On their own, most genetic variations linked to Alzheimer’s disease affect a person’s risk by a minuscule amount. Together, however, they pack a wallop, at least according to a follow-up study led by Rahul Desikan at the University of California, San Francisco, and published January 24 in Brain. In some 600 people without dementia, polygenic hazard scores correlated not only with their extent of amyloid and tau accumulation and cortical shrinkage, but also with their rate of cognitive decline over the following years. Importantly, the paper claims, a person’s polygenic score had predictive power above and beyond that offered by ApoE genotyping or neuroimaging, suggesting it might make a useful screening tool for trials. Curiously, the score also correlated with non-AD pathologies, including cerebrovascular disease and Lewy body inclusions. A commercial polygenic hazard score (PHS) is already being sold to the public.

  • Polygenic hazard score correlated with Aβ and tau accumulation, and cortical atrophy.
  • PHS associated with cognitive decline in non-demented people.
  • A model combining polygenic risk and biomarkers performed best.

Desikan and colleagues first described their PHS in 2017 (Mar 2017 news). This PHS is the sum of the weighted age-specific AD risk, compared with the general population, imparted by each single nucleotide polymorphism (SNP) at 31 locations in the genome. The 31 SNPs were picked because together, their genotypes correlated with the age of onset in participants in the Alzheimer’s Disease Genetics Consortium. Desikan’s 2017 study reported that the PHS correlated with low CSF Aβ42 and high total tau and, in autopsy tissue, with more plaques and tangles and with atrophy of the medial temporal lobe. The PHS correlated with cognitive decline in non-demented people, and the correlations held regardless of ApoE genotype.

In the new paper, the researchers expanded their in vivo analysis to include regional distribution of Aβ plaques by PET, as well as presence of non-AD pathologies such as Lewy bodies and cerebrovascular pathology upon autopsy. They also wanted to assess whether the PHS could improve on the diagnosis made by brain imaging alone. Drawing on data from 980 participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI), co-first authors Chin Hong Tan and Luke Bonham and colleagues reported that their PHS correlated with Aβ-PET standardized uptake value ratios across 34 brain regions. The associations held across the disease spectrum, including in cognitively normal people.

PHS Predicts Aβ, Atrophy. Polygenic hazard scores correlate positively with regional Aβ accumulation (left) and negatively with change in cortical volume (right). [Courtesy of Tan et al., Brain, 2018.]

How about neurodegeneration markers? As judged by longitudinal structural MRI data from 607 participants, their PHS correlated with atrophy. It did so most strongly in areas predisposed to neurofibrillary tangles, including the entorhinal cortex, inferior parietal cortex, and inferior and middle temporal cortices. The PHS correlated with atrophy even in people who were cognitively normal.

The researchers also checked the PHS against neuropathological data from a separate cohort. Among 485 postmortem brain samples in the Religious Orders Study and Memory and Aging Project (ROSMAP), the PHS correlated with the extent of Aβ pathology; once again, this was independent of ApoE genotype. Checking the PHS against tau pathology, the scientists found it correlated with neurofibrillary tangle burden across all brain regions of interest regardless of ApoE genotype, except for the hippocampus, where the correlation held only in ApoE4 carriers.

But does a person’s PHS associate with cognitive decline? To address this, the authors tapped serial cognitive data and baseline neuroimaging measures from 632 ADNI volunteers, of whom 229 were cognitively normal and 403 had mild cognitive impairment. The PHS correlated with slippage on tests of executive function, memory, and dementia—the latter as measured by the CDR sum of boxes (CDR-SB). The associations remained after controlling for frontal Aβ load, entorhinal cortex volume, and other covariates including ApoE genotype, sex, age, and years of education, suggesting PHS contributed independently to cognitive decline. Cognitively normal people with a high PHS declined faster on the CDR-SB than did people with a low PHS, the authors report.

A diagnostic model made of Aβ-PET, entorhinal cortex volume, and the PHS better captured cognitive decline than a model without the PHS, suggesting to the authors that the PHS adds clinical predictive value beyond neuroimaging data.

