Using Explainable AI To Ensure Drug Discovery Safety

Clinical drug development has a 90% failure rate – a disheartening reality for drug development scientists, pharmaceutical companies and patients alike.

Safety is a major contributing factor to this high failure rate. Many drugs progress through preclinical stages of development, only to fail when their safety is assessed in human subjects. This process costs millions of dollars and shatters hope.

While in vitro tests and animal models can give some indication of potential safety issues, they do not always translate well to humans. Ignota Labs is tackling this challenge.

The start-up company deploys a variety of different data-driven techniques to address the different components of drug safety. Its state-of-the-art in silico models can identify the root causes of toxicity – giving scientists the information they need to improve a drug’s safety profile.

At the European Laboratory Research and Innovation Group (ELRIG)’s Research and Innovation 2024 event, Technology Networks sat down with Ignota Labs’ CEO Sam Windsor to learn about the company’s proprietary platform, SAFEPATH, and the impact AI is having on drug discovery, development and commercialization.

Molly Campbell (MC): Can you discuss the different components of drug safety that Ignota Labs is developing data-driven techniques to address?

Sam Windsor (SW): Drug safety has a range of different stages of checking. There are initial checks with in vitro assays, followed by animal testing and then finally human trials. We are particularly focused on trying to get all the way to the understanding of how a drug will act in a human being; that's our main aim and that’s also what differentiates us from some of the cheminformatics platforms, which are essentially just replicating the in vitro tests.

To get to that understanding, we do a range of different things. We do have cheminformatics platforms that look at in vitro replicas, but we also have a big bioinformatics platform that integrates that data with a range of biologic data types so that we can understand how a drug might react in an animal test and in humans, and importantly, the difference between those two.

A human is not a rat, and a human is not a mouse, but the regulators insist (with good reason) on animal tests before a drug is administered to humans. You can learn things from those studies; however, you need to understand the biological differences. That way, you can achieve useful insights as to what will happen in a human being, assess if the toxicity is likely to translate across and ultimately assess if a drug project is viable.

MC: SAFEPATH is your proprietary platform. Can you explain how it works, and how quickly it can obtain information regarding toxicity compared to more traditional methods?

SW: I think the unique edge of our platform centers around that combination of cheminformatics, bioinformatics, cutting-edge AI and our unique data moat.

SAFEPATH is very good at understanding various aspects of biology in the body, such as different genes, different proteins and how they interact with each other, and can create a Knowledge Graph, organizing the data in a way that it can be queried. We then have causal reasoning on top of it, so it's not just “This piece of data was linked to this other one,” it's asking, “In what direction is there an effect, and what is causing that?” The challenge around that insight is that sometimes there are data gaps. Biology is very vast and complex, and data is often very sparse. We use the cheminformatics platform to plug in gaps where empirical data doesn't exist.

MC: Once you have the infrastructure of your data and your relationships, how do you then conduct the best analysis on top of it?

SW: We have been working very hard, scanning not only the more traditional machine learning methods that have been around for a while – spoiler alert, some of these are still the best in some situations – but, more recently, we have been leveraging some of the techniques you may have heard of since Large Language Models such as ChatGPT burst onto the scene. We have seen a notable increase in all of our relevant accuracy metrics, so we know that, actually, the value can come from using those approaches – but you need highly specialized people to do that – many of whom don’t currently work in the biotech space.

We’ve also got proprietary data that gives us an “unfair advantage”. For example, we worked with Newcastle University and the Wellcome Centre for Mitochondrial Disease to develop the world’s largest machine learning grade data sets for mitochondrial toxicity and cytotoxicity. Mitochondrial toxicity in particular is an underlying cause for lots of safety issues but can be hard to pin down with traditional methods.

MC: Ignota Labs places emphasis on identifying why a compound is toxic, which cannot be inferred by simple models and in vitro assays. Can you discuss the significance of this for drug discovery/ development?

SW: Safety is a huge cause of failure, 56% of projects are failing from preclinical trials onwards due to safety. Yet, when we started the company, safety seemed to be being overlooked as a problem area to point some of these interesting AI and data tools at.

We've had many theories as to why that might be. I think one factor is that some scientists just don’t consider it as exciting; instead, it’s seen as a barrier that you’ve simply got to get over. But if you look at the economics of safety and failure rates, it’s a huge problem. We decided that it was underserved and that it was an area we should be focusing on, which is why we went there. The further we dug into it, the more we realized that there was so much potential from an interesting science and commercial angle.

MC: AI has become a “buzzword” as of late, but drug discovery/development seems to be a space where such technologies are really having an impact right now. Can you discuss your thoughts on this?

SW: It’s a word that can be used to cover a lot of things, and it’s not surprising therefore that people who are not in the field get confused by it.

When I’m speaking to family members who aren’t working in AI, I always use an example from school to explain it. Remember when you created charts in school that plot everybody in the class’s heights against their arm span, and you draw a line straight through it? Get a computer to do that and update it over time, that’s a basic machine-learning algorithm.

AI is vast and fast moving – and takes specialism to understand the difference between cutting-edge two years ago and cutting-edge today.  

I think it's important to articulate what we also mean by an “AI drug”. People have hailed that this drug will be the “first AI-developed drug”, but in reality, there's not really any drug that's getting to market now that hasn't had some combination of what you might call AI, and traditional wet lab science; they’re all a blend.

One of the challenges around AI in drug discovery is that there are several companies who have managed to make a big song and dance around getting a drug to the clinic faster – “we managed to take something that would normally take 10 years, and we got the candidate within 10 months” – that sort of thing. From what I can tell, there is little evidence to suggest those drugs are of better quality or that their failure rates decrease once they reach the clinic.

There's some very interesting work by Abraham Heifets, the CEO of Atomwise, who looked at the different types of drugs being “found quickly” and their similarity to their nearest competitor on the market. He discovered that there are typically very tiny changes that have been made to chemical structure, but the company claims to have found a new drug 10 times faster – it’s not a fair comparison.

I think that clinical translation will become the next big focus area. Are these AI drug discovery companies really getting better quality drugs to the clinic? Or just getting average-quality drugs to the clinic faster.

We’ll be able to know this soon as they're all starting to enter the clinic now. At Ignota Labs, we are trying to get ahead of that and place emphasis on clinical translation today.

About the interviewee:

Sam Windsor is the CEO at Ignota Labs. Ignota Labs' interdisciplinary team of world-leading experts in AI, drug discovery and toxicology has created state-of-the-art in silico models that accurately predict key toxicology endpoints.