Großes AlphaFold-Upgrade bietet Auftrieb für die Arzneimittelforschung. Die neueste Version der KI modelliert, wie Proteine ​​mit anderen Molekülen interagieren – DeepMind schränkt jedoch den Zugriff auf das Tool ein.

From the article:

>To create AlphaFold3, Jumper, DeepMind chief executive Demis Hassabis and their colleagues made large changes to its predecessor: the latest version depends less on information about proteins related to a target sequence, for instance. AlphaFold3 also uses a type of machine-learning network — called a diffusion model — that is used by image-generating AIs such as Midjourney. “It’s a pretty substantial change,” says Jumper.

>AlphaFold3, the researchers found, substantially outperforms existing software tools at predicting the structure of proteins and their partners. For instance, scientists — especially those interested in finding new drugs — have conventionally used ‘docking’ software to physically model how well chemicals bind to proteins (usually with help from the proteins’ experimentally determined structures). AlphaFold3 proved superior to two docking programs, as well as to another AI-based tool called RoseTTAFold All-Atom4.

>Uhlmann’s team has used AlphaFold3 to predict the structure of DNA-interacting proteins involved in copying the genome, a step that is essential to cell division. Experiments in which proteins are mutated to alter such interactions suggest that the predictions were usually spot on, Uhlmann says. “It’s an amazing discovery tool,” he adds.

>“The structure-prediction performance of AlphaFold3 is very impressive,” says David Baker, a computational biophysicist at the University of Washington in Seattle. It’s better than RoseTTAFold All-Atom, which his team developed4, he adds.