Does this PHS capture non-AD pathology? For this question, the scientists tapped postmortem samples stored at the National Alzheimer’s Coordinating Center (NACC). Based on tissues from 603 people, the PHS correlated with three types of cerebrovascular pathology: cerebral amyloid angiopathy, atherosclerosis, and hemorrhages and microbleeds. It also correlated with the presence of Lewy body pathology.

Importantly, it did not correlate with tangles in non-AD tauopathies, including frontotemporal lobar degeneration (FTLD) or progressive supranuclear palsy. It also did not correlate with TDP-43 pathology in people with FTLD.

Why would a genetic risk score for AD predict certain non-AD pathologies? The researchers think this reflects overlap between genetic risk factors for Alzheimer’s and those for cerebrovascular and Lewy body diseases (Broce et al., 2019; Dec 2017 news). Alternatively, AD might be a risk factor for these other morbidities. The pathologies often occur together, with recent ROSMAP data implicating AD pathology as an instigator of cardiovascular and Lewy body pathology (Bennett et al., 2018). 

Tan, who is now at Nanyang Technological University in Singapore, told Alzforum that the PHS could come in handy as a tool for clinical trials. For therapies targeting Aβ or tau, the PHS would enrich trials for people likely to harbor those pathologies, meaning fewer people would undergo costly PET scans or invasive lumbar punctures at screening. The researchers recently reported that 60 percent of people with a PHS in the top quartile would test positive for Aβ accumulation, compared with 40 percent of people with a PHS in the bottom quartile. The PHS could also enrich trials for people most at risk of cognitive decline, Tan added. The ability of Aβ positivity to predict clinical AD rose from 75 percent in people with a low PHS to 100 percent in people with a high PHS (Tan et al., 2018). 

Other polygenic scores have already proved useful in gauging AD risk. To generate such scores, scientists lower the significance threshold in AD GWAS, allowing them to pool the AD risk associated with thousands of SNPs beyond the few dozen top hits. In 2015, researchers led by Valentina Escott-Price of Cardiff University, U.K., reported that collective associations of 29,000+ variants with AD accounted for far more AD heritability than just the 21 most highly associated variants. Indeed, a polygenic risk score (PRS) based on these thousands of SNPs accounted for 84 percent of the variance in risk of pathologically confirmed AD (Escott-Price et al., 2015; Escott-Price et al., 2017). Similarly, researchers led by Elizabeth Mormino at Massachusetts General Hospital in Boston reported that a PRS incorporating genotypes from more than 16,000 AD GWAS SNPs correlated with cognitive and structural brain changes that precede the onset of dementia, as well as with risk of AD (Jul 2016 news). Mormino is now at Stanford University.

Unlike Desikan’s PHS, however, these versions of PRS did not estimate age-specific risk. The SNPs used to calculate PRS are selected based on their association with AD cases compared with controls in GWAS, while the 31 SNPs used to calculate PHS were selected based on their combined correlation with age at onset in a separate cohort. In a 2018 review, Tan and Desikan argued that because AD is a continuous, age-dependent process rather than a binary state, PHS might be more useful in the clinic than PRS (Tan and Desikan, 2018). 

This argument was tested in a paper published February 19 in Annals of Clinical and Translational Neurology. Escott-Price and collaborators compared the PRS and PHS head-to-head. Using AD GWAS data, along with age-at-onset data from the Genetic and Environmental Risk for Alzheimer’s Disease (GERAD) cohort, the researchers reported that whether SNPs were selected based on age-specific hazard or overall AD risk, collectively they correlated with age at onset similarly well. For PRS, only SNPs that associated with AD with a p value below 10-3 contributed to these age-specific predictions (Leonenko et al., 2019). The researchers added that while PRS and PHS performed similarly in their study, PHS could be advantageous if age-at-onset data were available in a giant AD GWAS, thus allowing the prioritization of SNPs that associate with age at onset on a genome-wide level.

