Novel genetic profile linked to cognitive decline in Hispanics/Latinos

Abstract

BACKGROUND

The APOE ε4 allele has weaker association with ADRD outcomes in Hispanic/Latino populations compared to non‐Hispanic White and East Asian populations.

METHODS

We used whole‐genome sequencing data from 13,025 Hispanic/Latino adults to infer population groups (clusters) from short DNA sequences at the APOE gene region. We estimated the associations between these clusters and mild cognitive impairment (MCI) and age‐adjusted significant cognitive decline assessed via the Study of Latinos‐Investigation of Neurocognitive Aging (SOL‐INCA) study (n = 5334). We compared these associations to those of the conventional APOE alleles.

RESULTS

We identified three clusters of individuals. Relative to the largest Cluster 1, being in Cluster 3 had higher likelihood of having significant cognitive decline, with odds ratio (OR) = 1.47, 95% confidence interval (CI) [1.18, 1.84], p < 0.001. The association with having prevalent MCI at the SOL‐INCA exam was not statistically significant, with OR = 1.33, 95% CI [0.97, 1.83], p = 0.08. Associations of the ε4 allele with these traits were not statistically significant.

DISCUSSION

The new APOE profile has stronger risk association with age‐adjusted significant cognitive decline in SOL‐INCA Hispanics/Latinos compared to the ε4 allele.

Highlights

• We clustered Hispanics/Latinos using short genomic sequences from the APOE region.

• One cluster had higher prevalence of age‐adjusted significant cognitive decline.

• APOE ε4 and ε2 alleles were not associated with significant cognitive decline.

• This approach may explain genetic heterogeneity of APOE allele associations.

1. BACKGROUND

The strongest known genetic risk factor for Alzheimer's disease (AD) is the ε4 allele of the apolipoprotein E (APOE) gene, while the relatively rare ε2 allele has a protective effect. ¹ , ² The APOE gene is also strongly associated with other health outcomes including cardiovascular disease and longevity. ³ Studies have demonstrated that APOE alleles’ associations with AD and related cognitive outcomes vary by sex, age, race, and ethnicity, ⁴ , ⁵ as well as genetic ancestry. ⁶ For instance, evidence from the Study of Latinos–Investigation of Neurocognitive Aging (SOL‐INCA) suggests that APOE ε4's effects on cognitive outcomes differ across Hispanic/Latino heritage groups and are related to levels of continental genetic ancestry, with individuals with high proportions of Amerindian (indigenous American) ancestry having a weaker effect of the APOE ε4 alleles. ⁷ A meta‐analysis covering 20 studies of Hispanic/Latino populations confirmed an association between APOE ε4 and increased risk of Alzheimer's disease and related dementias (ADRD) and showed that the strength of this association varied across different Hispanic/Latino heritage groups. ⁸ Another study from HCHS/SOL suggests that while ancestry‐enriched genetic variants (including Amerindian ancestry‐enriched variants) are present in the APOE region, there are no variants with large effect sizes interacting with APOE ε4 on mild cognitive impairment (MCI) in the APOE region, underscoring the complexity of genetic influences in this population. ⁹ Some previous efforts, including of Caribbean Hispanics from the National Institute on Aging's Late‐Onset AD and the National Cell Repository for AD Family study, admixed Brazilians from the Biobank for Aging Studies (BAS) from the University of São Paulo Medical School (Brazil), and Puerto Rican and African American AD cases and controls from multiple sources, successfully showed APOE alleles’ interaction with genetic ancestry in the APOE region but did not identify specific genetic variants driving this interaction. ¹⁰ , ¹¹ , ¹² One study used data from nine studies of African American participants, whole‐genome sequencing (WGS) from Ibadan individuals from Nigeria, all having some levels of African genetic ancestry and identified an African‐enriched variant modifying the association of the APOE ε4 allele with AD. ¹³ While the search for ancestry‐specific variants implies the existence of relatively rare variants interacting with APOE, it is possible that more common genetic structures explain the genetic ancestry‐related heterogeneity of the APOE alleles’ associations with AD and related outcomes. For example, polygenic risk scores (PRSs) constructed from AD GWASs have been shown to predict MCI in Hispanic/Latino older adults, with stronger associations observed when including variants from the APOE region – yet not the conventional APOE alleles themselves – highlighting the potential importance of other, potentially common, APOE variants in this population. ¹⁴

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional sources. Prior work demonstrated that the associations of apolipoprotein E (APOE) ε4 and ε2 alleles with ADRD traits differed by race/ethnicity and by genetic ancestry. However, this heterogeneity is not yet fully explained by known genetic measures. Relevant publications are appropriately cited.

2. Interpretation: We applied a recently developed method to use short genomic sequences (of length 21 in primary analysis) on data from the HCHS/SOL. Short sequences from an APOE‐surrounding genomic region were used to separate Hispanic/Latino individuals into three groups (clusters). Individuals in each cluster had similar genetic profiles, defined by these short sequences. Individuals in one of these clusters were more likely to have age‐adjusted significant cognitive decline compared to individuals in the other clusters. The association was weaker or did not exist in clusters based on control regions.

3. Future directions: Identification of specific genetic sequences and variants enriched in the cluster with higher levels of significant cognitive decline, and assessment of their associations with ADRD traits in other Hispanic/Latino populations.

Population structure analysis has traditionally relied on model‐based and statistical approaches using single nucleotide polymorphisms (SNPs). ¹⁵ Hispanic/Latino populations demonstrate rich genetic diversity reflecting complex ancestral contributions from European, Amerindian, and African populations. ¹⁶ , ¹⁷ While current methods for estimating population structure have advanced our understanding of genetic diversity in general and in Hispanics/Latinos specifically, novel methods may yield additional information regarding variation overall as well as in functionally important regions like that of the APOE gene.

