Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain
the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in
Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles
and JavaScript.
AI & robotics briefing: AlphaFold predicts thousands of possible psychedelics
For the first time, predicted protein structures have been shown to be just as useful for drug discovery as experimentally derived ones. Plus, a plant-inspired robot that grows like a vine and how an AI figured out geometry.
A robot that 3D prints itself grows like a climbing plant, winding around structures at a few millimetres per minute. Its growth can be influenced by gravity, light and shade — for example, it can be programmed to grow towards the light. The researchers were inspired by plants’ ability to “conquer very challenging and mutable environments,” says roboticist Emanuela Del Dottore, who co-developed the machine.
The protein-structure-prediction AI AlphaFold has helped to identify thousands of potential psychedelic molecules, which could help to develop new antidepressants. This is the first time predicted protein structures have been shown to be just as useful for drug discovery as experimentally derived ones, which can take months to determine. In one out of three cases, an AlphaFold prediction could jump-start a drug discovery project by a couple of years, estimates pharmaceutical chemist and study co-author Brian Shoichet. Others caution that AI-generated molecules seem to be helpful for some biological targets but not others, and it’s not always clear which applies.
In a small study, a Google AI system was more accurate than human physicians in diagnosing respiratory and cardiovascular conditions. It was also rated more empathetic by the 20 testers who didn’t know whether they were texting a human or a machine. However, the chatbot hasn’t been evaluated for biases or tested by people with real health problems. “We want the results to be interpreted with caution and humility,” says clinical research scientist and study co-author Alan Karthikesalingam.
The World Health Organization (WHO) warns that developing and deploying health-care AI mustn’t be left to tech companies and those in wealthy countries. This could lead to a “race to the bottom” in which firms seek to be the first to release applications, even if they are dangerous or simply useless, particularly for people in lower-income countries. The WHO’s new guidelines recommend, among other things, mandatory audits of medical algorithms to ensure they protect both data and human rights.
Radiologists collaborating with a cancer-screening AI system are around 3% better at detecting breast cancer than when they work alone. The AI tool screened almost 83,000 old mammograms, flagging those that seemed concerning to a human. When the AI tool was left to make assessments on its own, it performed worse than a doctor. “In the proposed AI-driven process, nearly three-quarters of the screening studies didn’t need to be reviewed,” which could be particularly useful where there is a shortage of specialists, says radiology researcher Curtis Langlotz.
There is little research on whether robotic arms are better than human hands holding surgical instruments. Robots are frequently used when it’s difficult to view the area, such as in urology, gynecology, and ear, nose and throat procedures. Patients who undergo robotic surgery often have less postoperative pain and can leave the hospital sooner. But cost is a major issue: an average robotic platform costs €1.4 million, plus supplies and maintenance. “The real advantages will be if we can get robots to do something we cannot physically do,” says neurosurgeon Peter Konrad — reconnecting nerve fibres, for example.
AI is redefining intelligence as succeeding in tasks that rely on pattern recognition and prediction from data, writes a trio of communication researchers. This, they argue, limits intelligence to the ability to look backward and condemns us to repeat past mistakes. “We fear this could set limits for human aspirations and for core ideals like knowledge, creativity, imagination and democracy — making for a poorer, more constrained human future.”
An AI tool called AlphaGeometry could (theoretically) win a bronze medal in the International Mathematical Olympiad. When presented with maths problems, existing large language models often struggle to make sense when asked to show their workings. So the team trained their model from scratch. “We were able to generate 100 million theorems and proofs so that the machine can learn all of these by itself, and then it can learn to generalize the new problems,” deep learning researcher and former maths Olympiad competitor Thang Luong tells the Nature Podcast. To win an Olympiad gold medal, the algorithm would have to become equally good at disciplines beyond geometry, such as number theory.
AI can help researchers speed through repetitive tasks — but it would be better to reassess whether these tasks benefit quality research in the first place, argue education researcher Richard Watermeyer, anthropologist Donna Lanclos and education technology specialist Lawrie Phipps. (Nature | 3 min read, Nature paywall)
