Artificial intelligence (AI) is certainly in the news constantly; however, it’s been used in drug discovery for some time. A new collaboration between artificial intelligence drug discovery company Insilico Medicine and University of Toronto biochemist and molecular geneticist Igor Stagljar will test AI-designed molecules against “undruggable” cancer targets.
The research will test 15 to 20 undruggable targets – including KRAS, the most frequently mutated cancer-causing gene (or oncogene), across all cancer types.
As many as 85% of all human proteins are thought to be “undruggable.” This is because they have smooth surfaces that lack easy pockets for small molecule drugs to bind, making drug design a huge challenge.
But are these targets undruggable, and how does AI work in the drug discovery process?
This week, we have a conversation with Kyle Tretina, Alliance Manager of AI Platforms at Insilico Medicine, on a wide range of subjects including drug discovery, undruggable targets, the collaboration with the University of Toronto, and more.
Table of contents
About Insilico Medicine
Insilico Medicine is a global clinical stage biotechnology company powered by generative AI. The company has developed AI platforms that utilize deep generative models, reinforcement learning, transformers, and other modern machine learning techniques for novel target discovery and the generation of novel molecular structures with desired properties.
Insilico Medicine is developing breakthrough solutions to discover and develop innovative drugs for cancer, fibrosis, immunity, central nervous system diseases, infectious diseases, autoimmune diseases, and aging-related diseases.
AI has been used in drug discovery for around 10 years. Tretina said in the mid 2010s, new deep learning systems started outperforming humans on tasks such as image recognition, voice recognition and text recognition.
“In 2019, Insilico launched their own platform, including a SaaS product. They also work in the service spaces as well. As well as in 2019, they started their own therapeutic programs. This was really important for the company because especially at that time, some companies were focused mostly on being an AI software company,” Tretina said.
“Other companies were really focused on building out their pipeline. But Insilico Medicine always really focused on both. It turned out to be a competitive advantage because you can use the pipeline to validate your software.”
Since 2019, Tretina said, there has been an explosion at the company in terms of ongoing commercialization and the development of generative AI algorithms.
“We now have a robust and growing pipeline; 31 programs against 29 targets, all developed internally with our own platforms.”
AI’s promise in drug discovery and development
There may be fears over the use of AI in some circles, however, Tretina said in terms of really hard problems like drug discovery and development with masses of data, AI outperforms humans.
“This is not something that humans are very good at,” he noted.
“Most drugs fail. Something like 90% fail. If you look at the percentage of things that even make it into clinical trials, it’s even worse than that. We have to be able to see patterns that humans just can’t see clearly.
“And also, there’s a repetitious aspect to certain parts of drug discovery and development in terms of data analysis, aggregation, and building systems that AI can help with in terms of efficiency. AI is kind of built for this. It’s just a matter of shaping it in the right way to actually make it effective.”
Insilico’s Pharma AI suite: Revolutionizing drug development
Tretina explained that the company utilizes the Pharma AI suite, which is comprised of three separate platforms. There is Biology 42, which focuses on identifying the targets the company is going to develop drugs against. This also looks at biological questions such as the relevance and impact of targeting different parts of biology and different diseases, different pathways or specific genes, or the impact of certain drugs on that biology.
There is also Chemistry 42, which focuses on making small molecules that are going to be effective in the clinic. The third component is Medicine 42, which is focused on clinical trial design and prediction.
Tretina said the company picks its own targets and designs the drugs. When it gets to the point when a drug is ready to go into the clinic, Insilico partners with companies to get the drugs into the clinic and through the next stage.
“It’s kind of a magical process that usually happens from both ends. We have our public pipeline on our web page. Anyone can go to that web page, click on the different molecules that are ready to be discussed with other groups.”
Picking drug targets: The vital role of multi-omics data
Picking a drug target, Tretina emphasized, is the most important part in the process. He said it is also the starting point.
“Our general approach at a really high level is to incorporate huge amounts of multi-omics data and text data. Multi-omics data are derived from human patients with or without the disease. And really what we’re looking at is at the molecular level, what are all the different changes that happen during disease states in the relevant tissues?
“And we’ve also looked at text data as well, which we think is actually really, really important. Kind of undervalued by people who maybe aren’t as familiar with drug discovery and development.”
Tretina said that often, companies will want to look at the competitive landscape of a given target or indication when deciding if they want to invest in it further.
“In developing a drug, there’s sort of the biology of research, but there’s also the business side of things. Companies want to make sure that if they pick that target, it’s going to fit into their portfolio, that it fits according to their expertise, that it fits into their mission.”
Leveraging AI for tailored molecule creation
The company tries to design small molecules that will target in a very specific, selective, and strong way to change the biology.
Tretina explained that, traditionally, if there is a target, and a company wants to design a drug, they will use a ‘docking approach.’
“AI is kind of built for this. It’s just a matter of shaping it in the right way to make it effective.”
“We have these libraries, either previously-made molecules or perhaps molecules that could be easily synthesized, and there’s slight variance. And then on a computer, they’ll simulate those molecules binding to the correct part of the protein. Now, this works and has worked in some cases, but the power of generative AI is that it can look at the problem and imagine new solutions.
“AI in this particular space can be extremely useful because it’s thinking of things that haven’t been imagined before.
“For example, how novel do we want it to be? How potent do we want it to be? How metabolically stable and druggable? We can also bias for safety measures. And then it imagines molecules from scratch, pretty much indefinitely. You can run these cycles of imagination as many times as you want. And each cycle will generate some number of molecules based on how you configure it, and then you go into the lab and actually test and validate them.”
Harnessing AI for undruggable targets
Tretina believes that an undruggable target simply means we can’t drug it at the moment.
“There’s a lot of different things that can make a target undruggable, and AI can really help with quite a few of them. So first would be structural unavailability. When a protein is acting in a signaling pathway that is related to the development of a disease, there’s usually certain parts of the protein that are involved in that process more than other parts.
“If it’s an enzyme, there’s maybe a little pocket that is involved in changing other molecules and enabling signaling. So it might be actually the case that the drug target may be structurally unsuitable for drug binding within that pocket where you want the small molecule to bind for chemical reasons.
“This is really where I think AI has been extremely powerful because of that imaginative ability. With AI, you can usually get to an answer much faster. Some targets are also a bit more complex in how they’re involved in various biological processes, where targeting something, it might effectively treat the disease, but it might also cause toxicity just by the nature of the fact that that one protein is involved in some really important biological process.”
Navigating target selectivity and disease heterogeneity
Tretina explained that this is an area still being worked on. He also stressed the issue of selectivity.
“The drug target may not be unique to the disease-causing cells if it is a certain cell that’s causing the disease. Disrupting that target may also affect other healthy cells. Even within a certain cell, there may also be other targets that look very similar to the one that you’re trying to target. And so coming up with this idea of finding something that’s going to only target what you want to target is really difficult.”
To make matters even more complicated, Tretina said there may be physical barriers, functional redundancy, and drug resistance to consider.
“In general, heterogeneity of diseases is a problem. Disease is often classified by how it looks at the patient level. And maybe at the molecular level, there are some really major differences between patients. But if you treat a group of patients with a drug, maybe only some of those patients will actually respond to that drug because of what’s actually causing it at the molecular level,” Tretina said.
The future of AI in drug discovery
AI is, of course, still developing. As it evolves, it is possible that drugs will be discovered even more quickly, and at a lower cost. However, the same process for clinical trials needs to be followed, and there are no shortcuts. And there will still be failures.
“I think a lot of our expertise has been early in the drug discovery process where we believe the biggest impact in terms of chance of success can come through. When you look at why do clinical trials fail, focusing in especially on small molecules, there are really two overwhelming reasons. They’re not effective or they’re not safe,” Tretina said.
“We think that in both of those cases, a lot of these drugs are not effective and/or not safe because they’re picking the wrong target.”
He added that in trials, much of the inefficiency involves patient selection, trial organization, training, and data recording.
“We’re still talking several years here. I think down the road, that’s really the area where AI could have the biggest impact on delivering drugs faster to patients.”
To learn more about this topic
Here are some links to more articles on the subject of AI, undruggable targets, and drug discovery.