AGENDA

REWS Diversity in Datasets Meeting 6
2nd November 2022 @ 12pm UTC

Zoom Recording [password: j5&&v0+f]

Slides

In preparation for the meeting please read the following:

• Arjun Manri et al, ‘Genetic Misdiagnoses and the Potential for Health Disparities’ (2016) The New England Journal of Medicine 375, 655-665.
• Aime Lumaka et al, ‘Increasing African genomic data generation and sharing to resolve rare and undiagnosed diseases in Africa: a call-to-action by the H3Africa rare diseases working group,’ (2022) Orphanet Journal of Rare Diseases 17
• Matchmaker exchange
• Ambroise Wonkam, ‘Sequence three million genomes across Africa’ (2021) Nature

Zoom Details:

Join Zoom Meeting

https://us02web.zoom.us/j/88982232645?pwd=eDVRUmdKdU9UaGpBeHpBRWRQT1BZUT09

Meeting ID: 889 8223 2645

Passcode: 319500

Attendees:

1. Maili Raven-Adams
2. Lindsay Smith
3. Anna Lewis
4. Beatrice Kaiser
5. Bimal Chaudhuri
6. Amie Lumaka
7. Alessia Costa
8. Ariana Gonzalez
9. Claude Kamau
10. Hanshi Liu
11. Yvonne Bombard
12. Mutiat Afolabi
13. Elena Ghanaim

Minutes:

AL:

Recap

Background doc, want to start developing policy deliverable, want to focus in on some use contexts and ID what diversity should mean in those contexts, will talk about 3 - Interpretation of Mendelian variants and other two are up for discussion, will be revisited

want  to identify what diverse data means and what should it mean, to what end

Want to avoid focusing on a group because they are biologically unique

Data lifecycle - can be adapted if necessary

This week: for each use cases be asking what diversity in datasets should mean and think about implications for that throughout the lifecycle.

Interpretation of Mendelian variants

In interpretation what we need to do in the context of returning genetic test results to anyone with suspected genetic condition, need to be able to take a variant ID’d through sequencing and see if identified with phenotype, important is frequency in databases, i.e. NOMAD

Important to know if variant has high freq, indicates not involved in diseases, if rare might be compatible with being involved in disease.

Databases carve up humanity into diff groups

For rare variants have examples of variants reported as pathogenic later found not to be - all looked very rare, in Black Americans, reported as pathogenic, but had it been there were lots of Black Americans in the database they wouldn't have been reported as pathogenic because frequency was high - shows why need diversity in datasets

In order to dismiss all variants not causative of disease has a “filtering frequency” that exclude bottlenecked populations. Can filter on the highest of those frequencies.

How do they define a bottleneck population?

Subject to founder effects, a point at which there weren’t that many of them, there was interbreeding, to the extent that they may harbor variants that haven’t had a chance to be selected out.

In the design of gnomad, they have tried to enrich by any means a variant that could be of interest. They did not include young children. Initially, categories were not included, but the more you can include about populations, it changes the way you evaluate the frequency of that variant.

It’s powerful to be able to assume these people are healthy individuals.

BC: Part of the logic of excluding the bottleneck populations isnt because you shouldn’t use them, it’s that they pre-computed that value - it’s based on size of population, certain distributions. Many people are using gnomod to asses someone not from those populations

ALe: In the original guidelines they suggested using race. By my understanding that the field has moved toward the max frequency for any population.

BC: 99% true, but breaks down for highly bottlenecked populations.

AiL: +1 Bimal. In practice, looking at population specific still applies. One of the common cause of these classifications have been the novelty. People make assumptions about that novelty, but if you look at the populations level, it actually may be relatively common. Most people see diversity as continental, but in Africa, it should be seen with two different layers. There are natural barriers that have segmented African populations.

ALe: Getting to subpopulations are important for the filtering frequency, but I took this as the frequency in the sub is important to matching to patients

BL: No, many of the reasons you exclude break down if the patient is from that population

AL: Covering ethnolinguistic, regional and other groups to capture the full scope of Africa’s genetic variation. Social and geographical barries. We’re all knitted together in an ancestral graph, but the structure is lumpy. “The high sensitivity of African data is often identified as the main limit - like lack of proper consent”.

Everything we have said so far is about frequencies. But we need to share individual patient level information attached to certain variants in the hope that other have the same features.

BC: We care about allele frequencies for all of the reasons we covered. Lack of diversity also contributes to inequity. Rare variants are more common in continental Africa. Uncommon variants, look rare. But if no one from [x] population is represented in gnomad, it can look rare. However, there actually are more rare variants in African ancestry populations, it’s not just the dataset. My best guess, is that less than 10% of data in the African American population is actually from Africa, it’s admixed.

If the process for making a RD diagnosis is to conduct seqeucning, and then identify the set of rare variants. The last mile is harder if that patient is of African ancestry because of the rare and apparently rare variants. Have to sort out what is actually rare, and gnomad is not helpful because it’s not diverse.

Patients, depending on what their history is, they may not know that information to a certain extent. Happens most often in families that have somewhat isolated populations.

The consequence turns out to be that misclassification. Burden fell on the African patients. The other way that this manifests is that you can’t make a diagnosis. The lab just reports a lot of rare variants.

Without good allel freq estimates, rare disease genomics in Africa will be a huge challenge. Increasing the diversity of alleles represented in databases will require African genomes. Increasing African genome representation more likely to help African Americans than more African American datasets.

ALe: I think our main challenge is to call for more data, and take the global picture and think about the RE pieces. One type of thing that can be helpful is thinking about what those populations actually are.

Breakout Groups

Next Steps: