Healthcare is one of the sectors likely to see the greatest benefits from the application of advanced AI. A number of companies are now using AI to develop drugs faster, cheaper, and with fewer failures along the way. One of the leading members of this group is Insilico Medicine, which has just announced the first AI-developed drug to enter phase 2 clinical trials. Alex Zhavoronkov, co-founder of Insilico Medicine, joined the London Futurists Podcast to explain the significance of this achievement.
Idiopathic Pulmonary Fibrosis
The drug in question is designed to tackle Idiopathic Pulmonary Fibrosis, or IPF. “Fibrosis” means thickening or scarring of tissue, and “pulmonary” refers to the lungs. The walls of the lungs are normally thin and lacy, but IPF makes them stiff and scarred. It is a common disease among the over-60s, and is often fatal.
Insilico is unusual among the community of AI drug development companies in that most of them go after well-known proteins, whereas Insilico has identified a new one. In 2019, Insilico’s AIs identified a number of target proteins which could be causing IPF, by scouring large volumes of data. They whittled the number down to 20, and tested five of them, which resulted in one favoured candidate. They proceeded to use another set of AI models to identify molecules which could disrupt the activity of the target protein. This second step involved the relatively new type of AI that is called generative AI.
GANs and GPTs
The first generative AIs were introduced in 2014 (the same year that Zhavoronkov founded Insilico Medicine), and are known as Generative Adversarial Networks, or GANs. This involves two AI models competing with each other – one to create an image, and the other to criticise it until it is essentially perfect. The second, and better-known class of generative AIs are transformer AIs, which were introduced in a 2017 paper by Google researchers called “Attention is all you need.” These are familiar to us all from ChatGPT and GPT-4: GPT stands for Generative Pre-trained Transformer.
To identify a molecule which can disrupt the target protein, Insilico gives the crystalline structure of the protein to as many as 500 different generative AI models, and instructs them to design molecules which will bind with the protein productively. Over a few days, these models compete to find the best molecule for the job. Human chemists in around 40 Contract Research Organisations (CROs), mostly in China and India, review the most promising 100 or so of the resulting molecules, and around 15-20 of them are synthesised and tested. The characteristics of the best performing molecules are fed back into the array of generative AI systems for further review. This was all done in 2019.
The resulting molecules were tested for both efficacy and safety in mice and other animals, including dogs. By 2021 the company was ready for phase zero of the clinical trial process, which was a preliminary test for safety in humans, conducted on eight healthy volunteers in Australia. This was followed by a phase one clinical trial, which is a large-scale test for safety in humans. This was carried out on healthy volunteers in New Zealand and China, and had to be particularly thorough because IPF is a chronic condition rather than an acute one, so people will be taking a drug for it for years rather than weeks or months.
Now, Insilico is able to proceed to the phase two study, dosing patients with IPF in China and the USA. Part of the challenge at this point is to find a large number of patients with good life expectancy, and the company is still recruiting.
Savings and consolidation
Overall, Zhavoronkov thinks that Insilico has shaved a couple of years off the six-year discovery and development process. But more importantly, 99% of candidate molecules fail, so the most important improvement offered by AI drug discovery and development lies in reducing this failure rate.
A couple of years ago, the community of companies applying AI to drug development consisted of 200 or so organisations. Biotech was a hot sector during Covid, with lots of money chasing a relatively small number of genuine opportunities. Some of that heat has dissipated, and investors have got better at understanding where the real opportunities lie, so a process of consolidation is under way in the industry. Zhavoronkov thinks that perhaps only a handful will survive, including companies like Schrödinger Inc., which has been selling software since the 1990s, and has moved into drug discovery.
New technologies, new opportunities
For the companies that survive this consolidation process, the opportunities are legion. For instance, Zhavoronkov is bullish about the prospects for quantum computing, and thinks it will make significant impacts within five years, and possibly within two years. Insilico is using 50 qubit machines from IBM, which he commends for having learned a lesson about not over-hyping a technology from its unfortunate experience with Watson, its AI suite of products which fell far short of expectations. Microsoft and Google also have ambitious plans for the technology. Generative AI for drug development might turn out to be one of the first really valuable use cases for quantum computing.
The arrival of GPTs has made Zhavoronkov a little more optimistic that his underlying goal of curing aging could be achieved in his lifetime. Not through AI-led drug discovery, which is still slow and expensive, even if faster and cheaper than the traditional approach. Instead, GPTs and other advanced AIs hold out the promise of understanding human biology far better than we do today. Pharmaceuticals alone probably won’t cure aging any time soon, but if people in their middle years today stay healthy, they may enjoy very long lives, thanks to the technologies being developed today.