Recent methodological advances offer new opportunities to understand these population‐specific patterns. Hrytsenko et al. have developed an alignment‐free approach using k‐mers – short DNA sequences of fixed length k – to infer population structure. ¹⁸ This method captures local genetic variation by analyzing the frequencies of these DNA substrings across the genome, providing a different view of genetic diversity without requiring predefined markers or specific genetic assumptions. Here, we leveraged whole‐genome sequencing (WGS) data from Hispanic Community Health Study/Study of Latinos (HCHS/SOL) and applied this k‐mer‐based approach to characterize k‐mer profiles and cluster individuals with similar genetic patterns specifically in the APOE region. We hypothesized that k‐mer‐based analysis in the APOE gene region would reveal distinct, genetically based, clusters among Hispanic/Latino adults that reflect their diverse genetic ancestral compositions around the APOE region. We further hypothesized that this structure would differ from patterns captured by genetic principal components and from the APOE alleles. More specifically, we hypothesized that APOE allele (ε2, ε3, and ε4) frequencies would differ across the detected clusters and that the new clusters of individuals will be relevant for ADRD (i.e., they would be associated with MCI and age‐adjusted significant cognitive decline in SOL‐INCA).

2. METHODS

2.1. The study of Latinos‐investigation of Neurocognitive aging

The Hispanic Community Health Study/Study of Latinos (HCHS/SOL) is a longitudinal cohort study established in 2006 by the National Heart, Lung, and Blood Institute (NHLBI). The first wave or recruitment took place in the period 2008 to 2011, where 16,415 Hispanic/Latino adults aged 18 to 74 were recruited by four field centers in the metropolitan areas of Bronx, New York, Miami, Florida, Chicago, Illinois, and San Diego, California. Individuals were recruited to the study via a two‐stage sampling design, stratified according to socioeconomic status, with sampling units defined according to high concentrations of Hispanic/Latino surnames. ¹⁹ , ²⁰ The HCHS/SOL data coordinating center developed weights to be used in statistical analyses so that proportions, means, and association effect estimates were representative of the HCHS/SOL target population of Hispanics/Latinos in the four metropolitan areas. Individuals took surveys in either English or Spanish, based on their own preference. A review of various measures available in HCHS/SOL and its contribution to cardiovascular research is available. ²¹

At the baseline HCHS/SOL exam, 9593 participants who were 45 or older responded to four cognitive tests, including a six‐item screener, ²² a Brief‐Spanish English Verbal Learning Test (B‐SEVLT), ²³ a word fluency (WF) test, ²⁴ and a Digit Symbol Substitution (DSS) test. ²⁵ The second main HCHS/SOL exam took place between 2014 and 2017. It did not include cognitive testing; however, an ancillary study to HCHS/SOL called the Study of Latinos‐Investigation of Neurocognitive Aging (SOL‐INCA) was established and recruited 6377 HCHS/SOL participants who were at least 50 years old during the second exam. ²⁶ Weights were developed for the SOL‐INCA study as well to reflect the specific target population of SOL‐INCA and account for participation patterns in the ancillary study. The SOL‐INCA exam occurred 7 years, on average, after the baseline cognitive exam. SOL‐INCA administered the same cognitive tests as were applied during the HCHS/SOL baseline visit and expanded the battery of tests to include tests on additional cognitive domains (Trail Making Test A and B ²⁷ ), self‐reported measures of cognitive decline (Everyday Cognition‐12 [ECog12] ²⁸ ), and Instrumental Activities of Daily Living (IADL) scale ²⁹

Based on these tests, SOL‐INCA investigators developed measures of cognitive aging. ²⁶ Cognitive decline (change) is a continuous measure of global cognitive change developed based on a latent change score model and relying on cognitive test scores in the baseline and SOL‐INCA exams. The latent construct (global cognitive score) was based on four tests (B‐SEVLT sum, B‐SEVLT recall, WF, and DSS, called manifest variables in latent class analysis jargon) measured at both the baseline and SOL‐INCA exams. Latent factor scores at each time point were estimated from the latent factor model after establishing longitudinal invariance. Subsequently, factor scores were used to calculate cognitive decline using a regression‐based change approach whereby (1) global cognition at V2 was regressed on global cognition at V1, (2) residualized V2 cognition was calculated as the difference between observed and predicted V2 global cognition from this model, and (3) the residual was divided by the model's root mean squared error (RMSE) to generate a standardized change index ³⁰ (z‐score), with negative values indicating higher‐than‐expected cognitive decline. ²⁶ , ³¹ Significant cognitive decline, which defined substantial cognitive decline in the context of a community‐based cohort (and does not refer to a statistical measure of significance), was operationalized as a binary measure, defined as a change that exceeds −0.055 standard deviation (SD) yearly in global cognitive function, as measured according to the cognitive decline variable. Significant cognitive decline is a core Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, and International Classification of Diseases, Tenth Revision, feature of dementia and of MCI. ³² MCI was defined according to the National Institute on Aging‐Alzheimer's Association criteria ³³ at the time of derivation, requiring (1) self‐reported cognitive decline according to the ECog12 questionnaire; (2) significant cognitive decline and a low score for any single cognitive test relative to SOL‐INCA internal norms and after accounting for age, sex, education, and for the picture vocabulary test ³⁴ ; and (3) no or minimal functional impairment according to the IADL questionnaire. We included in this MCI group individuals categorized as MCI+, that is, suspect cognitive impairment. Note that while our significant cognitive decline measure reflects a change from the baseline to the SOL‐INCA exams, the MCI outcome is prevalent at the SOL‐INCA exam, and not incident.

