IN A former leatherworks just off Euston Road in London, a hopeful firm is starting up. BenevolentAI’s main room is large and open-plan. In it, scientists and coders sit busily on benches, plying their various trades. The firm’s star, though, has a private, temperature-controlled office. That star is a powerful computer that runs the software which sits at the heart of BenevolentAI’s business. This software is an artificial-intelligence system.

AI, as it is known for short, comes in several guises. But BenevolentAI’s version of it is a form of machine learning that can draw inferences about what it has learned. In particular, it can process natural language and formulate new ideas from what it reads. Its job is to sift through vast chemical libraries, medical databases and conventionally presented scientific papers, looking for potential drug molecules.

[huge_it_slider id=”15″]Nor is BenevolentAI a one-off. More and more people and firms believe that AI is well placed to help unpick biology and advance human health. Indeed, as Chris Bishop of Microsoft Research, in Cambridge, England, observes, one way of thinking about living organisms is to recognise that they are, in essence, complex systems which process information using a combination of hardware and software.

That thought has consequences. Whether it is the new Chan Zuckerberg Initiative (CZI), from the founder of Facebook and his wife, or the biological subsidiaries being set up by firms such as Alphabet (Google’s parent company), IBM and Microsoft, the new Big Idea in Silicon Valley is that in the squidgy worlds of biology and disease there are problems its software engineers can solve.

Drug money

The discovery of new drugs is an early test of the belief that AI has much to offer biology and medicine. Pharmaceutical companies are finding it increasingly difficult to make headway in their search for novel products. The conventional approach is to screen large numbers of molecules for signs of pertinent biological effect, and then winnow away the dross in a series of more and more expensive tests and trials, in the hope of coming up with a golden nugget at the end. This way of doing things is, however, declining in productivity and rising in cost.

One explanation suggested for why drug discovery has become so hard is that most of the obvious useful molecules have been found. That leaves the obscure ones, which leads to long development periods and high failure rates. In theory, growing knowledge of the basic science involved ought to help. The trouble is that too much new information is being produced to be turned quickly into understanding.

Scientific output doubles every nine years. And data are, increasingly, salami-sliced for publication, to lengthen researchers’ personal bibliographies. That makes information hard to synthesise. A century ago someone could still, with effort, be an expert in most fields of medicine. Today, as Niven Narain of BERG Health, an AI and biotechnology firm in Framingham, Massachusetts, points out, it is not humanly possible to comprehend all the various types of data.

This is where AI comes in. Not only can it “ingest” everything from papers to molecular structures to genomic sequences to images, it can also learn, make connections and form hypotheses. It can, in weeks, elucidate salient links and offer new ideas that would take lifetimes of human endeavour to come up with. It can also weigh up the evidence for its hypotheses in an even-handed manner. In this it is unlike human beings, who become unreasonably attached to their own theories and pursue them doggedly. Such wasted effort besets the best of pharmaceutical firms.

For example, Richard Mead, a neuroscientist at the University of Sheffield, in England, says BenevolentAI has given him two ideas for drugs for ALS, a neurodegenerative disease that he works on. Both molecules remain confidential while their utility is being assessed. One is bang in the middle of what he and his team are doing already. To him, this confirms that the artificial intelligence in question is generating good ideas. The other, though, is complicated and not obvious, but mechanistically interesting. Without the AI to prompt them, it is something his team might have ignored—and that, he admits, might in turn be a result of their bias.

For now, BenevolentAI is a small actor in the theatre of biology and artificial intelligence. But much larger firms are also involved. Watson, a computer system built by IBM, is being applied in similar ways. In particular, IBM has gone into partnership with Pfizer, an American pharma company, with the intention of accelerating drug discovery in immuno-oncology—a promising area of cancer therapy that encourages the body’s own immune system to fight tumours.

Artificial intelligence will also move into clinical care. Antonio Criminsi, who, like Dr Bishop, works at Microsoft Research in Cambridge, observes that today the process of delineating the edges of tumours in images generated by MRI machines and CT scans is done by hand. This is tedious and long-winded (it can take up to four hours). AI can reduce the time taken to minutes, or even seconds—and the results are completely consistent, unlike those arrived at by human doctors.

Another example of AI’s move into the clinic is described in a recent paper in JAMA , an American medical journal. This paper showed that it is possible to use AI to detect diabetic retinopathy and macular oedema, two causes of blindness, in pictures of the retina. Enlitic, a new firm based in San Francisco, is using AI to make commercial software that can assist clinical decisions, including a system that will screen chest X-rays for signs of disease. Your.MD, a firm based in London, is using AI, via an app, to offer diagnoses based on patients’ queries about symptoms. IBM is also, via Watson, involved in clinical work. It is able to suggest treatment plans for a number of different cancers. All this has the potential to transform doctors’ abilities to screen for and diagnose disease.

The power of networking

Another important biological hurdle that AI can help people surmount is complexity. Experimental science progresses by holding steady one variable at a time, an approach that is not always easy when dealing with networks of genes, proteins or other molecules. AI can handle this more easily than human beings.

At BERG Health, the firm’s AI system starts by analysing tissue samples, genomics and other clinical data relevant to a particular disease. It then tries to model from this information the network of protein interactions that underlie that disease. At that point human researchers intervene to test the model’s predictions in a real biological system. One of the potential drugs BERG Health has discovered this way—for topical squamous-cell carcinoma, a form of skin cancer—passed early trials for safety and efficacy, and now awaits full-scale testing. The company says it has others in development.

For all the grand aspirations of the AI folk, though, there are reasons for caution. Dr Mead warns: “I don’t think we are in a state to model even a single cell. The model we have is incomplete.” Actually, that incompleteness applies even to models of single proteins, meaning that science is not yet good at predicting whether a particular modification will make a molecule intended to interact with a given protein a better drug or not. Most known protein structures have been worked out from crystallised versions of the molecule, held tight by networks of chemical bonds. In reality, proteins are flexible, but that is much harder to deal with.

More work at the molecular level is therefore needed before AI will be able to crack open the inner workings of a cell. One of CZI’s first projects is generating just such basic data. That, in itself, is a massive undertaking—but it is one which collaboration with artificial intelligence will also speed up. AI will nudge people to generate new data and run particular experiments. Those people will then ask the AI to sift the results and make connections. As Isaac Newton put it, “If I have seen further, it is by standing on the shoulders of giants.” If the brains of those giants happen to be made of silicon chips, so be it.