Retorno só o título traduzido. Não adicione explicações, notas ou texto extra Por décadas, eles têm acertado o recorde em biologia. Próximo: crise de reprodutibilidade da ciência.

Even after decades of study, scientists are still divided on what drives diseases such as Alzheimer’s. Why? Turns out, the literature is littered with different hypotheses, some pointing in different directions. How do we decide which ones are worth believing?

That’s a question University of Maryland professor John Moult has wanted to figure out for years – how to decide what evidence in the literature to trust. Which papers support which hypotheses on, say, how the APOE4 gene affects Alzheimer’s? Which experiments support the conclusions of those papers? Which of those experiments – whether in humans, mice, or cells – have more dubious or more reliable conditions and statistical analyses? Could we rule out some of the major hypotheses about the APOE4 gene and get to treatments faster, if we only knew which experiments to trust?

Moult has more than three decades of experience winnowing the wheat from the chaff in structural biology – he helped found the Critical Assessment of Structure Prediction, or CASP. It’s the blind challenge measuring stick that paved the road on which DeepMind proved the true prowess of its tool AlphaFold, and eventually led to DeepMind’s 2024 Nobel Prize in chemistry. With the advent of large language models, he thinks perhaps it is now finally possible to bring a similar objective measuring rod to bear on the scientific literature.