Intelligent Drug Discovery

Pharma must be the only industry on the planet with a 98 percent failure rate. One recent study found that 52 percent of drugs terminated at phase II or III were because of a lack of efficacy (1), which tells me we’re either not getting the right molecules or not getting the right targets – I believe the latter is more likely. Worryingly, the study also found that 25 percent of terminations were due to strategic or commercial reasons, which suggests the industry is also not very good at making key decisions.

Having spent most of my career in the pharma industry – working for GSK and running a center of excellence for drug discovery, – I’ve seen first-hand how decisions are made without having access to all the available evidence, often relying on people’s narrow perspectives. Knowledge is incredibly important, yet our ability to mine it is limited.

Artificial intelligence bridges that gap rather well. It can enable the industry to use evidence without bias, enhancing our ability to make good decisions – this is the essence of Benevolent AI.

We have built an AI platform that can crawl through scientific papers, patents, and structure databases, and recognize and ground certain entities. Based on a dictionary we compiled, the platform can recognize, for example, that CB2 is a cannabinoid receptor (and not a postcode in Cambridge, UK!). It then uses natural language processing to look at the sentences and paragraphs to identify a relationship between one entity and another: is this receptor related in any way to a certain enzyme? Does gene X downregulate protein K? The AI can also determine that “AD” in the context of one paper referring to “atopic dermatitis,” but in another context referring to “Alzheimer’s disease.” We take all that known information to build a knowledge graph consisting of over one billion relationships. The idea is to ask, given this known information, what can be inferred about what should be known, but currently isn’t. Essentially, the platform is a hypothesis generator, whereby we can link new targets and molecules to different diseases. It also allows us to mine for negative information, which quite often only appears in very obscure journals or meeting abstracts. Once we have a hypothesis, the scientists can see if it makes sense, and then investigate it. It augments the ability of a scientist to come up with new ideas – using information from outside of their limited sphere of knowledge.

Typing a disease into the platform will pull up hundreds of clinical trial results, thousands of related diseases, thousands of molecules, as well as symptoms and potential targets – and this only takes a few minutes. We’re currently collaborating with experts to validate some hypotheses, and we’re looking to see whether we can repurpose drugs or start our own drug development programs.

Doing things differently

Our platform is just one example of how AI can be used in pharma, but we can’t change the industry by ourselves. I think often the issue of being slow to adopt new technology in pharma isn’t tech readiness, but people readiness. There are a lot of organizational structures that have been built up within pharma and they must be deconstructed to allow some of these emerging technologies to flourish. I’ve been through a couple of mergers in pharma and you need strong direction from the top to say, “This is how it’s going to be done differently.” Without the support of leadership, things stay as they are.

It’s always going to be difficult to change things from within and that’s why we’ve created a completely different structure at our company. We have cross-functional teams with drug developers, engineers and data analysists all working together. We like to see ourselves as “discoverers,” as opposed to “tech people” or “engineers.” Sometimes it’s hard for people to get their heads around these latter terms as applied to drug discovery.

Having said that, AI is already an integral part of healthcare. People track and analyze the data generated by wearable tech, and AI is being used to stratify patients for personalized medicine. In addition, pharma companies are increasingly thinking about how they might investigate the data that they generate – especially from previously unsuccessful compounds.

With any new technology there are always people concerns. For example, there is always concern that technology and automation may take people’s jobs and livelihoods, but the 98 percent failure rate demonstrates that pharma’s current model is unsustainable. The increasing number of mergers we are seeing isn’t the solution – we need more innovation. In my view, the only way we’re going to become more innovative is to mine and use data more effectively.

Jackie Hunter is CEO at BenevolentBio, the bioscience arm of BenevolentAI.