Priority policy actions for AI implementation in genomic medicine

Advances in artificial intelligence, particularly machine learning and deep learning, are offering ever more powerful approaches to genomic analysis that many hope will drive forward the next era of personalised medicine. However, like in other fields, AI for genomic research has been susceptible to unrealistic expectations and a great deal of hype.

A recently published PHG Foundation report examines what factors are limiting progress in genomic medicine, and what resources and collective commitment will be required to realise the full potential of AI in the clinic. The authors outline their practical recommendations to healthcare policymakers for accelerating clinical application and minimising risk to patients.

What has driven the recent rise of AI?

In recent years the convergence of increasingly powerful high-performance computing and the expansion of national programmes to collect population datasets have driven the use of these technologies in genomic research. Improvements in software and hardware, including deep learning algorithms and graphics processing units (GPUs), have made possible the rapid advances in genomic medicine in a matter of years as the technologies are our most powerful tool to process large sequenced datasets. For instance, in clinical diagnostics, AI-driven image analysis has been shown to be more accurate than human physicians.

What are the leading applications of AI in genomic medicine?

Methods have been developed across the genomic data pipeline to aid genomic analysis by machine application. These approaches can be broadly split into two groups: those that directly facilitate clinical genome analysis, and those that seek to improve our understanding of genetic variation underlying human health and disease.

Algorithms are frequently being developed by data scientists to gain a clearer understanding of genetic variants, such as somatic and copy-number variants, which are traditionally difficult to detect accurately. Similarly, developing tools for predicting the effect of such variants, and how they may impact downstream biochemical systems and molecular processes, is a key focus of scientists. Other tools such as deep learning based image recognition systems can be used to mine phenotypic data from electronic health records, to inform clinical diagnosis.

And while we have ever more powerful methods of genomic analysis, we’re still far from having a complete understanding of many diseases and their pathways. To better understand the relationships between genetic variation and disease, and thus inform the most appropriate clinical interventions, AI is being used to study molecular changes in tumour development to improve the efficiency and accuracy of CRISPR, a technique used to investigate the role of genes in disease; and integrate genomic data with other sources of biodata.

What is limiting the clinical uptake of AI?

AI, frankly, has yet to demonstrate significant improvements in genomic medicine, largely due to a number of interconnected factors:

• Data quality and accessibility: Think rubbish in, rubbish out. If the data used to train machine learning algorithms isn’t complete, accurate, or up to date, the performance and reliability of the tool will be substantially impacted.
• Bias: Bias in data often reflects the bias we inherently have in our societies; with minority and ethnic populations often under-represented in genomic studies and national datasets. This and algorithmic bias, as a result of data processing and integration, has the potential to exacerbate existing health disparities.
• Expectations: Machine learning, in particular, has often been referred to as a black-box tool, where data is imputed and an analysis reading is produced, but the researchers themselves don’t understand the process behind the algorithms. This inevitably causes uncertainty and can make replicating methods and results difficult.
• Skills and infrastructure: Genomics itself, such a broad discipline, is only multiplied in its breadth by the addition of data science and AI, meaning that most researchers don’t have all the skills, expertise, and resources, to deliver all the potential benefits of AI in genomic medicine in its entirety. Instead, a focus should be taken across multiple sectors.
• Privacy and security: Concerns relating to the secure, confidential and ethical use of patient or participant data, can impede the use of these technologies.
• Regulation and clinical governance: Whether or not the AI algorithm developed qualifies as a medical device can determine it’s regulatory status. Adaptive algorithms often fall into a grey regulatory area, where they need to be certified and if something were to go wrong who would be liable.
• Uncertainty: Again this relates to the problems arising from black-box algorithms, where without transparency it’s unclear if these tools meet EU General Data Protection Regulation for clinical decision making.

Many of these challenges are already being tackled by researchers and policymakers alike. However, the PHG Foundation, a not-for-profit think tank, sees addressing these above priorities as vital to accelerating the adoption of clinical AI tools, and therefore have addressed policymakers with their top seven recommendations:

The full AI for Genomic Medicine report (63 pages) can be found here.