While multiple research groups are still exploring the polygenic contribution to AD risk, a commercial version of the PHS is already being sold directly to consumers. Two of Desikan’s co-authors, Chun-Chieh Fan and Anders Dale of the University of California, San Diego, though not Desikan himself, worked with the San Diego-based healthcare company HealthLytix to develop a PHS for public use. HealthLytix then teamed up with San Francisco-based Dash Genomics to market the product. For $119, customers who have previously submitted their DNA to 23andMe or Ancestry.com can upload their raw data file to the Dash Genomics website. After calculating a PHS, Dash discloses the customer’s odds of developing AD within a year, the age at which they will face a 50 percent chance of showing signs of the disease, and how the customer’s AD risk measures up to population averages at a given age. The company began offering the service in August 2018, but declined to disclose to Alzforum how many customers have used it so far. This company advertises on Alzforum’s Virtual Exhibit Hall

UCSD’s Fan, who is also a chief scientist at HealthLytix, said that the proprietary PHS he and Dale helped develop is different than the 31-SNP PHS described in their scientific papers. Instead of relying on the 31 variants with a demonstrated link to AD age at onset, the commercial PHS weaves in genotypes at the 600,000 SNP positions that are included in the raw data files from 23andMe and Ancestry.com. The researchers then correlate genotypes at these positions with age at onset in a separate cohort. The vast majority of these SNPs have no relationship with AD risk, while some have minuscule contributions. Fan claims that pooling them leads to a closer approximation of AD onset age than the original PHS published by Desikan et al.

Fan’s statement, and the differences between the academic and the commercial PHS, have not been published or subjected to peer review. Efforts to reach Dale for comment were unsuccessful. Desikan told Alzforum he is not involved in helping these companies derive their version of the PHS. However, he noted that incorporating more genotype information should result in an improved PHS, and he plans to use a similar, more extensive version in future research (see Desikan Q&A below).  

Ancestry.com and 23andMe disclose to their customers a limited list of specific genotypes—such as ApoE4—that are validated for accuracy. However, the vast majority of the 600,000+ genotypes in a customer’s raw data file generated by these direct-to-consumer companies are not validated, and this file is the one from which HealthLytix generates its PHS (Jul 2017 news). Fan told Alzforum that before calculating a customer’s PHS, they culled their list of 600,000 SNPs by weeding out those known to be prone to error, or “miscalls” by 23andMe and Ancestry.com.

Multiple commentators noted that while PHS was useful in research, they consider it not ready for commercial use. “There is no consensus among the scientific community on how PHS can be used for clinical management, much less for DTC tests,” Rita Guerreiro of the Van Andel Institute in Grand Rapids, Michigan, wrote to Alzforum (full comment below). The prediction accuracy and predictive utility are not clear, wrote Escott-Price (full comment below). “The diagnostic and prognostic accuracy of the PHS are not yet known. For these and other reasons, it is premature to share PHS scores with consumers,” agreed Jessica Langbaum of Banner Alzheimer’s Institute in Phoenix (full comment below). Carlos Cruchaga of Washington University in St. Louis raised similar points, adding that PHS may explain only a small fraction of AD risk. For his part, Desikan is comfortable with the marketing of the PHS to the public, but he also noted that genetic susceptibility for complex conditions should not be viewed in isolation, but rather as one component along with environmental and lifestyle factors, that together drive risk.

John Hardy of University College London, further cautions that the predictive value of PHS is especially weak in non-Caucasian populations. “At the moment, essentially all the data we have is on northern European Caucasians, and predictive values in other populations will be unreliable because we do not have the background genetic data to make the predictions from,” he wrote to Alzforum. The Dash Genomics website mentions this in its “Limitations” section. —Jessica Shugart


Q&A with Rahul Desikan

Q: Do you feel comfortable that a modified version of the PHS is being sold to the public via Dash Genomics?

A: Yes.

Q: Do you think the way Dash is calculating PHS—i.e., incorporating data from 600,000 SNPs instead of just 31—is an accurate and technically valid approach? 

A: I like to think about PHS in terms of versions. What we did in our original paper was the best we could do given the technology at the time. Now we can do better. PHS 2.0 performs better than the original version, because you are taking the whole genome into account and using annotations.

Q: Did HealthLytix or Dash tap into any of the genotyping, neuroimaging, and/or clinical data from the cohorts you used in your papers to validate their version of the PHS? 

A: I can’t tell you what Dash or HealthLytix did because I’m not involved with either. I can tell you scientifically that we will be using the PHS 2.0 going forward because it gives better results. On a personal note, the folks at HealthLytix are good friends and top-notch scientists.

