AGENDA

REWS Diversity in Datasets Meeting 5
Tuesday 11th October 2022 @ 3pm UTC

Zoom recording [password: gg5eYb&.]

Zoom Details:

Join Zoom Meeting

https://us02web.zoom.us/j/86150013124?pwd=UmxuQjdoM2RoM0k5NEhKdGFNbnkrZz09

Meeting ID: 861 5001 3124

Passcode: 131979

Attendees

1. Lindsay Smith
2. Maili Raven-Adams
3. Alessia Costa
4. Tina Hernandez-Boussard
5. Kate Orviss
6. Claude Mwangi
7. Donrich Thaldar
8. Anna Lewis
9. Mutiat Afolabi
10. Marietjie Botes
11. Beatrice Kaiser
12. Alham Saadat
13. Dsrivastava

Minutes

THB: Recap from last meeting and go over some of the key points. We talked about research data lifecycle, components on where bias can enter. Went through and spoke about how we’re going to move froward. Today we will build on this and talk about the lifecycle, there are many lifecycles in the literature. Mapping diversity at each stage.

AL: We aligned on our strategy for moving forward, and now going to see where that strategy takes us.

MRA: In the last 4 meetings we’ve addressed that background context. In the document, outline where we are to date. From this meeting, we want to start addressing the questions that need to be addressed. Started with 3 use cases.

For each of the use cases, identify what diversity means. Doing this for each stage rather than what diverse data means as a larger concept.

AL: Not sure is use case is exactly the right term, different considerations might come up.

MRA: We came across this lifecycle that nicely identifies points in a data lifecycle.

We put together a workflow that includes a bit about the purpose of why someone is doing the research. Last time we talked about bias, so I put a slide together on the types of bias to think about as we work through things.

THB: Many types of bias throughout the entire lifecycle. May have different names in each discipline. Every way we use the data and in the decision we make, there’s opportunity for bias. Need to be aware of it and be transparent about it. Each bias affects the way we can capture, interpret or use the data.

AS: One things that stands out is the purpose bias. Thinking about the initial ideation around what is the scientific question that’s being asked. That purpose piece is incredibly important, but who is defining purpose and how it is being evaluated throughout.

THB: Good example is that we’ve had clinicians come to us to ask about who is missing appointments. Instead, we suggested looking at barriers to making appointments.

AL: Rather than attempting to define what diversity in datasets is, we wanted to break it down and think about what you’re trying to do and what considerations might be in that case. What are the implications. These use cases capture a lot of what goes in genomics. If we can show what this type of thinking looks like, hopefully arms people to think through their own use case.

Developing an understanding of complex traits, typically GWAS.

Why we might care about diversity when we think about interpretation of rare variants. When you’re trying to diagnose someone with a suspected condition, the task of genetics is to locate which variant(s) might be causing that condition. Need sequence data and have to work out which variants might be causative. Here diversity matters because the frequency of that variant is the single most important piece of information for causality. But they are not uniformly distributed across our species. Some cases where some people who have African ancestries have a variant interpreted as pathogenic, but turns out it only looked pathogenic because there wasn’t enough people with african ancestry in the databases.

Work out who across this group thinks about this, make a subgroup to work on this.

When we think of genetics and health, there are the rare variants, and those tend to have big effects. Most conditions also have a genetic component, and we discover that via different analyses. Find there are hundred of variants that each have a tiny effect, but if we aggregate those variants, we get a score that is somewhat predictive of the trait in question. The proposal is that these are getting ready for prime time in the clinic. The issue is that if you divide humans up into continental categories, it’s much more predictive for European populations. This may have closer analogies to non-genomics areas. The training data impacts how your models perform.

There’s this notion of ancestry, which is a dominant way people speak about genomics, but it can also be a dangerous way to talk. Mostly because the continental categories are an oversimplification. We don’t just want to say “oh we just need more data from non-European ancestries”, that’s not all of it. We haven’t actually discussed this clinical PGS score as group, any overall thoughts about this area?

AC: When we talk about PGS or complex traits, there is a non-genetic component. How do you come to understand the risk and how you assign responsibility to the genetics vs lifestyle for example.

AL: Gets to the heart of some of the things we’re going to want to say. The short answer is that GWAS just looks for correlations in the data they happen to have. Big question on how this data is generalized to those who are not genetically represented in the dataset. Various different camps that think different things about these topics. Additive genetic effects that are largely independent from the environment.

DT: What you’ve mentioned about data being reflective of Europe, that I think is a fascinating research question. It can be approached from different angles, from an innovation or economic perspective. It might be because of competition with public databases. The question is can private enterprise provide the solution from an African perspective or not. Can approach from a global public policy perspective. I suggest the solution should come from Africa.

AL: Do you have a sense of what the important pieces of data protection legislation that are needed?

DT: The leading nations in Africa have data protection legislation, but we are generally very critical. Drafted on par with what’s in Europe and the GDPR. Paradigm shift of working in compliance with this legislation. TO my knowledge, no country in Africa that has this legislation has taken the step for standard contractual clauses. Dealing with data protection legislation in Africa is still very new. These are preconditions to fall in place before focus can shift to researchers, funders, etc. to create a big African database.

AL: This group could highlight what aspects of the data are important from a sharing point of view. It’s not necessarily that everyone needs access to everything. I think we can spell these things out. In the data created box, the need for summary of data protection requirements.

DT: If you can specify exactly what it is that is most needed, it’s an invitation for scientific entrepreneurs in Africa to supply that need.

KO: This whole issue around GDPR, privacy, the interface between how data is generated and accessed. This point about the interface between genomics and the GDPR has not been solved 100%. If we stop sharing and grant access, that could have a beneficial impact. I find it bewildering that every jurisdiction is trying to answer this. Need to try and help in terms of accessing the thinking in other places.

AS: Wanted to add that being cognizant that Africa is a large continent made of up diverse countries, it’s really important not to oversimplify that there’s an easy way to solve these problems. What are the foundational pieces we should be thinking about that we could generalize across diverse global populations. Starting from a place of understanding that complexity, all of this plays into how we think about ethical practices and policy.

AL: We spoken a lot about what it takes to share data. Should think about integrating data. The stuff around what are the relevant legislative frameworks around where data is being creating and what impacts how it is shared.

MRA: We often think about this ancestry question, but maybe need to think broader about gender categories, for example.

AL: If it comes down is that we need aggregate frequencies should be labeled in such and such a way, that had implications for data sharing. Whereas, if we’re talking about clinical PGS, these dimensions matter much more and may have different considerations.

AC: Possible to share papers in advance about these topics? This would be helpful to think about this in granular and practical terms.

AL: Great idea. Let’s plan on doing that.