How pioneering research project could one day help influence the future of medicine
Imagine a world in which computer processing power will not only revolutionise the way drugs are made, but will also identify the markers of a disease before it matures in your body. How is that possible?
Before we pick apart how all of this is supposed to work, a quick disclaimer: Insilico Medicine is not a start-up, but it is still on its way to validate itself as a business model. As anyone familiar with the Theranos controversy will understand, it is really difficult to audit the business model and the science behind medical start-ups.
Theranos was a “blood testing” start-up that received US$400 million in funding from serious Silicon Valley investors, and a hefty US$9 billion valuation, before it exited the retail lab business.
So what follows is not an analysis of the Artificial Intelligence (AI) medical sector – it is still too early for that, possibly even in the US market. Instead this is a case study exploring the technology and science behind an AI medical concept that looks to be on its way towards validating and commercialising in the near future. As such, Insilico Medicine and its team are ideally placed to help our readers understand how this technology works, and how this sector might develop here in the UAE in the near future.
With these caveats aside, Mehta provides an introductory crash course in how preventative medicine works, with Zhavoronkov following up with some insights into exactly how AI and biotechnology are merging.
“There are enough studies that have proven, although more research is needed, that simple de-stressing techniques like meditation and yoga, even if done rigorously for twenty one days, showed molecular changes in the body that are established markers for de-stress”, says Mehta. “For example, your cortisol level goes up in the body, if you are under stress. The first step is teaching an artificial intelligence to ‘learn’ how to identify these ‘biomarkers’.”
Why? Because by being able to identify changes taking place in the body earlier than ever before, it might just be possible to reverse those changes and, one day, even cure them.
“If you are stressed, molecular changes are taking place in your body that will eventually lead you to all of the chronic diseases, which include cancer,” continues Mehta, adding: “If you are not having proper food, this will eventually lead you to chronic diseases such as diabetes, and other metabolic diseases.”
These are day-to-day examples all of us are familiar with. What has changed is that advances in genomic science allow doctors to take a “molecular snapshot” or “image” of the tiny changes in the body that can lead to illness. Diabetes or heart disease don’t always manifest right away - but according to Mehta “at the molecular level you can already see the changes.”
So, let’s say we have an “image” of a particular molecular structure or matrix of structures. We already know that this “image” can lead to a particular disease, let’s say cancer. What if we could train a computer to automatically scan for and recognise that image at the molecular level, for every single human being on Earth? See where we’re going with this? As with everything else in the new Millennium achieving this historic goal begins with two things. Lots of data, and cat pictures. Yes, cat pictures.
It turns out that teaching a computer to recognise a snapshot of a molecule, is not that much of a leap from teaching it how to distinguish a cat’s face out of a dataset of two hundred and fifty million other images. Or in other words, teach a computer to recognise a cat, and at some point in the future, it might be able to help you cure cancer.
If that sounds completely bonkers, it is because it is.
This data-magic is known as “deep learning”, and it is a central component upon which all forms of AI - from self-driving cars, to Prisma app and Google Image Search - are built.
Back to cats. “The input could be a cat’s face,” explains Mehta. “Now there are multiple data-points in a cat’s face. The eye of a cat is different from the eye of a human or eye of a dog. The size of a cat is different. So the computer is programmed in such a way using the neural network to recognise the cat and provide the output of: ‘okay, that is the image of a cat.’
“If that sounds easy, it isn’t. That programming is not easy - it’s not linear programming. Way back we had this simple programming where you input this data and that is the outcome… but now computers have to analyse various data-points to recognise that pattern which is of a cat.”
Eventually the process of deep-learning, leads to what you or I would today label AI. “So, when the programming is done and when the system is trained to recognise the same problem - the cat in a video, or in a picture, or in real-life, this is where the computer becomes intelligent enough to recognise the cat anywhere. Even if you just show the eye or marker of a cat, it will immediately recognise it,” concludes Mehta.
The superhuman potential of AI, is something Zhavoronkov also talks about. During a Skype presentation he exhibits a series of slides demonstrating the speed at which this technology is emerging. “The real renaissance in this field started in kind of 2010-ish,” he says. “In 2015, deep learning surpassed human accuracy in image recognition, last year it came close to human accuracy in text recognition, human or almost superhuman accuracy in voice recognition.”