2.2. Whole‐genome sequencing

DNA extracted from blood samples of 13,025 HCHS/SOL participants was sequenced via two projects: the Trans‐Omics for Precision Medicine (TOPMed), supported by NHLBI, and the Centers for Common Disease Genomics (CCDG), supported by the National Human Genome Research Institute. WGS for HCHS/SOL took place at the Baylor College of Medicine Human Genome Sequencing Center (HGSC). WGS followed TOPMed protocols, as described at the link https://topmed.nhlbi.nih.gov/topmed‐whole‐genome‐sequencing‐methods‐freeze‐9. Alleles were then called for HCHS/SOL together with another study that participates in TOPMed and CCDG (the Atherosclerosis Risk in Communities study), also at the Baylor College of Medicine HGSC, using Illumina's DRAGEN pipeline. ³⁵ The WGS data are provided in genome build hg38. To compute genetic PCs and genetic relatedness over this dataset, PC‐AiR ³⁶ and PC‐Relate ³⁷ were used, implemented in the GENESIS R package. ³⁸ Based on the generated kinship matrix, we applied the pcairPartition function (also in the GENESIS R package) to identify a maximal set of unrelated individuals (degree 3), defined as those with kinship coefficient less than 2^−9/2 $\approx$ 0.044.

2.3. Characterization of global and local genetic ancestry in HCHS/SOL

To assess potential connection between the new APOE measures and admixture patterns, we utilized a previously developed measure of local admixture ³⁹ and developed a new measure of global admixture. HCHS/SOL individuals are primarily admixed with three ancestral populations: European, African, and Amerindian. A few (<20) HCHS/SOL individuals were also previously shown to have substantial levels of Asian ancestry. ¹⁶ Here, we estimated global ancestry proportions (i.e., proportions of genome inherited from European, African, Amerindian, and Asian ancestry for each individual) by first adding a dataset of 1473 WGS of reference Asian individuals from public datasets (1000 Genome, ⁴⁰ the Human Genome Diversity Project [HGDP] ⁴¹ ), and the Simons Genome Diversity Project [SGDP] ⁴² ). These individuals were added in order to have sufficient data to prevent miscalibration due to outlier effects (where outliers could be the handful of HCHS/SOL participants with substantial Asian ancestry). Proportions of ancestry were estimated using ADMIXTURE software. ⁴³ Because ADMIXTURE requires unrelated individuals, we subsetted the data multiple times to separate datasets, where each such dataset included unrelated individuals, but these separate datasets overlapped and all of the HCHS/SOL individuals with WGS were included in at least one dataset. We applied ADMIXTURE on each such dataset. As a quality control, we checked and verified that ancestry proportions obtained by separate analyses for the same individual resulted in similar values (all such values were <0.01 from each other, for all ancestries) and that ancestry proportions for European, African, and Amerindian ancestries were similar, with a correlation >0.99, to those previously obtained on a subset of HCHS/SOL individuals. ¹⁶ Local ancestries (i.e., counts of alleles inherited from European, African, and Amerindian ancestry in a given genomic position for each individual) were previously estimated using RFMix ⁴⁴ for the set of individuals with chip genotyping data ³⁹ and were shown to have a performance similar to that of local ancestry inference performed using a more recent approach on a large subset of these individuals. ⁴⁵

2.4. K‐mer analysis in APOE gene surrounding region

Figure 1 illustrates the k‐mer analysis procedure that we applied. This section describes the first step. We subsetted the WGS data and generated new Visual Contact Files limited to the variants between the APOE region start and end sites (chr19:44,905,781–44,909,394 in GRCh38) and an additional 500,000‐base‐pair flanking region on either site (overall region chr19:44,405,781–45,409,394). This region covers the coding regions of additional, non‐APOE genes, including TOMM40 and APOC1, which are also suggested to be involved in ADRD. ⁴⁶ , ⁴⁷ , ⁴⁸ We next phased the data used Beagle ⁴⁹ version 5.2, using a genetic map from chromosome 19 provided on the Beagle software website. For each individual and chromosome, we extracted their two APOE alleles from the phased data. Next, we indexed the phased data using the faidx indexing command in samtools ⁵⁰ and then generated FASTA sequences for each individual using BCFtools, ⁵⁰ combining reference genome sequences with the individual‐specific variants, generating two chromosomal sequences for each individual covering the APOE region and the flanking regions. We next applied Jellyfish, ⁵¹ a tool for k‐mer counting, to derive k‐mer profiles for the individuals in the dataset. A k‐mer profile is a list of all k‐length sequences appearing in the phased genetic file and their counts. We selected a k‐mer length of 21 nucleotides, mimicking the published work. ¹⁸ Using Jellyfish, we used canonical k‐mer counting (‐C parameter), which combines reverse complements, and required a minimum frequency threshold of 2 (‐L 2 parameter), filtering out singleton k‐mers that appear only once in an individual.

FIGURE 1.

Illustration of k‐mer‐based procedure to derive genomic profiles in APOE region. The figure illustrates the process used to identify sets of individuals with similar genetic patterns in the APOE gene region. In step 1, we converted variant format data into sequence data, which we used to derive k‐mer profiles (set of k‐mers and their counts) for each individual in the dataset. Step 2 demonstrates how sets of individuals were identified: We first aggregated the individual k‐mer profiles into a single table, retaining only k‐mers with counts varying across individuals. We applied PCA on this dataset, and then applied k‐means clustering on these PCs. Not conveyed in the picture is two selections of sets of k‐mers, multiple clustering, and consensus clustering over the results. APOE, apolipoprotein E; PCA, principal component analysis; PC, principal component.

