AGENDA

REWS Diversity in Datasets Meeting 7
13th December 2022 @ 9pm UTC

Zoom Recording [^*Y&dY9P]

Add Slides

In preparation for the meeting please read the following:

• Variants Informing Rare Disease Collation from our last meeting
• Polygenic Scores Explainer
• Alicia Martin et al, ‘Clinical Use of Current Polygenic Risk Scores May Exacerbate Health Disparities’ (2019) Nature Genetics 51
• Anna Lewis and Robert Green, ‘Polygenic Risk Scores in the Clinic: New Perspectives needed on Familiar Ethical Issues’ (2021) Genome Medicine 13

Zoom Details:

https://us02web.zoom.us/j/85858589470?pwd=dlRLdE5tejhOK2pBazA3UHhXa2VLUT09

Meeting ID: 858 5858 9470

Passcode: 012179

Attendees:

Anna Lewis, Donrich Thaldar, Alessia Costa, Hanshi Liu, Alyna Khan, Bimal Chaudhari, Aimé Lumaka, Gabriela Chavarria, Sarah Nelson

Minutes:

What is needed to support achieving calls for greater diversity in genomic datasets

So far:

• Background document summarising important elements
• Building on that to develop policy deliverable

• ID what Q’s of diversity and at what points in data life cycle, 2-3 use cases

Use Case 1: interpretation of variants associated with RD (document)

Today - discuss Use Case 2: Clinical use of polygenic risk scores

• What is the purpose of the actual use case (why genetic data is used)?
• Why is diversity important for this particular use case?

Next - other elements of the framework

Use Case 1

Diversity to what end: progress toward equity in receiving an accurate genetic diagnosis

• New research studies currently getting off the ground, probably won’t be able to share all their data, to achieve equity - what data do we need access to? Given that, can we say practical things about what to ask during consent?
• Key role of filtering frequency - need suitable groups that those frequencies are defined as part of, can’t be defined just by population labels that people self-identify as, need genetically inferred groups - pushes need for data sharing beyond just aggregate frequency data, what are these groups, how do they need to be defined to be used for this function?

Thoughts / comments on the doc :

• Donrich Thaldar (DT):

• very often esp in global ethics circles, accepted that some kind of benefit sharing needed, concern (esp legal),
• debate hasn’t been settled from normative perspective, esp not the kinds of benefit sharing acceptable / undue enticement, and how it should be met; problematic to negotiate certain benefits with community leaders, pressure on members of the community to participate, bc if don’t can be seen as reason why benefit doesn’t materialise. Be sensitive to the nuances of benefit sharing, not a blanket good thing.
• Certain types of benefit sharing a simply unlawful, in S. Africa, if a person donates blood or tissue is given a financial reward is unlawful and clinical, except for costs incurred; positivist perspective, red flag for researchers worldwide, look at more nuance

• Gabriela Chavarría (GC): same is true in Costa Rica
• DT: communicating with a community rather than with individuals can be quite tricky, suggestion to have more discussion about this, perhaps not today
• Anna Lewis (AL): Once we’ve had a stab at 2-3 use cases, then we can extract what is common across all of the use cases, then have a session across all of those, think benefit sharing falls into that category

Use Case 2: Clinical Use of PRS (until 27.5 mins before end of hour)

PRS calculated on basis of GWAS, our use case is more specific than that

PRS aggregate effects of many genetic variants linked to a condition, link all variants asstd to the condition of interest, if you aggregate them, the distribution of number looks different between controls and cases; meaningless number until contextualized, for clinical relevance.

• Can report patient’s risk within a percentile within a population/group of interest,
• some measure of relative effect “based on PRS, this patient has an odds ratio of 1.7 / relative risk,”
• or try and integrate PRS into some measure of absolute risk, lifetime/five year risk of developing a condition, “for this patient, 32% risk of developing the disease over rest of lifetime”

Examples:

• Tyrer-Cuzick Risk Assessment, 10 year risk for breast cancer, many different factors considered, your number may dictate the type of screening recommended to the patient
• Coronary artery disease, commonly used risk model put in specific pieces of info and spits out 10 yr risk of ascvd events, determines cost effective treatment approach going forward

How can PRS be used in clinic?