Q: Do you think the benefits of learning PHS outweigh any downsides to the customer?

A: Like any new test, polygenic scores have to undergo rigorous evaluation before being adopted clinically. Polygenic scores must be tested for their ability to measure absolute risk for a single person from the general population. Provided a collection of DNA samples, polygenic scores could retrospectively assess whether response to therapy is conditional on genetic risk in existing drug trials. Polygenic scores should be examined for their ability to prospectively predict clinical outcomes (e.g., time to progression to dementia) in asymptomatic individuals. With PHS we have shown that we can get at absolute risk and prospective prediction in older people who are cognitively normal.

Finally, genetic susceptibility for complex conditions should not be viewed in isolation but as another risk factor for disease, and combined with lifestyle and environmental factors in multivariate evaluation of disease risk. With PHS we are doing this right now and assessing how genetic information can be integrated with other markers.


  1. There is no consensus among the scientific community on how PHS can be used for clinical management, much less for DTC tests. One can easily find experts who contest the accuracy and utility of PHS, particularly regarding their utility for personalized screening. There is no regulation or public education framework (that I know of) specific for PHS, so I think it is still very early to make such tests available to consumers, even if all limitations and disclosures are made available on websites and reports.

  2. The PHS is associated with amyloid PET and various markers of neurodegeneration and may be useful in a research setting to track disease progression.

    The diagnostic and prognostic accuracy of the PHS are not yet known. For these and other reasons, it is premature to share PHS scores with consumers. For instance, the SNP results from various direct-to-consumer labs may not be accurate, leading to miscalculation of the PHS. Even if the calculation of the PHS is accurate, it is just a risk factor and not an indication that someone will (or will not) develop the disease.

    Just as people without the e4 variant of APOE still develop dementia due to Alzheimer’s, those with a low PHS may still develop the disease. As with other genetic testing results, it is important to work with a genetic counselor or other healthcare professional to discuss family, insurance, and emotional considerations of learning results.

  3. Tan and colleagues report an association of polygenic hazard score (PHS) with numerous factors, including Aβ and tau accumulation, neurodegeneration, cognitive decline, etc. Prior to PHS analyses, Desikan et al. selected common genetic variants (SNPs) based on their association with AD at p-value≤10-5 in the publically available IGAP dataset (Lambert et al., 2013). They only use 31 SNPs, which mostly represent genome-wide AD associated loci. The effect sizes for age-specific, or “hazard” risk, of these SNPs are very similar to the general AD risk effect sizes. We have recently demonstrated that the PHS constructed as suggested by Desikan et al. is a shortened version of the polygenic risk score (PRS) (Leonenko et al., 2019) and, similar to PRS, is capable of predicting AD and related phenotypes over and above APOE (as shown in Escott-Price et al., 2015Escott-Price et al., 2017). 

    However, the prediction accuracy of the PHS was not reported in Desikan et al. and therefore the predictive utility of PHS is not clear. Furthermore, although APOE and GWAS SNPs show strong association with late-onset AD, the genetic heritability explained by these loci is not high (h2=5.1 percent [95 percent CI: 3.9 percent to 6.3 percent]) (Escott-Price et al., 2017) as compared with genome-wide estimates (h2=24 percent to 53 percent) (Ridge et al., 2016; Ridge et al., 2013; Lee et al., 2013). The maximum prediction accuracy which can be achieved based upon APOE and GWAS significant loci is limited (maximum prediction accuracy AUC=66 percent [95 percent CI: 64 percent to 67 percent]), and is insufficient for clinical applications or trials (Lewis and Vassos, 2017). 

    PRS as defined in Escott-Price et al., 2015 shows prediction accuracy of AUC= 75 percent to 84 percent in clinical and pathology confirmed samples, respectively (Escott-Price et al., 2015Escott-Price et al., 2017). These AUC estimates are very close to the maximum prediction accuracy that can be achieved based upon SNP-based heritability captured by the whole genome (Escott-Price et al., 2017), and may potentially be used for AD risk prediction with more confidence.