2.5. Derivation of k‐mer profile‐based clusters

We derived clusters from principal components (PCs) applied on the k‐mer profiles. A challenge arose from the fact that each individual's k‐mer profile was stored in a separate file, the set of k‐mers slightly differed between individuals due to biological variation, and the number of k‐mers is very large, so that merging all files while preserving all k‐mers in the data is computationally expensive. Thus, for computational efficiency, we performed an analysis to guide the selection of a smaller set of k‐mers to use in PC analysis (PCA). First, we assessed the variability of k‐mer counts across individuals, with the goal of selecting k‐mers that had some variability across individuals. To this end, we performed 100 random samples of 500 individuals each. For each sample, we listed all available k‐mer sequences and computed the SDs of the counts of these k‐mers in the sample. We summarized the information into four sets of k‐mers: those in which the SD of their counts was non‐zero in all 100 random samples of individuals and those in which the SD of their counts was non‐zero in at least 90%, 85%, and 80% of these samples. Note that these are nested lists, with the list of k‐mers that have non‐zero SD of their counts in at least 80 of the samples being the largest. Next, we considered each of these sets of k‐mer sequences and performed PCA on each. When performing PCA, we used only the maximal set of genetically unrelated individuals (N = 10,772) and then projected the PCs over the remaining individuals (N = 2253). Based on the heuristic “elbow” method (visual inspection of the “break” in the scree plot of PC variances), we selected a number of informative PCs for each of these sets of k‐mers. Finally, we applied k‐means clustering on selected sets of PCs and consensus clustering to aggregate the information across results from separate clustering analyses.

2.6. Characterization of APOE k‐mer count PCs and clusters

We assessed the population structure captured by the PCs of the k‐mer counts using density plots of PCs and scatter plots of pairs of PCs and coloring these figures, highlighting Hispanic/Latino heritage groups and groups defined by patterns of APOE alleles. We compared these patterns to those captured by standard genetic PCs.

To characterize the clusters in terms of genetic ancestry, we used the previously derived measures of global and local ancestry and visualized the distribution of global proportions of European, African, and Amerindian genetic ancestries across individuals in each of the clusters and the proportions of chromosomes in which the local ancestry at the APOE region was European, African, or Amerindian across each of the clusters’ participants.

2.7. Statistical analysis

We computed the frequency of APOE alleles in the APOE k‐mer clusters using genetically unrelated individuals. We estimated and compared the associations of APOE k‐mer clusters to the associations of conventional APOE alleles with cognitive aging outcomes. The cluster with the largest number of individuals was set as reference. For association analysis, we used the survey R package ⁵² , ⁵³ to perform survey logistic regressions of significant cognitive decline and MCI, adjusted for baseline age and education (three‐level variable: less than high school, high school, and more that high school), sex, study center, time between the baseline and the SOL‐INCA exam, and testing language. The survey regression accounts for the survey design of the HCHS/SOL (and, consequently, SOL‐INCA), that is, multi‐stage stratified sampling from a “target population,” so that the resulting estimators are applicable to (representative of) the target population and robust to potential correlations stemming from having individuals from the same census block unit and household potentially more similar to each other than individuals living in other census block units. Models were fit separately to the APOE clusters, where Cluster 1 served as a reference and coefficients were estimated for Clusters 2 and 3 and for the APOE alleles, where the APOE ε3 allele was the reference and coefficients were estimated for dosages of the ε2 and ε4 alleles. For each cognitive outcome, we recorded the Akaike information criterion (AIC) ⁵⁴ to compare model fit. In secondary analysis, we repeated the association analyses further adjusting for counts of African and Amerindian ancestry in the APOE region and adjusted for global proportions of African, Amerindian, and Asian ancestries.

2.8. Sensitivity analyses to k‐mer analysis

While we report as the main analysis the one following our original plan, focusing on the k‐mer methodology with k = 21 and the APOE surrounding region, we performed additional sensitivity analyses. First, we repeated the complete pipeline of the k‐mer analysis using k = 17 and k = 25 (i.e., reducing and increasing k’s length by 4). Second, we repeated the complete pipeline of the k‐mer analysis in three additional “control” genomic regions (all coordinates in GRCh38): (1) chr19:48,405,781–49,409,394 (original region shift by 4Mbp); (2) chr18:44,405,781–45,409,394; (3) chr17:44,405,781–45,409,394 (same coordinates as the original region surrounding APOE, on chromosomes 18 and 17). For the new regions we used k = 25 because the results in the APOE region were slightly stronger with k = 25. Third, we used the first k‐mer count PC as the variable in the association analysis with cognitive outcomes instead of clusters. The PC was derived from the k‐mer counts computed over the sets of k‐mers with SD > 0 in 80% and in 100% of the subsamples. This was done for all APOE and sensitivity regions. Fourth, we extracted the top k‐mer from PC1 of each region (i.e., the k‐mer with the highest PC loading) and used them individually in association analysis.

2.9. Code availability

Code for the k‐mer pipeline will become publicly available on the GitHub repository https://github.com/tamartsi/APOE‐kmers‐neurocog.

3. RESULTS

3.1. APOE k‐mer‐based population clusters

Across N = 13,025 individuals with WGS data, the number of distinct k‐mers per individual was about 2.4 million. Over 100 random samples of 500 individuals in each, there were 26,554 k‐mers with non‐zero SD in all samples and 31,040, 33,313, and 36,284 k‐mers with non‐zero SD in 95%, 90%, and 80% of the samples, respectively. We generated PCs based on the four sets of k‐mers defined by having non‐zero SDs in at least 100%, 95%, 90%, and 80% of the random samples, limiting the computation to genetically unrelated individuals (N = 10,772) and projecting the PCs on the remaining individuals (N = 2253). Figure 2 demonstrates that the distribution of k‐mer PCs somewhat varied by groups defined by APOE alleles, with the largest variability observed for PC1. Figure 2 uses the first six PCs generated based on the set of k‐mers with non‐zero SD in at least 80% of the random samples, Figures S1 to S3 demonstrate similar patterns for the other sets of k‐mers. Some structure is also observed when highlighting groups defined by self‐reported Hispanic/Latino heritage (Figure S4, focusing on the set of k‐mers with non‐zero SD in at least 80% of the random samples). To contrast this with standard genetic PCs, Figures S5 and S6 provide similar plots when using genetic PCs, showing much stronger structure by Hispanic/Latino heritage and less by APOE genotype.

