Hello Nature readers, would you like to get this Briefing in your inbox free every week? Sign up here.
An AlphaFold3 model of a bacterial enzyme bound to a chemical.Credit: Isomorphic Labs
AlphaFold gets major upgrade
A new version of DeepMind’s AlphaFold gives scientists the ability to predict protein structures during interactions with other molecules. The AI tool could be transformative for drug discovery because it can predict the shape of proteins that contain function-altering modifications, or their structure alongside those of DNA, RNA and other cellular players that are crucial to a protein’s duties. “This is just revolutionary,” says biochemist Frank Uhlmann. “It’s going to democratize structural-biology research.” Access to the AlphaFold3 server, however, is limited — partly to protect the advantage of DeepMind’s own drug-discovery spin-off company.
Algorithm spots 27,500 new asteroids
An algorithm called THOR (Tracklet-less heliocentric orbit recovery) has discovered 27,500 solar-system bodies by digging through hundreds of thousands of archive images of the night sky. Around 150 of these asteroids seem to be on paths that bring them close to Earth’s orbit — though none are in danger of colliding with our planet. Asteroids are discernable because they move against the backdrop of stars, an effect that usually requires two easy-to-compare images taken on the same night, by the same telescope. THOR can recognize an asteroid even in quite different images, which could speed up the work of asteroid-seeking telescopes and open up the hunt for Earth-threatening rocks to include data collected by any telescope.
The New York Times | 6 min read
AI maps brain slice in spectacular detail
Researchers have created an exquisitely detailed atlas of a tiny piece of one woman’s brain, which had been removed during surgery to treat her epilepsy. The sample was cut into thousands of nanometre-thick slices and each was imaged with electron microscopes. AI tools then classified different structures and cells, and created a 3D reconstruction of the sample. “I remember this moment, going into the map and looking at one individual synapse from this woman’s brain, and then zooming out into these other millions of pixels,” says neuroscientist and study co-author Viren Jain. “It felt sort of spiritual.”
Rendering based on electron-microscope data, showing the positions of neurons in a fragment of the brain cortex. Neurons are coloured according to size.Credit: Google Research & Lichtman Lab (Harvard University). Renderings by D. Berger (Harvard University)
Features & opinion
AI needs to see the ‘ugly’ side of science
The absence of negative results in the scientific literature is affecting AI tools trained on published data. Publishing failed experiments is often seen as not worth the time and effort, even though they play an important part in forming scientists’ intuition — something that’s lacking in AI models trained only on successful data. Techniques to make up for the scarcity, such as exposing AI models to multiple copies of available negative data points during training, can introduce fresh biases. Open data repositories and alternative journals now allow scientists to share more of their negative results. “Machine learning is changing how we think about data,” says chemist Keisuke Takahashi.
Why animals still outrun robots
Mechanical components are often more resilient and powerful than bones and muscles, yet no robot runs as fluidly as people and other animals. In a comparison of five movement subsystems — power, frame, actuation, sensing, control — robotic technologies meet or outperform their biological counterparts in all but one (control, because brains outclass computers). Animals’ advantage seems to come from integrating all of these subsystems. “Rather than focusing on the latest, newest, fanciest, most expensive component that’s going to make my robot better, maybe we could take a step back and think more carefully about the parts we have, and do better with those,” says robotics researcher and review co-author Samuel Burden.
Reference: Science Robotics review
How AI helps plan for a renewable future
How should Ghana change its infrastructure to meet its future energy needs in a renewable way? Using computer simulation linked to AI algorithms, researchers showed that expanding full speed on solar and wind power would wreak havoc on ecosystems, food security and human health. The country would need to use existing hydropower sources whenever there isn’t enough power from wind and sun, which would make the rivers prone to dangerously erratic flooding. “In the end, [the AI model] invested in a system that put quite a bit of biogas and solar power in the north of Ghana,” explains water management researcher Julien Harou.
Nature Careers Podcast | 24 min listen
Infographic of the week
Amanda Montañez; Source: Andrew Gray
The year after ChatGPT’s release, a lot of ‘AI buzzwords’ — words that appear more often in AI-generated than in human-written texts — started turning up in scientific papers. Control words such as ‘furthermore’ and ‘consider’ didn’t experience the same increase, according to a preprint analysis by librarian Andrew Gray. He estimates that at least 60,000 papers published last year could contain AI-generated text, which is slightly more than 1% of all articles. (Scientific American | 6 min read)
Reference: arXiv preprint
