An AI That Reads DNA Cell by Cell, Obesity Beyond BMI and Rubber Blades for Tidal Power

Read the episode (Narration script)

About two percent of your DNA is genes in the familiar sense: recipes for proteins. Most of the other ninety-eight percent is instructions about the recipes. When to use one, how much, and in which cells. Most DNA differences linked to common diseases sit out there, and working out what any one of them does has been slow bench work.

In 2024 a team led by James Lee at the Francis Crick Institute in London ran down one such difference, linked to inflammatory bowel disease. It lies in what geneticists call a gene desert, a long stretch of DNA with no genes in it at all. They showed it works like a dimmer switch, turning up a distant gene called ETS2, and specifically in macrophages, a type of immune cell, where it drives inflammation. That took a year and a half of experiments.

Yesterday Google DeepMind announced AlphaGenome 2, an AI model that reads a stretch of DNA and predicts how it behaves in a particular kind of cell. Pointed at that same region, it singled out the same variant, and predicted it would turn up ETS2 in gut macrophages and do very little anywhere else. Lee told Nature the model found in thirty minutes what took his lab eighteen months, and that he finds this reassuring.

The cell type is the advance. The first AlphaGenome predicted what a DNA change does across a blended sample of tissue. But a liver cell and a neuron carry the same genome and read it very differently, so a variant that matters in one can be silent in the other. The new model was trained on measurements from more than a hundred million individual cells, from 68 tissues. It can also be handed a few dozen cells in a state nobody has catalogued, and make predictions for those.

Reproducing a known answer is the easy test, though. On a harder one, using immune cells, the model identified the variant truly responsible for a change in gene activity 30 percent of the time. The previous best was 16. So nearly double, and still wrong more often than right. The paper hasn't been peer reviewed, and the model is for research, not diagnosis. What it changes is where the bench work begins: with a shortlist, instead of a desert.

A humbler instrument also had a good week: the tape measure. Body mass index, a ratio of weight to height, has defined obesity for decades, though it can't tell fat from muscle or say where the fat sits. Last year a Lancet commission proposed something else: confirm excess fat with a waist measurement, then separate clinical obesity, where organs or daily life are already affected, from preclinical obesity, where they aren't.

A team led by An Pan at Huazhong University of Science and Technology in Wuhan, and including the commission's own chair, has now tested that scheme on 844,000 adults in China, Britain, the United States and Spain. In Nature, they report that people who were obese by BMI but had ordinary waists, a few percent of everyone, were at no greater risk of dying than people of normal weight. Meanwhile people who were only overweight by BMI but large around the middle, between five and nine percent of everyone, did carry extra risk. BMI waves them through.

The second distinction mattered even more. Clinical obesity came with a 69 percent higher risk of death during the study. Preclinical obesity, 19 percent. And most of the preclinical group, between three-quarters and nine-tenths depending on the country, had not progressed to the clinical kind ten years on. These are observational records, showing who is at risk rather than what treatment achieves. But it's a strong case for diagnosing obesity by what fat is doing to a body, not by how much the body weighs.

Yesterday we brought you a study arguing that plate tectonics took hold about 2.7 billion years ago. On the same day, Nature published a rock with a different view. Hangqiang Xie and colleagues at the Chinese Academy of Geological Sciences describe a gneiss from Hebei province in northern China that is 3.95 billion years old. Only one other place on Earth, in northern Canada, has rocks confirmed to be older than 3.9 billion.

Its chemistry suggests it began as basalt altered by water near the surface, then carried something like fifty kilometres down, fast enough to stay cool, and partly melted there. Quick, cold burial of old seafloor is what subduction does, when one slab of crust dives beneath another. The authors call it a transient event, and that word lets both papers stand. The planet may have tried subduction in fits and starts for more than a billion years before it became the way the whole surface works.

To the sea. Tidal power has one great virtue, which is that tides run to a timetable, and one great problem: moving seawater is heavy and turbulent, and it batters turbine blades. Big turbines cope by swivelling each blade with motors and sensors, expensive machinery to keep alive underwater.

Ignazio Viola's group at the University of Edinburgh, with engineers from the tyre maker Michelin, built blades with a stiff front and a rubber trailing edge. When the current surges, the rubber bends, the blade flattens, and the extra force spills away without anything being switched on. In a 25-metre test tank, a scale-model turbine with rubber blades collected as much energy over a simulated tidal cycle as rigid blades with active control, and the rapid load swings that wear blades out fell by up to a third. Full size is the next question. The idea is borrowed from fish and trees: don't fight the flow, give a little.

Decent advice for a turbine, and not bad for a Thursday. See you tomorrow.