FIGURE 2.

Principal components of APOE region k‐mer counts and APOE allele count group. Distribution and pair scatter plots of PCs of k‐mer counts. K‐mers were selected as those with SD of their counts being non‐zero in 80 of the 100 subsamples of HCHS/SOL individuals. APOE, apolipoprotein E; PC, principal component; SD, standard deviation; HCHS/SOL, Hispanic Community Health Study/Study of Latinos.

Because there was no reason to prioritize any sets of k‐mers of those examined, we separately applied k‐means clustering four times: over the first two and over the first six PCs generated from the set of k‐mers with non‐zero SD in at least 80% of the random samples and similarly over the first two and first six k‐mers with non‐zero SD in 100% of the random samples. For each of these clustering efforts we selected the k parameters, determining the number of clusters, by visually examining the scree plot. These scree plots are presented in Figure S7. We chose three clusters for each of these k‐means applications. Finally, we applied consensus clustering to generate a single cluster assignment for each individual, resulting in three APOE k‐mer clusters. Table 1 characterizes the study population (by providing sample sizes) and the target population (via weighted‐estimated means of continuous measures and proportions of categorical variables) stratified by the three clusters. We ordered the clusters by size, such that Cluster 1 had the largest number of individuals and Cluster 3 the smallest. Demographic variables often differed between clusters, likely relating to genetic patterns, which in turn, in this dataset, correlated with demographic heritage (see Figure 3 in Rao et al. ⁵⁵ demonstrating genetic ancestry similarity within Hispanic/Latino heritage groups) and patterns of migration to the United States. (For example, individuals self‐reporting with Cuban heritage tended to be older and more educated and more often were recruited in Miami; see Figure 2 in Rao et al. ⁵⁵ Such individuals are enriched in Cluster 1.)

TABLE 1.

Characteristics of SOL‐INCA target population by k‐mer cluster.

Variable	Cluster 1	Cluster 2	Cluster 3	Missing %
N (unweighted)	2457	1770	1107
Age at baseline (years)	55.27 (7.37)	54.58 (7.08)	55.61 (7.26)	0
Years between visit 1 and SOL‐INCA visit	6.90 (1.16)	7.04 (1.15)	6.90 (1.16)	0
Mild cognitive impairment (MCI)	246 (10.8%)	182 (9.9%)	133 (14.3%)	1.8
Significant cognitive decline	715 (29.1%)	517 (31.1%)	363 (37.4%)	0.1
Sex
Female	1521 (51.4%)	1157 (55.9%)	713 (55.8%)	0
Male	936 (48.6%)	613 (44.1%)	394 (44.2%)
Field center
Bronx	472 (20.3%)	249 (24.8%)	580 (53.7%)	0
Chicago	461 (8.6%)	605 (20.3%)	171 (6.4%)
Miami	912 (49.3%)	294 (21.4%)	258 (29.8%)
San Diego	612 (21.8%)	622 (33.5%)	98 (10.0%)
Education
No high school diploma or GED	891 (32.0%)	798 (42.9%)	447 (41.0%)	0.3
At most high school diploma/GED	530 (21.2%)	375 (22.7%)	250 (21.6%)
Greater than high school/GED	1033 (46.8%)	591 (34.4%)	406 (37.4%)
Hispanic/Latino heritage
Central American	192 (5.3%)	213 (9.7%)	94 (6.6%)	0.2
Cuban	683 (39.9%)	109 (10.1%)	169 (22.3%)
Dominican	167 (6.4%)	42 (3.2%)	312 (26.8%)
Mexican	770 (24.8%)	1047 (56.3%)	130 (11.7%)
South American	149 (3.6%)	153 (7.9%)	40 (2.5%)
Puerto Rican	440 (16.0%)	181 (10.0%)	333 (25.8%)
More than one/other heritage	54 (4.1%)	21 (2.8%)	25 (4.3%)

Note: Numbers are unweighted, while means and percentages are weighted with SOL‐INCA weights to provide estimates characterizing the SOL‐INCA target population. Hispanic/Latino heritage is self‐reported.

Abbreviation: SOL‐INCA, Study of Latinos–Investigation of Neurocognitive Aging.

FIGURE 3.

APOE k‐mer clusters and standard measures of ancestry. (A) Visualization of standard, genome‐wide, genetic PC 1 versus 2 for HCHS/SOL individuals, colored by their APOE k‐mer cluster. (B) Frequency (or proportion) of genomic intervals (two intervals, for each chromosomal copy, per person) overlapping with APOE gene region, associated with each of the HCHS/SOL continental ancestral populations: Europe, African, and America. (C) Violin plots of distribution of global ancestry proportions for individuals in each of the APOE k‐mer clusters. APOE, apolipoprotein E; HCHS/SOL, Hispanic Community Health Study/Study of Latinos.