• Primary prevention, lifestyle change
• Secondary prevention, prevent existing condition getting wors
• Tertiary, improve quality of life / reducing symptoms for someone with disease
• Low PRS populations - other applications

Two problems with PRS

• Portability Problem

• Distribution of aggregated frequencies look wildly different depending on the group in which you’re looking at them, which don’t seem to be biologically meaningful
• Perform better and are more predictive in those populations that are more similar to the training data that were used for constructing the risk scores in the first place.
• Can just look at genetic distance from the training population, performance drops off in continuous fashion (see preprint)
• Predictive performance can vary across sex, age, and socio-economic status
• What causes these PRS differences?

• Data handling differences

• Uncorrected population stratification in discovery data or training samples
• Biases in the methodologies, and the way they aggregate effects
• Random error

• True differences

• Drift, selection, genetic effects due to environmental differences

• Eg., do the training in a specific environment, height in UK, most people getting adequate nutrition, in different environment  you’ll find something completely different

• Ethics point of view, differences seen between population

• Interpretation Problem

• PRS capture direct genetic effects, molecular mechanisms of what’s really going on, such that the underlying genetic variants, impact on proteins, influences phenotype AS WELL AS other signals

• Gene-enviro correlations
• Gene-by enviro interactions

• What is a GWAS aiming to do? Statistical geneticists struggling with this question

• ID loci involved in the trait, so that it can initiate a process of mechanism discovery, lead to drug discovery
• Help unpick multi-factorial contributions to complex traits, what might be confounding the signals observed and how should they be controlled for? >> particularly relevant for our use case

• Effect size matters, not just are they ahead or not? Will aggregate all biases and errors of different types

• Portability problem is the #1 issues focused on in the community, rightly highlighted that with current state of affairs driven by data we have access to, PRS are more predictive in populations of european ancestries, but extra logic to go from that fact that this will contribute to disparities, also need to show that PRS are beneficial, and that the benefits depend on predictive power

• Predictive power can never be guaranteed across groups, and even if predictive power equal, other health disparities are involved, more complicated story

• Data relevant to this question:

• GWAS data
• All data necessary to link PRS to clinical risk
• Evidence whether sharing PRS is actually clinically useful
• Perhaps there are other types of data that the group finds relevant

• Considerations around state of the science that might be useful?

• Aimé Lumaka (AiLu): coming from RD field, has there been an attempt to look for portability in the other direction (eg., starting from african population going to the european)? If we look on people from african background, may have better perspective on variation and impact, portability in other direction may have less issue than when starting from a subset (eg., European descent), testing association in population with much larger diversity.
• AL: sample size matters hugely, also meta-analysis out-performs stratifying, that’s starting to become more obvious. Excellent question, but impression I get is to define methods that use all the available data we have and ensure the extra diversity we’re getting from ind’ls with african ancestry isn’t drowned out
• GC: when talking about complex disorders, where the sample comes from becomes an issue; in latin america very little data on LA samples in general in GWAS and when included, coming from ind’ls of LA ancestry living in USA, in complex disorders-  where we live is very important! Cant just be sample recruitment based on ancestry, important to include diversity of current geography

• Interpretation of results, when you think about these kind of samples, eg., lung cancer, picking up something that has to do with biology and something that has to do with… [had a hard time understanding this point]

• Sarah Nelson (SN): location of recruitment is important, concern that the way we think about portability problem has become “calcified” Martin paper is quintessential evidence, but also way we’re contextual

• Genetic inference based on 1K genomes data
• Visualising transferability issue as a continuum is a good mood, if we’re focused on ancestry as the access of difference overlooking other geographical, social and other differences, important for building more robust scores
• How are we even framing the issue?

