Using AI to develop medicines with a higher probability of success
Drug development pipelines are slow with low success rates. New technologies, robotics and AI can all make a difference.
Article information and share options
Individuals across the world are suffering from diseases that could be cured. Discovering cures historically has been a long process. The discovery phase for drug development can involve up to 10,000 potential compound combinations. By the time of pre-clinical trials, up to 100 potential drugs could be in the mix, resulting in 10 candidate cures moving to phase 1 clinical tests. By the time phase 3 trials have been undertaken, and one candidate is finally approved, the whole process can take 10 years.
The COVID-19 pandemic has forever changed the way we regard medical development pipelines. Faced with the emergency of an infectious condition, aided by ample development funds and a supportive regulatory regime, vaccine candidates were being rolled out within 10 months of the pandemic being declared.
Eroom's law suggests an inverse relationship between R&D funding and medical success. A new molecular entity had become the product of millions of dollars of funding; and the chances of success of a new candidate from inception to becoming a final product being only 5%.
In order to increase the effectiveness and efficiency of the medical research pipeline, smarter use of technology is being made.
Human genetics are a rich tool for medicine development. Target developments with high quality genetic evidence are around three times more likely to succeed in the clinic. Biobanks and the use of high-powered analytical tools can help us create maps of cellular causal mechanisms. Gene editing technology, such as CRISPR, allows scientists to apply alterations to compare to untreated controls. Creating maps of cellular causal mechanisms could see a three-fold increase in the chances of a candidate treatment coming to market.
Robotics are increasingly used in labs. Robots can run thousands of petri dish experiments on a daily basis, generating evidence at scale. That evidence can then be analysed, again at scale, through the use of artificial intelligence. Deep intelligence can help isolate perturbations that move the molecular phenotype towards a healthy state. Machine learning can further help streamline other data sources, with the use of: natural language processing to examine scientific literature; genome wide association studies; and single cell RNA sequencing.
Finally, the amount of computation available to scientists allows for the construction of large scale experimental feedback loops. Machine learning models can use experimental data to create causal effect maps; these can be used to generate new experimental designs in turn. Every iteration of the cycle provides new insights and learnings of the underlying disease biology. Machine learning achieves results at scale.
There are currently around 8,000 potential treatments in preclinical development. Around 5% will yield new medicines. If that number could be pushed up just one more percentage point – to 6% - there would be 80 more cures that could be potentially life changing for diseased individuals.
Summary based on Swiss Re Institute event: Algorithms for hope.
See also other summaries from this event
- Kate Darling - The new breed: What our history with animals reveals about our future with machines
- David Dao - Using artificial intelligence to help restore the natural world
- Frida Polli - Using artificial intelligence to reduce bias in recruitment
- Cathy O'Neil - Algorithmic accountability will lead to a better world