3.2. APOE k‐mer cluster associations with other measures of ancestry and with APOE alleles

Figure 3 relates the APOE k‐mer clusters to standard genetic PC‐based measures of ancestry, proportions of global continental ancestries, and frequencies of local continental ancestries. All measures of ancestry differed by APOE‐based cluster. Comparing the clusters, Cluster 3 had the highest median value of European global ancestry, Cluster 1 had the highest median value of African ancestry, and Cluster 2 had the highest median value of Amerindian ancestry. However, local ancestry patterns differed between these genome‐wide averages. Cluster 1 was dominated by intervals associated with European ancestry (88% of the chromosomal copies across individuals); Cluster 2, agreeing with the global ancestry patterns, had the highest proportion of Amerindian ancestry intervals (64%), while Cluster 3 had a majority of African ancestry‐inherited intervals (61%). African local ancestry was very low ($\le 3\%$) in Clusters 1 and 2. Table 2 demonstrates that the frequency of APOE alleles differs between the clusters, with the frequency of both APOE ε4 and ε2 alleles being highest among Cluster 3 individuals, with minor allele frequency (MAF) = 0.18 for APOE ε2 (compared to 0.09 in Cluster 1 and 0.04 in Cluster 2), and MAF = 0.40 for APOE ε4 (compared to 0.20 and 0.22 in Clusters 1 and 2, respectively).

TABLE 2.

APOE alleles distribution across APOE‐based k‐mer clusters.

Cluster	N	APOE $\epsilon 2$ count	APOE $\epsilon 4$ count	APOE $\epsilon 2$ frequency	APOE $\epsilon 4$ frequency
1	5793	527	1112	0.091	0.192
2	4496	169	987	0.038	0.220
3	2736	478	1089	0.175	0.398

Note: APOE allele frequencies were estimated using unrelated individuals only.

Abbreviation: APOE, apolipoprotein E.

3.3. APOE k‐mer cluster associations with MCI and significant cognitive decline

We estimated the association between APOE k‐mer clusters and the SOL‐INCA cognitive aging measures and compared them with the associations of conventional APOE alleles (Table 3). Figure 4 presents the analytic sample used for association analysis. Figure 4A demonstrates the overlap between the WGS and SOL‐INCA datasets, while Figure 4B describes the exclusions leading from the SOL‐INCA ancillary study to the analytic dataset used for association analysis with significant cognitive decline and with MCI. Individuals in Cluster 2 did not differ in their rate of MCI and significant cognitive decline compared to individuals in Cluster 1. Individuals in Cluster 3 had non‐statistically significant 33% higher odds (95% confidence interval [CI: 0.97 to 1.83) of having MCI compared to those in Cluster 1 and a statistically significant association of 47% higher odds (95% CI: 1.18 to 1.84) of having significant cognitive decline. The conventional APOE alleles were not associated with MCI in these data. Each additional APOE ε4 allele (compared to the reference of two APOE ε3 alleles) was associated with 17% higher odds of significant cognitive decline (95% CI: 0.98 to 1.41). For both cognitive outcomes, the AIC, measuring model fit, was better (lower) for models that included the APOE k‐mer clusters as explanatory variables compared to models that included dosage of the conventional APOE alleles. In secondary analysis, we performed the same association analysis while further adjusting for African and Amerindian counts of local ancestry in the APOE region and adjusting for global proportions of African, Amerindian, and Asian ancestries (Asian ancestry, not contributing to most study individuals, was available only for the global ancestry analysis). Results are provided in Table S1, showing that effect sizes are essentially the same, even though p values are less significant in local ancestry‐adjusted analysis, likely due to the smaller sample size. 

TABLE 3.

Associations of APOE k‐mer clusters and APOE alleles with MCI and with significant cognitive decline.

APOE variable	Estimated odds ratio	95% CI	p‐value	Model AIC
Association with APOE k‐mer clusters with MCI
Cluster 2	0.92	(0.65, 1.30)	0.64	3597.87
Cluster 3	1.33	(0.97, 1.83)	0.08
Association with APOE alleles with MCI
APOE ε2	0.79	(0.53, 1.19)	0.26	3603.11
APOE ε4	1.00	(0.77, 1.30)	0.999
Association with APOE k‐mer clusters with significant cognitive decline
Cluster 2	1.09	(0.88, 1.36)	0.414	6401.23
Cluster 3	1.47	(1.18, 1.84)	< 0.001
Association with APOE alleles with significant cognitive decline
APOE ε2	1.09	(0.82, 1.45)	0.55	6418.01
APOE ε4	1.17	(0.98, 1.41)	0.08

Note: Analyses were adjusted for age at baseline, time between the baseline and the SOL‐INCA visit, sex, education level, and testing language. Analyses used survey regression procedures and were weighted using SOL‐INCA weights to generate estimates representative of the SOL‐INCA target population. The sample size for MCI analysis was n = 5236 and for analysis of significant cognitive decline it was n = 5317.

The sample size for the association analyses with MCI is n = 5236.

The sample size for the association analyses with significant cognitive decline is n = 5317.

Association models were adjusted for age at baseline, time between the baseline and SOL‐INCA visits, sex, study center, education level, and testing language.

Abbreviation: APOE, apolipoprotein E; MCI, mild cognitive impairment; SOL‐INCA, Study of Latinos–Investigation of Neurocognitive Aging.

FIGURE 4.

Sample sizes and derivation of the analytic datasets. (A) Visualization of overlap between the two subsets of HCHS/SOL data used: the WGS dataset and the SOL‐INCA ancillary study. (B) Derivation of analytic dataset used in association analyses, starting from SOL‐INCA ancillary study participants and excluding individuals with missing WGS, education, and cognitive outcomes of interest. HCHS/SOL, Hispanic Community Health Study/Study of Latinos; SOL‐INCA, Study of Latinos–Investigation of Neurocognitive Aging.

