How is AI driving drug discovery?

An unarguable fact about the modern era, is that we are awash with data; and when it comes to healthcare, the same applies, particularly since many countries have moved over the last decade onto electronic patient record systems. This means that we have massive health libraries we can trawl for new drug leads. But, paradoxically, we’ve gone from a paucity of information to a situation where we have so much data that processing it meaningfully is a problem. And this, Aaron Wenteler, from Queen Mary University of London, argues is where AI and machine learning is coming into its own, especially when it comes to drug discovery. Seeing biochemical needles in haystacks is what this technology is supremely good at, and in the next decade should see our successful new drug hit rate climb, partly by helping us to refine our choices of targets…

Aaron - One of the biggest challenges in drug development is that many clinical trials fail. This isn’t only because drugs are poorly made, but often because they target the wrong biology. The biological target—often a protein or a gene—may not actually be involved in the disease. In fact, a large portion of clinical trial failures can be traced back to the initial choice of a drug target. So, in this review paper, we essentially asked: how can we harness AI and this vast amount of biological data to improve drug target discovery and validation?

Chris - The pharmaceutical industry says it gets it right about 10% of the time. We hear the number 10 a lot—10 years, 10 billion dollars, and a 10% success rate. You're trying to improve those odds using AI?

Aaron - Yes, exactly. That’s precisely the idea. We’re motivated by the huge wealth of biological, genomic, proteomic, and clinical data available. There are many ways we can use it better to inform target discovery.

Chris - So where’s the missing link? Given we already have all that data, what are we doing wrong that’s stopping us hitting the mark more often?

Aaron - That’s a great question. First, we’re still in the early stages. Recent breakthroughs in AI have only just empowered us to do certain things that were previously impossible. AI is already being used in many biological areas—genomic sequences, protein structures, even biomedical literature through large language models. One of the biggest missing links in biology is the integration of all these different data types. That’s still not being done enough—but it’s something researchers are actively working on.

Chris - So you’re saying we’ve got all the building blocks, but as humans, we struggle to see the connections among millions of data points. This is something machine learning and AI systems are especially good at—spotting patterns we can’t?

Aaron - Exactly. Not only that—they help us spot patterns across different types of biological data. Biology is highly multimodal—operating on multiple levels: cellular, tissue, DNA. AI allows us to integrate all of this and find patterns that support drug discovery.

Chris - But we still have to train it to know what to aim for. How do you go about training it?

Aaron - Great question. It depends on the application. For example, if you’re working with electronic health records, you often have massive datasets covering tens of thousands of patients. You train the model on that data. This brings its own challenges—such as bias. Many large datasets are only representative of a small fraction of the population. We need to ensure the data we use is representative across different demographics so we build models that benefit as many people as possible. At the end of the day, your AI model is only as good as the data you train it on.

Chris - Indeed. It can spot associations, but it can’t see what we haven’t discovered yet. So there's still a role for new research and data collection. Can it help identify knowledge gaps as well?

Aaron - Absolutely. That’s how we should be thinking about AI. Media narratives often suggest AI will replace humans. We’re not there. What AI can do is help us explore the boundaries of our current knowledge—spotting patterns that humans or conventional algorithms can’t.

Chris - What’s the timeline for this? It sounds exciting and promising, but also like a huge undertaking.

Aaron - Yes, it is. It’s hard to give precise numbers, but we’re already seeing AI having an impact. In our paper, we mapped out current drug candidates from so-called AI-first drug discovery companies. We found dozens of compounds that, at the time of writing, were in clinical trials. That’s a good proxy for the impact AI is already having. So, in the next five to ten years—if you want to be conservative—I expect we’ll see AI make a significant difference.