APPLES, mushrooms and pork sounds a promising recipe for a kebab, but the average barbecuer might balk at adding strawberries. According to John Gordon of IBM, however, the result is delicious. Dr Gordon is one of the leaders of that firm's cognitive-computing team, responsible for a machine called Watson which is able to digest and analyse large amounts of English text and then draw inferences from it. When, in March, Watson was fed reams of recipes and texts about food, it reasoned that these four ingredients would complement each other, based on their sharing a number of flavoursome chemical compounds. And Dr Gordon, at least, thinks Watson's suggestion is a winner.
Devising new recipes sounds a trivial use for a multimillion-dollar piece of kit. But Dr Gordon's culinary experiment neatly demonstrates the idea of automated hypothesis generation—and the possible uses of that are certainly not trivial. More than 90 groups of scientists are now developing hypothesis-generation software. They hope to use it not on recipe books but on the vast corpus of scientific literature (by one tally at least 50m scientific papers) that has piled up in public databases.
The power of the technique was demonstrated by research published in August by Olivier Lichtarge of Baylor College of Medicine, in Houston, Texas, and his colleagues. In collaboration with Dr Gordon's group, they employed it to hunt for proteins called kinases that activate another protein, p53, which curbs the growth of cancers. They used the software to read the abstracts of 186,879 papers and produced a list of the most promising kinases for experiments. The twist was that the papers in question were all published before 2003. That meant Dr Lichtarge could check to see if the Watson-based approach came to the same conclusions as those arrived at by human researchers over the subsequent ten years. And it did. Of the top nine kinases the software picked, seven have subsequently been shown to activate p53.
Anne Poupon of the French National Institute for Agricultural Research, in Tours, heads another group working on automated hypothesis generation. Her software, Méthode d'Inférence, crunches research on hormones and the 1,500 types of receptor molecules with which they interact. Sometimes, it recommends looking more closely at certain of these interactions because the literature on them contains contradictory results that need to be resolved. On other occasions it deems interactions ripe for closer examination because the hormone and receptor types involved are similar to those of known pairings that have already proved medically valuable. Even though Méthode d'Inférence is still a work in progress, it has already prevented the duplication of work within the institute and has produced a novel hypothesis about the mode of operation of follicle-stimulating hormone (a substance that helps, among other things, to govern the menstrual cycle).
A third example of automated hypothesis generation at work is brainSCANr, devised by Bradley Voytek of the University of California, San Diego, and his wife Jessica. BrainSCANr is designed to help neuroscientists choose research projects. It has, among other things, revealed a promising path for migraine research. By sifting through more than 3.5m papers, the software suggested that clues to the origin of migraines may be found in the levels of serotonin, a signalling molecule, that are released by neurons in a region of the brain called the striatum.
In Dr Lichtarge's view, hypothesis-generation software works in part because science writing tends to be free of humour, sarcasm and "emotive or literary overlay" that could trip it up. That points to another source of text it can analyse in search of hypotheses to test. Web searches typically lack complex grammar and even verbs that could confuse software. By examining words typed into Microsoft's Internet Explorer web browser and Bing search engine by people wondering why they feel ill, computers at Microsoft Research, in Redmond, Washington, are producing hypotheses on potentially harmful pairings of medications. Eric Horvitz, Microsoft Research's head, says America's Food and Drug Administration has formed a team to use these "early warning" hypotheses to produce better designs for laboratory experiments on potentially dangerous drug combinations thus revealed.
It all, then, looks rather promising, both for science and for IBM—which launched a commercial version of its automated-hypothesis-generation software in August. Dr Gordon hopes Discovery Advisor, as this service is known, will be a money-spinner for the firm. If that does happen, it will probably be because it has also proved an ideas-spinner for science.