3.4. Sensitivity analysis for k‐mer analysis

We performed additional analyses changing the k parameter used to set the length of the k‐mers (k = 17, 25) and additionally considered other genomic regions as controls: a chromosome 19 shifted region, with the APOE coordinates shifted by 4 mega base‐pairs, and regions on chromosomes 18 and 17 with the same coordinates as the primary APOE‐surrounding region. The results are presented in Table S2, showing that association results were weaker when k = 17 and about the same when k = 25; association results were weaker (but present) in the APOE‐shifted control region (Cluster 3 OR = 1.33, p = 0.01) with significant cognitive decline, and associations with chromosome 18 and 17 control regions were null. Table S3 presents the results from association analysis using the top k‐mer count PC as the variable of interest instead of clusters derived from PCs of k‐mer counts. We used both our primary and control regions and PC1 computed based on two sets of k‐mers (those with non‐zero SD on 80 and 100% of the subsamples). One‐SD increase in k‐mer PC1 (k = 25, k‐mers set based on non‐zero SD in 80% of subsamples) in the APOE surrounding region had OR = 1.12, 95% CI of 1.03 to 1.22, p = 0.006 for significant cognitive decline. These results were consistent with PC1 of other k‐mer parameters, thought slightly weaker. K‐mer PC1 of the APOE‐shifted control region also had marginal association with MCI (OR = 1.13, 95% CI: 1.00 to 1.28) per 1‐SD increase in the PC. Finally, associations of the top k‐mer from the PC by region (k = 25 always) are presented in Table S4. While the k‐mers from the chromosome 18 and 17 control regions were not associated with cognitive outcomes, counts of the top k‐mer from the APOE‐shifted control region, AAGTGCTGGGATTACAGGCGTGAGC, were protectively associated at the nominal levels with MCI: OR = 0.93 per increase in 1 k‐mer dosage (p = 0.03) and with OR = 0.95 for significant cognitive decline (p = 0.05). Counts of the top k‐mer in the APOE region also had protective association with cognitive outcomes, but they were not statistically significant (p ≈ 0.10). ,  ,  

4. DISCUSSION

We used k‐mers, short genomic sequences representing genomic variation, to study genomic structure in the APOE gene region. Using cluster analysis, we derived three sets of individuals with differing genetic profiles in the APOE region. One of these clusters had a relatively strong association with age‐adjusted significant cognitive decline (OR = 1.47) compared to individuals in the reference cluster. In the conventional approach, APOE ε4 has a weak association with significant cognitive decline in this dataset, as was previously shown in HCHS/SOL using an earlier, smaller version of genetic data. ⁵⁶ This finding is important because it shows that there is an APOE structure that is more predictive of significant cognitive decline in Hispanics/Latinos, compared to the APOE ε4 allele. Prior work, motivated by the observed weaker APOE ε4 association with ADRD in Hispanics/Latinos and in high‐African‐ancestry individuals, focused on identifying variants that interacted with the APOE ε4 allele to result in this weaker association.

K‐mers are different from standard genetic variants, which are aligned to a reference genome to annotate specific positions and alleles. In comparison, k‐mers in the approach we implemented were derived from sequence data (also using a reference genome to “fill in the gaps” between variants), which may have different lengths for each individual in the data as it accounts for insertions and deletions (indels). Thus, k‐mers rely on the specific ordering of alleles on the same chromosome. K‐mers are being increasingly used in bioinformatics, including genome assembly and taxonomic classifications. K‐mers are also being used for cancer diagnosis, when applied on RNA sequences extracted from tumors. ⁵⁷ Notably, these applications of k‐mers are different than the approach we implemented; in most of these settings, k‐mers are used directly from raw sequence reads (prior to variant calling). K‐mer counting, as implemented in our approach, was performed on the genomic sequence after allele calling to reduce errors and sequencing artifacts. ¹⁸ This approach was only recently demonstrated to capture population structure, and this is likely the first application of k‐mers to identify structures within a pre‐specified genomic region for subsequent association analysis with an aging‐related disease.

There are a few limitations of the work with k‐mers, specifically, arbitrary choices that we had to make to balance computational needs. First, the length of the k‐mers was selected as in primary analysis 21, consistent with prior work. ¹⁸ In sensitivity analyses we also considered k = 17 and k = 25 and still every k that is pre‐specified is arbitrary. In principle, shorter k‐mer lengths would result in fewer k‐mers, and longer lengths would result in more, as the number of potential sequences is exponential in k: It is $4^k$ because there are four nucleotides and each can theoretically be at each of the k positions. Further, as we derive the individual genomic sequences using the reference genome to “fill in the gaps” between variants, many of the k‐mers are the same across individuals – consistent with the well‐known observation that most of the genome is the same in all humans. Thus, we implemented a sampling‐based approach to select a subset of k‐mers in which counts vary across individuals and proceeded with two potential choices of those considered (k‐mers with counts that show a non‐zero SD in all 100, and in 80, of the random samples of individuals – the two most extreme scenarios considered). Various choices can be made in this regard, but this is not quantitatively different from placing allele frequency thresholds on variants, as is done in most genetic analyses. Finally, we chose to cluster individuals based on PCs of k‐mer counts, using a k‐means algorithm. We performed clustering four times and applied consensus clustering to generate a single set of clusters. Consensus clustering identifies whether a set of people tend to cluster together, and if so, they are kept together in the final clusters. As no standards exist yet for k‐mer‐based inference of genome structure, we think that our approach is reasonable. Importantly, the resulting clusters indeed demonstrated differences with respect to conventional APOE alleles and associations with cognitive aging outcomes. Still, we also performed sensitivity analyses estimating associations with PC1 (the top PC) of the k‐mer counts and with counts of the k‐mer with the highest contribution to PC1 (highest PC loading). PC1 in the APOE region had a statistically significant association with significant cognitive decline, but the top k‐mer did not.

