It has been said that COVID-19 is a generational challenge. In many ways it is. Not just because it is affecting humans and human systems around the globe in unprecedented ways, exacting pressures that many of us have never felt. But because it seems to be one of those watershed moments, a catalyst for collective action, innovation and discovery. Realizing the full potential of artificial intelligence for drug discovery and repurposing may be one of these innovations.
To turn the tide against COVID-19, treatments and vaccines must be developed at breakneck speeds. Drug repurposing, or using an existing drug to treat a new disease, is our best – and fastest – hope. But imagine the expansive number of drug-to-target combinations. Even within the antivirals cohort, a needle in a haystack seems like a generous characterization. And landing on the drug-target hypothesis is only half the battle. That hypothesis must be tested through clinical trials and ‘wet lab’ environments - a process that is frustratingly slow at best, and excruciatingly long at worst. Biochemical researchers are constrained by the data in front of them and siloed from others who may be making progress. Discovering and validating the right drug-target pair (i.e. the Achilles’ Heel of a disease) is laborious. It is code breaking at its most frustrating. Its most manual.
Fortunately, in the case of COVID-19, we aren’t starting from scratch. We have the benefit of collective experience and knowledge from other coronaviruses, most recently SARS and MERS. We aren’t against a complete unknown, as we were with HIV. But that collective intelligence is difficult to mobilize.
Here is the perfect problem set for artificial intelligence. The goal – a modern-day Bombe (the machine used by Alan Turing’s codebreakers in World War II) that can digest thousands of combinations of drug-target pairs to decrypt the disease code and point our research, our trials, and ultimately our treatments in the right direction. An ingenious combination of data, knowledge and creativity to solve the problem faster and at the scale and complexity of human health.
This goal is not confined to our current epidemic. For years, researchers and data scientists alike have been working on ways to speed up drug discovery, using rapidly expanding data and technology to find new ways of testing outcomes, understanding efficacy, or identifying risk. Deep learning models have been successful in producing a short list of potential drugs to treat a condition, and virtual screening can test these drug and protein interactions in silico.
But to date, applications have been limited and constrained by complexity. It is no small ask for a biochemical scientist to transfigure into a computational researcher, adept at the language of data science. There are also myriad models out there, using different methods for coding data and making predictions, and different data sets to “train” the algorithms.
IQVIA’s Analytic Center of Excellence joined data scientists, computational researchers, and public health experts from Harvard University, the Georgia Institute of Technology and the University of Illinois, to develop an elegantly simple line of code that accelerates the application of deep learning to drug repurposing and virtual screening. The toolkit – called DeepPurpose – makes deep learning and state-of-the-art data science methodologies readily accessible to the biochemists and industry leaders making the critical decisions within drug discovery.
Imagine a researcher wants to explore which antiviral drugs could be repurposed to potentially treat COVID-19. To start the process, she need only input a single line of code. From there, the toolkit uses five pretrained aggregation models and automatically outputs top-ranked drug candidates from the antiviral drug dataset. The simplicity of this compared to legacy approaches cannot be understated. And DeepPurpose was already put to work in the search for effective drug candidates for COVID-19. Taking a library of 81 antiviral drugs, DeepPurpose identified 13 with a high binding affinity score, or a high probably of efficacy against the SARS-CoV-2 virus. Among the top candidates are Ritonavir, Darunavir and Lopinavir – all of which are currently in clinical trials.
Alan Turing and his team of codebreakers were pioneers, innovators whose work enabled a better future for everyone. Human Data Science promises a similar kind of hope with the acceleration of discovery and innovation to advance human health – the integration of data science and human science to uncover new patterns, new ideas and new solutions for some of our biggest challenges. DeepPurpose is a Human Data Science achievement that delivers an immediate boost to the fight against COVID-19, and a tantalizing glimpse at what is possible as we address other enigmas of human health.
For more information about DeepPurpose, read the full paper here.