It is compelling stuff - but will it make money? As medical costs rise across the UAE, the USA and the world, can AI really save big pharma? How can relatively small companies, such as Insilico Medicine, hope to compete in an industry where the R&D for a single over-the counter-drug can cost billions of dollars?
There is a case to be made that many of these companies have become victims of their size. “Basically, they send emissaries to the jungle of Amazon and to all kinds of rare islands to find new molecules, and they construct libraries of those molecules consisting of millions of compounds and they blindly test those using huge robots, on human cell lines and on bacteria,” observes Zhavoronkov, referencing an Insilico’s recently-published paper which won the American Chemical Society Award in 2016.
In it, the Insilico team suggest a new drug delivery pipeline powered by AI. By using the above techniques, Zhavoronkov believes they can generate new molecules very quickly and test them in mice, and then in human models. “And you can essentially replace entire pharma companies,” he says. “That’s why you don’t need those huge robotic facilities anymore, you can have a small basement somewhere in the UAE where you have a lab to validate your predictions.”
By contrast, 95 percent of the clinical trials in cancer fail and often take decades,” Zhavoronkov points out. “So, most of the time pharmaceutical companies are used to failure.” This is why, he argues, companies can charge up to US$200,000 for the treatment of one patient, which costs them nothing, “because they want to recover their R&D costs.”
Mehta describes the “drug pipeline” process Insilico Medicine is working towards in three simple steps. “First, we collect actionable data, for example a blood test, a urine sample, or other indicators. Then we train a machine using neural networks to analyse that data over a period of time and based on that data the trained system can come up with a potential drug pipeline for various chronic diseases. The final stage is to validate this new drug in a clinical trial, and then place it in the market.
“The advantage of an AI is what would have taken 20 years to identify the right molecule, pass it through animal modelling to the preclinical stage, now the rate of failure and the duration is reduced.”
In many parts of the world the medical system is broken. What remains to be proven is whether or not AI really will revolutionise the drug pipeline, making faster and cheaper to produce effective drugs. As Zhavoronkov himself admits during our conversation, it is very likely that once Insilico Medicine validates its model, the existing pharma giants will simply swoop in and deliver a buyout offer the team can’t refuse.
This scenario is made more likely by the fact that many of the global pharmaceutical giants are home to a treasure trove of proprietary Big Data that is not in the public domain. Because this data is crucial for developing the sophistication of an AI engine through deep learning techniques, a market synergy between mainstream pharma and cutting-edge biotech looks likely in the future.
This will raise questions over the cost of drugs, corporate culture and the OpenAI standards that have been championed by the likes of Elon Musk.
OpenAI, launched by Elon Musk, founder of Tesla, and Sam Altman, president of tech incubator Y Combinator, aims to make patents and research in the field of AI open to the public for iteration and development.
Because of the potential power of AI to reshape government and entire industries, the organisation is also intended to act as a form of informal series of checks and balances to ensure the technology does not simply become a tool of the powerful. Should private interests in the medical sector, for instance, control both the data and the access to AI technologies, history might suggest that would be bad both for competition, and for progress as a whole.
Meanwhile Insilico, as far as we know, is the first company in the world to have developed a “Nutraceutical” that has been formulated by an AI. The technical term Zhavoronkov uses is “Generative Adversarial Network” based on research by Ian J. Goodfellow.
What if an AI doesn’t simply just locate a snapshot of the molecular change; what if it develops then builds the actual cure? If this sounds wholly unlikely, then you might want to consider that computers today can already not just locate a picture of a flower, but they can generate a composite photorealistic image of that flower that is entirely new, based on a simple text command. AI is teaching computers to imagine.
“Once we validate we will be able to open the kind of cornucopia of drugs for multiple disease indications, not just cancer, but cancer is going to be the first frontier,” states Zhavoronkov. “Our second pillar is that we want to develop much more accurate biomarkers of health status, so we’ll link everything to age.
“Very soon, we’ll launch a product that will look at multiple features of datatypes within your body and tell you how healthy you are and how long you are going to live”.
With the healthcare market in the UAE reportedly projected to reach US$19.5 billion by 2020, and an average annual growth rate of 12.7 percent and a renewed focus on preventive care due to the recent fall in oil prices, it remains to be seen whether the AI revolution in medicine will be able to develop a niche in the Middle East.