing. Although there is no much functional evidence reported about this gene product, it is well known that phospholipase A2 releases arachidonic acid, which is a substrate used for the synthesis of potent proinflammatory factors such as prostaglandins. Thus, it can be hypothesized that phospholipase A2 inhibition caused by the PINLYP product could favor an anti-inflammatory effect.

Limitations of this study include the fact that the oligogenic model was derived from a very unique form of fAD, and therefore it might only be applicable to carriers of the PSEN1 mutation. However, the identification of one of these associated variants in patients with sAD shows a common gene effect predisposing to AD to that in the E280A pedigree, and suggests that these variants could be real major players modifying the natural history of the illness. The analysis of other AD cohorts from around the world would be the next focus of our research.

In summary, we have defined major mutations modifying ADAOO in members of a multigenerational extended family carrying the PSEN1 E280A mutation. One of the modifier genes reported herein was also associated with the ADAOO in sporadic cases from the general population. This suggests that the ADAOO modifier effect is not unique to the Paisa pedigree; it may be a general finding applicable to other forms of AD. Of major importance is the highlighting of the APOE*E2 allele as an ADAOO decelerator. This finding is consistent with recent research,62 and suggests that PSEN1 and APOE may interact to modify ADAOO in these patients. Furthermore, using a subset of these variants, we constructed an accurate predictive framework to characterize AD patients in terms of early or late onset that can be used as a diagnostic tool during clinical assessment. Finally, the pathway enrichment analysis suggests that cell proliferation, protein degradation, apoptotic and dysregulation of immune processes are implicated in the variable onset of AD.