    In conclusion, a modified version of this PHS may have a limited predictive utility and low accuracy, and may be misleading for clinical trials and/or the general public.


    . Genetic assessment of age-associated Alzheimer disease risk: Development and validation of a polygenic hazard score. PLoS Med. 2017 Mar;14(3):e1002258. Epub 2017 Mar 21 PubMed.

    . Meta-analysis of 74,046 individuals identifies 11 new susceptibility loci for Alzheimer's disease. Nat Genet. 2013 Dec;45(12):1452-8. Epub 2013 Oct 27 PubMed.

    . Polygenic risk and hazard scores for Alzheimer's disease prediction. Ann Clin Transl Neurol. 2019 Mar;6(3):456-465. Epub 2019 Feb 18 PubMed.

    . Common polygenic variation enhances risk prediction for Alzheimer's disease. Brain. 2015 Dec;138(Pt 12):3673-84. Epub 2015 Oct 21 PubMed.

    . Polygenic risk score analysis of pathologically confirmed Alzheimer disease. Ann Neurol. 2017 Aug;82(2):311-314. Epub 2017 Aug 9 PubMed.

    . Polygenic score prediction captures nearly all common genetic risk for Alzheimer's disease. Neurobiol Aging. 2017 Jan;49:214.e7-214.e11. Epub 2016 Aug 5 PubMed.

    . Assessment of the genetic variance of late-onset Alzheimer's disease. Neurobiol Aging. 2016 May;41:200.e13-20. Epub 2016 Mar 3 PubMed.

    . Alzheimer's disease: analyzing the missing heritability. PLoS One. 2013;8(11):e79771. Epub 2013 Nov 7 PubMed.

    . Estimation and partitioning of polygenic variation captured by common SNPs for Alzheimer's disease, multiple sclerosis and endometriosis. Hum Mol Genet. 2013 Feb 15;22(4):832-41. PubMed.

    . Prospects for using risk scores in polygenic medicine. Genome Med. 2017 Nov 13;9(1):96. PubMed.

Make a Comment

To make a comment you must login or register.


News Citations

  1. Genetic Risk Score Combines AD GWAS Hits, Predicts Onset
  2. First Genome-Wide Association Study of Dementia with Lewy Bodies
  3. Are Early Harbingers of Alzheimer’s Scattered Across the Genome?
  4. Genetic Wild West: 23andMe Raw Data Contains 75 Alzheimer’s Mutations

Paper Citations

  1. . Dissecting the genetic relationship between cardiovascular risk factors and Alzheimer's disease. Acta Neuropathol. 2019 Feb;137(2):209-226. Epub 2018 Nov 9 PubMed.
  2. . Religious Orders Study and Rush Memory and Aging Project. J Alzheimers Dis. 2018;64(s1):S161-S189. PubMed.
  3. . Polygenic hazard score: an enrichment marker for Alzheimer's associated amyloid and tau deposition. Acta Neuropathol. 2018 Jan;135(1):85-93. Epub 2017 Nov 24 PubMed.
  4. . Common polygenic variation enhances risk prediction for Alzheimer's disease. Brain. 2015 Dec;138(Pt 12):3673-84. Epub 2015 Oct 21 PubMed.
  5. . Polygenic risk score analysis of pathologically confirmed Alzheimer disease. Ann Neurol. 2017 Aug;82(2):311-314. Epub 2017 Aug 9 PubMed.
  6. . Interpreting Alzheimer disease polygenic scores. Ann Neurol. 2018 Mar;83(3):443-445. Epub 2018 Mar 13 PubMed.
  7. . Polygenic risk and hazard scores for Alzheimer's disease prediction. Ann Clin Transl Neurol. 2019 Mar;6(3):456-465. Epub 2019 Feb 18 PubMed.

Other Citations

  1. Virtual Exhibit Hall

Further Reading


  1. . Common polygenic variation enhances risk prediction for Alzheimer's disease. Brain. 2015 Dec;138(Pt 12):3673-84. Epub 2015 Oct 21 PubMed.

Primary Papers

  1. . Polygenic hazard score, amyloid deposition and Alzheimer's neurodegeneration. Brain. 2019 Feb 1;142(2):460-470. PubMed.