Drug discovery, the process by which new medicines are weaved from various different chemical compounds, is a slow and expensive process: swallowing an average of 10 years and $2.6 billion to develop. But by marrying the fields of quantum physics and machine learning, startup GTN believes they are moving towards a more elegant solution. Noor Shaker,
Drug discovery, the process by which new medicines are weaved from various different chemical compounds, is a slow and expensive process: swallowing an average of 10 years and $2.6 billion to develop.
But by marrying the fields of quantum physics and machine learning, startup GTN believes they are moving towards a more elegant solution.
Noor Shaker, co-founder of GTN, trained as a computer scientist before leaving her native Syria to hold a series of academic positions in Belgium and Denmark.
She embraced the exploration of artificial intelligence. “I decided to pursue a career in a field that I believed would change the world,” she says. “I fell in love with systems that can be adapted.”
However, a pivotal moment in her life spun her trajectory off course. “I still remember the day when I got a call from my sister, that my mother was suffering from lung cancer,” she says.
Because of her mother’s advanced age, her treatment options were limited to chemotherapy. Shaker says this experience made her reevaluate her options and value time differently. At that point she’d accumulated 12 years experience in academia and machine learning.
“But I knew next to nothing about how to use my experience or knowledge and apply it in a way that could be impactful,” she says.
She soon moved from her position in Copenhagen to London, where she met quantum physicist Vid Stojevic, who would go on to be her business partner. “Together, we started looking at the overlap between AI and quantum physics.”
Of quantum physicists, Shaker says, “They’ve been working on our understanding of the smallest particles of atoms to the largest scale of the universe,” while simultaneously, machine learning scientists were trying to develop models to understand the world.
Together, Shaker and Stojevic developed potential ways of marrying up these two fields. However, they were uncertain as to what the best application would be. Around this time, they had a chance encounter with someone who spoke about applying machine learning to drug discovery and the penny dropped. “We immediately realised that the technology we had in our hands could provide a paradigm shift in the traditional ways of doing drug discovery,” says Shaker.
“For me, this realisation that I could apply my knowledge and expertise that I’d gathered through the years to something as impactful as saving lives was kind of a dream come true.”
Traditional drug discovery has involved searching through millions of different chemicals in the hopes of fusing just the right elements. Over 100 million have been registered today, but Shaker says that there are far more chemicals out there. However, they don’t have a means of accessing these chemicals and finding the ones that could cure diseases.
“At the moment, whenever you start a drug discovery programme, you start by screening chemicals in existing chemical libraries,” says Shaker. “But because those have been mined and exploited over many years it’s becoming harder to find something in those libraries. The challenge is becoming greater and greater.”
She says mapping this unknown expanse can be achieved with new machine learning techniques. In the field of image processing, a class of machine learning called deep generative and adversarial networks has shown promising results.
She describes the technology by comparing it to AI software that can create realistic but fake faces simply from the input of other images. She says in the same way, machine learning could be used to come up with new drug structures. “That’s actually true,” she says. “But, there is a catch.” She explains that this class of methods only works when fed with as much information as possible. In the case of the picture algorithm, it was fed millions of pure pixels of images.
Therefore, this schema cannot be neatly mirrored onto the world of drug discovery, mainly due to the complexity of the structure of drugs, which incorporates quantum physics, entanglement properties and tonic orbitals among other elements.
“We believe that current representation of chemicals are insufficient to train our machine learning models to go into uncharted areas in the chemical space,” says Shaker. For those working in biotech using machine learning for drug discovery they must simplify their representation of chemicals into one dimensional or two dimensional representations to feed into AI models.
There are lots of drug-like compounds in the uncharted chemical space. Yet “because we humans are limited to what we’ve seen, it’s very hard for us to imagine how those molecules look like,” explains Shaker.
Therefore GTN’s technology hinges on the ability to capture the quantum physical properties of chemicals and then create machine learning models that are compatible with this.
“We’ve demonstrated that our models can come up with some commercially viable chemicals for that project in less than a week,” says Shaker. “And we have already achieved up to 30 percent increase in accuracy in investigating some of the technical properties.
“We have lots of projects on our team – they’ve been trained on specific class of quantum physics simulation problems.”
However, this doesn’t mean the use of quantum computing. “We don’t really need quantum computers, we scale all the calculations on planet GPU.” At the moment they’re looking into oncology and neurodengeneration as the first classes of ailments to discover new drug treatments. In May 2018, the startup announced investment of £2.1 million from venture capitalists which it is channelling into the quantum future of drug discovery.