• AL: we should call for a consistent focus on multiple issues of difference, not just ancestry and not just continental ancestry categories, reason why need to talk about effect sizes and others as we think about input data, location is a proxy for other things that could be causal, eg., policies in place

• Not just the GWAS data and performance of PRS themselves that matters when thinkinga bout clinical use of PRS, other pieces that have to be in place

• SN via chat: I don’t know that this directly addresses Aime’s question, but it might be worth looking into publications from Segun Fatumo - here is an example publication - https://www.nature.com/articles/s41591-022-01835-x 
• Alyna Khan (AK): There is also this paper that might touch upon Aime’s question, as well: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7869832/pdf/main.pdf

Discussion

• AL: concerned that what we were just saying is we need all the data along all of these different dimensions, but very aware of a lot of projects that don’t want to share all data they have, willing to share some data, consent participants to share that data,
• GC: if we don’t want these diversity categories, continental, then we need whole data but then run into privacy issues
• AL: also straight out legal restrictions on what can cross jurisdictional boundaries,
• GC: [missed this comment] From the practical side, sharing large amounts of data can be difficult
• Alessia Costa (AC): in terms of data sharing, coming from non-scientific background, is problem we need to share more data or that there needs to be smaller scale, or new type of research, in some cases data don’t even exist, may be impossible at national scale, eg., postcodes for proxy for other things, not in itself what is needed, but to include some of the data are missing, possible to develop meaningful research knowledge, even if just at national level
• AiLu: Very interesting papers. Thank you. I’m finding my answers reading the abstracts already.
• AL: we know that sheer sample size counts for so much, but perhaps smaller deeper datasets will make more of a difference?
• Genetics community has known about integrating environmental data for 100+ years, and no progress has been made, but exciting time to start thinking about it

Clinical use of PRS

• Diversity to what end? Progress towards equity in receiving a clinically useful PRS
• Hanshi Liu (HL): access justice issue if you look at it from an ethics perspective

• Something of value, not a lot of incentive, like any state of the art medicine, in this case more than accessing the tech but also the value of the test once you do take it, very stratified

• AL: and along multiple dimensions,
• Equity = specification is needed here
• AiLu: are you talking about sharing PRS or sharing data that will drive the development of the PRS?
• AL: thinking about data that enables the clinical use of PRS, could be just the data for training, or could think more broadly than that.
• GC: Comparably predictive is a good statement that we could aim for; how useful they really are is another issue. But if we’re going to use them they should be comparably predictive.
• AiLu: big difference when we speak about RD like (UC1) and PRS (UC2), equity doesn’t mean the same thing, improve variant specification for RD, equity is immediate for all population groups, but sharing PRS immediate benefit not always evident, pharma - we need to access the data of your population, medication will be developed, what is benefit for population from which the data come, PRS not priority for many african countries right now… once that’s recognize, helps frame the discussion, especially with research participant
• AC: seems like a common issue, conceptual issue about equity, comparable predictiveness, access equity - not the same as reducing health disparities, not sure we always see these distinctions made very clear
• DT: at a global level, might make sense, when any research driven at national level, may think acceptable for researchers to focus on their own populations, so if someone doing research in India, justified in focusing on their own population group, no ethical obligation on them to also consider europeans, africans, etc. practical concern, especially from an african perspective, lots of push to focus on black african pouplations, quite justified, but I think when positioning it at a global level, struggle to see how that boils down, high level principle can be applied at national level
• AiLu: this discussion is same as one world has been having for 27 year, COP27 - same discussion, why should people from Africa consider that the forest be kept to heal the environment when other people are making money out of the population, africans able to develop , considered the lungs of the world to heal the world, equity discussion being held in cop 27, they’ve made lots of progress, if we look at what they’ve come to consensus, we may find ideas to apply to genomics community; eg., see integrating other data into covid discussion.