This work was also limited by the cognitive outcomes used. First, we did not have AD outcomes, where the APOE ε4 allele is the hallmark risk factor for AD. Second, our significant cognitive outcome criterion of decline in test performance at or exceeding −0.055 SD yearly could be different. This threshold was selected to match that used by the Atherosclerosis Risk in Communities (ARIC) study, ⁵⁸ but other thresholds could be used, in principle. It is also worth noting that we used significant cognitive decline as a criterion for the definition of MCI (as required by the National Institute on Aging‐Alzheimer's Association and DMS5 criteria), so that the two outcomes were not independent (most individuals with MCI are also categorized as having significant cognitive decline). While we used both outcomes because their APOE genetic determinants may have different contributions, it would be interesting to assess outcomes that capture more distinct features of cognitive aging, when available. Finally, because we did not have MCI outcomes ascertained at the baseline HCHS/SOL visit (as cognitive change information is not available), we only had a prevalent MCI outcome on the SOL‐INCA exam, rather than incident MCI.

An important question is whether these k‐mers could be used to trace information back to specific genetic variants. It is possible that a count of a specific k‐mer would be sufficient to explain the cluster association with cognitive aging outcomes. However, we only considered the top PC‐contributing k‐mer, and it did not have a statistically significant association with cognitive outcomes (however, the top k‐mer from the APOE‐shifted control region had a nominal association with cognitive outcomes). Considering all ∼35,000 k‐mers used by the PC analyses, the multiple testing burden is too high for our outcomes: We do not have sufficient power to conclude significance at a p value threshold that would properly account for this number of tests. Also, if a specific k‐mer were identified, it would be difficult to trace it to specific genetic variants and would require substantial modeling. K‐mer counts by themselves (rather than standard single nucleotide vaeiations) may be useful, but pipelines will need to be adapted to compute them in standard genetic analyses.

How can this work be extended and generalized? Opportunities lie in new data and modeling that are becoming available. For example, plasma AD biomarkers are becoming more available and can potentially serves as outcomes that, being continuous and available throughout life, may have more statistical power compared to MCI or cognitive decline in genetic analyses. ⁵⁹ Similarly, new artificial intelligence/machine learning methods such as genomic language models, ⁶⁰ which, like k‐mers, rely on genomic sequences, could be potentially trained or fine‐tuned using appropriate data to develop and explain predictions from the APOE region. Importantly, new models and analyses should use data from genetically diverse populations, including Hispanics/Latinos and individuals with African ancestry, given the known limitations of the conventional APOE alleles in these groups.

In summary, we implemented and evaluated a new approach to developing genetic parameters in a genomic region using k‐mers. We focused, in our primary analysis, on APOE‐surrounding region sequences and identified three clusters of individuals with similar k‐mer profiles. Individuals having one of these profiles (Cluster 3) had a higher likelihood of significant cognitive decline compared to individuals in the reference Cluster 1, and this relationship was stronger compared to the relationship of the APOE ε4 allele with these ADRD outcomes. More work is needed to evaluate this approach in other populations and to identify genetic variants that are easily ascertained and that determine these profiles and their associations with ADRD outcomes.

CONFLICT OF INTEREST STATEMENT

Dr. Ramos reports receiving grant support from Axsome Pharmaceuticals, consulting fees from Jazz Pharmaceuticals, honoraria from the Medical Education Speakers Network, and participation in the National Institutes of Health/National Heart, Lung, and Blood Institute's Sleep Disorders Research Advisory Board. Dr. HM González reports receiving consulting fees from University of Texas San Antonio Health Sciences Center Alzheimer's Disease Research Center, receiving a REMIND honorarium from UCI, and receiving support from the National Academy of Sciences, Engineering & Medicine, for attending a meeting. Dr. Tarraf reports having the role of a section editor for Alzheimer's & Dementia: The Journal of the Alzheimer's Association. All other authors report no conflicts of interest. Author disclosures are available in the Supporting Information.

DATA AVILIABILTY STATEMENTS

The HCHS/SOL fully supports data sharing for HCHS/SOL–approved manuscript proposals with outside investigators. All data sharing is conducted in accordance with HCHS/SOL study and NIH policies and governed by a Data and Materials Distribution Agreement (DMDA) between UNC and the external institution, ensuring the confidentiality and privacy of HCHS/SOL participants and their families. Alternatively, de‐identified HCHS/SOL data are publicly available at BioLINCC and dbGaP for the subset of the study cohort who authorized general use of their data at the time of informed consent. HCHS/SOL WGS data are available by application to dbGaP according to the study‐specific accession “phs001395”. Phenotype data are also available from dbGaP according to the study‐specific accession “phs000810”.

CONSENT STATEMENT

The HCHS/SOL was approved by the Institutional Review Boards (IRBs) at each field center, where all participants gave written informed consent, and by the Non‐Biomedical IRB at the University of North Carolina at Chapel Hill, to the HCHS/SOL Data Coordinating Center. The IRBs approving the HCHS/SOL study are as follows: Non‐Biomedical IRB at the University of North Carolina at Chapel Hill, Chapel Hill, NC; Einstein IRB at the Albert Einstein College of Medicine, Bronx, NY; IRB at Office for the Protection of Research Subjects (OPRS), University of Illinois at Chicago. Chicago, IL; Human Subject Research Office, University of Miami, Miami, FL; Institutional Review Board of San Diego State University, San Diego, CA. All methods and analyses of HCHS/SOL participants’ materials and data were carried out in accordance with human subject research guidelines and regulations. This work was approved by the Beth Israel Deaconess Medical Center Committee on Clinical Investigations.

Supporting information

Supporting information

Supporting information

Hrytsenko Y, Purushotham Y, Li Z, et al. Novel genetic profile linked to cognitive decline in Hispanics/Latinos. Alzheimer's Dement. 2026;22:e71855. 10.1002/alz.71855