Diversity in Datasets: Ensuring Applicable and Equitable Research Outcomes Based on Usability, Operationalization and Transparency of Datasets / Ensuring the Push for Diversity in Datasets Enables Greater Benefits from Genomic Research and More Equity in the Field

Ensuring diversity in datasets used in research, and consequently in research outcomes, remains a topic of great importance, including  genomic research, due to the over-representation of individuals recruited in Europe and the US. Although being tackled from many different angles on a national and international scale, the majority of research and progress in this area has focussed on community engagement to gather datasets from populations who may otherwise be excluded. In tandem, there has recently been an increased focus on the ‘categorisation’ of participants and their genomic samples to try and ensure ‘diverse’ and ‘representative’ samples are being curated and utilised.

However, there are some fundamental unanswered questions underlying this appeal for more diversity in genomic datasets. These include, but are not limited to: (i) What is diversity? (ii) What can be considered ‘representative?’ (iii) How can we ensure/measure that datasets and research outcomes are ‘diverse’? (iv) Whose responsibility is this? (v) Does our understanding change depending on context? (i.e. country to country and research vs clinical settings) (vi) How do we ensure research ‘diversity’ remains transparent to onlookers? (i.e. publishing of results) (vii) Is there a role for standards and regulatory oversight? (If so, what does this look like)? (viii) When are there good reasons for unrepresentative data?

Additionally, ensuring the full international scope is considered is imperative to ensure diversity efforts go beyond the borders of one country. For example, there are concerns that lots of work in this area is being done in the US, meaning that there is a risk the rest of the world will be excluded from these discussions and influenced by inapplicable US-centric norms. To avoid This, different countries’ practices and ethical considerations need to be understood and compared to gauge how best to foster international harmonisation in this area, while also recognising the reasons behind differences in global perspectives.

This is not only important for well-established genomic research practice, but it is also integral to continuously emerging technologies being used in genomic research, such as artificial intelligence (AI) and machine learning (ML). Utilising these new technologies may have large benefits to genomic research as it may enable the analysis of genomes and patterns with more efficiency, speed and accuracy than previously possible. However, without ensuring diversity in the data underlying use of the AI/ML, and transparency in the processes applied to this data, biases are only likely to be further frustrated by this technology and research results may be skewed and disproportionate as a result, benefiting specific populations over others. Therefore, this project can also be considered to constitute the first step towards a possible AI/ML and bias project in a future Regulatory and Ethics Work Stream (REWS) Roadmap.

REWS is well placed to tackle this issue due to its international scope. In fact, this work can be considered to further GA4GH’s mission, to activate the human right for all individuals to benefit from science equally [1].

Policy Outputs:

1. Defining the concepts that underpin this topic (e.g. ‘diversity’, ‘representation’ etc)

• What does this mean in practice, and how does this apply internationally?

2. Actionable recommendations for researchers to adopt that can uphold ‘diversity’ in their research and finding outcomes, touching upon:

• Data Acquisition (e.g. a representative capture of data across populations)
• Usability of the dataset (e.g. what to consider when selecting a data sample)
• Operationalization/Measurability of ‘diversity’ in research (also considering when there may be (good) reasons for datasets to be undiverse)
• Transparency in use, research and publication of research (e.g. publication of the dataset and ‘diversity’ metrics)