How is the power of artificial intelligence helping to enhance diagnosis and discovery?
Medical advances, better drugs, healthier lifestyles and safer workplaces have resulted in us living longer lives. However, late diagnosis of otherwise treatable illnesses is now one of the biggest causes of avoidable deaths. In addition, longer lives and larger elderly populations are inevitably associated with complex and chronic health conditions with limited treatment options.
Artificial intelligence (AI) algorithms were first explored in the 1960s, however they have only entered widespread use over the last decade. This is due to the increasing volume and availability of data and the exponential advances in computing power.
Today, AI now has the power to not only improve early diagnosis but also to support the discovery of vital new medicines.
Early diagnosis of disease is vital to increase both the length and quality of patients’ lives. Although significant progress has been made in diagnosing many forms of cancer, little progress has been made where symptoms do not present until a late stage, for example in pancreatic cancer. The application of AI to early diagnosis presents significant opportunities for both earlier intervention and improved efficiency.
Established AI image processing models are being adapted to interpret medical scans with comparable performance to expert clinicians. There has also been an increase in app-based symptom checkers, which use AI to interpret symptoms from users to give a suggested diagnosis. The increasing uptake of wearables and home monitoring devices will also generate additional data to further improve diagnosis and management of chronic conditions.
However, AI systems are not infallible and will inevitably make errors, raising issues around accountability and ethics. As adoption of patient-facing AI increases, there will be a key role for clinical professional bodies, healthcare regulators, for example the National Institute for Health and Care Excellence (NICE), and the Medicines and Healthcare products Regulatory Agency (MHRA) in ensuring public safety.
Machine learning is well established in the drug discovery industry and is used from the very early stages in a project to select molecules for progression, through to application in later preclinical stages to minimise the risk of toxic or metabolically unstable molecules reaching clinical trials.
Until recently, these learning approaches have largely been restricted to working with well-defined problems with data that can easily be transformed, or ‘featurised’, into a structured format that the learning algorithm can then process. They have also been used in a relatively passive mode – filtering or scoring ideas proposed by humans but unable to generate novelty themselves.
Recent developments in AI, however, are extending the power of machine learning beyond these boundaries and enabling more complex, fuzzy and less-defined questions to be addressed effectively. A key advance in the recent ‘deep learning’ approaches over previous learning methods is that they themselves can extract the relevant features from the data without prior reduction to simple features, extending the data sources that can be used, reducing bias and ultimately enabling the data to ‘speak for itself’.
AI is also able to generate novelty. Using nothing more than a database of existing molecules, AI algorithms can learn some of the rules of chemistry and create entirely new virtual molecules with the ability to steer generation towards those with desired properties. We are also seeing methods to plan chemical syntheses; a difficult multistep problem where AI must decide which of a multitude of branches to pursue.
AI is already having an impact on the drug discovery industry, but short-term hype about the power of AI needs to be kept in perspective. The very high level of inventive abstraction that humans are capable of without prior experience, to bring distant concepts together remains problematic for a machine (a human can easily imagine a dog in a lounge suit dreaming about a white Christmas, while a computer will struggle).
We should also be wary of unrealistic criteria for success. Where AI has been used in image recognition or driverless cars, performance near that of a human is a major success. Suggesting that AI will be successful only when the gene name of a therapeutic target is input, and a drug emerges from the other end, is to set an extraordinarily high bar. This would require spectacular outperformance over our existing discovery system, where most projects end in failure.
Well before that distant ‘Drug Machine’ vision, AI will impact where we work with it as a tool. We can anticipate increasingly sophisticated text-mining algorithms aiding biologists to keep track of the >800 thousand biomedical articles published annually and identify linkages between diseases and underlying biological processes. Unearthing new datasets with these methods will also help feed the data-hungry machine learning algorithms. Increasingly accurate predictions of a potential drug activity towards biological targets that may induce side-effects, will reduce the number of compounds that need to be synthesised and tested, enabling go/no-go project decisions to be made more quickly and with more confidence. Engaging human creativity and imagination to develop learning approaches that increase the effectiveness of AI is critical; for example, methods to capture increasingly sophisticated and accurate descriptions of protein-drug interaction
are being developed.
The UK is well-placed to take advantage of these developments, via a growing sector of small and medium sized enterprises (see below for some examples). Coupling these developments with the UK’s long-standing position as a leading centre for the pharmaceutical industry with the intricate knowledge of how to develop a drug in practice holds the potential for a high growth industry.
The field is competitive with major investments taking place in China and the US, but the UK ecosystem of entrepreneurial technology companies and practical drug discovery expertise presents a major opportunity.
Andrew Pannifer is Lead Cheminformatics Data Scientist at Medicines Discovery Catapult.
Matt Hodgskiss is Lead Data Scientist at Medicines Discovery Catapult.
Go to www.md.catapult.org.uk/
UK Companies working in AI
Babylon Health | www.babylonhealth.com
Ask Babylon – an AI symptom checker
Skin Analytics | www.skin-analytics.com
Novel AI models to identify skin cancer
GTN | www.gtn.ai
Pushing the boundaries of describing how drugs react with their targets
ExScientia | www.exscientia.co.uk
At the forefront of objective decision-making and compound design
NextMove | www.nextmovesoftware.com
Cutting-edge software to automatically identify and extract key data in scientific documents
Intelligens | www.intelligens.ai
Methods to enable machine learning algorithms to process ‘sparse’ data
Optibrium | www.optibrium.com
Working to incorporate sparse data algorithms into software applications
for drug discovery