YouTube video by Pawthon Labs

As we bring forth our friend AI in a universe designed for cognition.

A consortium of pharmaceutical companies has demonstrated that incorporating proprietary protein structure data into AI models significantly improves their performance in predicting protein-drug interactions. The study, conducted by AbbVie, Astex, and other firms through the AI Structural Biology (AISB) Network, trained an AI model called OpenFold3 on over 20,000 proprietary protein structures bound to potential drugs, outperforming both public data-only models and those trained on individual company datasets. The AISB model achieved high accuracy on more than half of 1,056 test protein-ligand structures, compared to one-third for the publicly available OpenFold3 version and around 40% for another open-source model. This research highlights a critical gap in public databases like the Protein Data Bank (PDB), which contains relatively few examples of protein-drug interactions—estimated at around 10,000 structures—compared to the vast internal repositories of pharmaceutical companies. The findings support the need for publicly accessible datasets to further advance AI-driven drug discovery, with initiatives like the OpenBind project, funded by UK government with up to £8 million, already releasing hundreds of new protein structures. The study, described in a blog post but not yet peer-reviewed, underscores the potential of private industry data to address limitations in current AI protein-folding models, which are increasingly vital for accelerating drug development and understanding molecular interactions.

Conformational changes in proteins are vital to their function yet remain challenging for state-of-the-art artificial intelligence, such as AlphaFold3, to predict. Researchers at the Institute for Mol...

Last weekend, I found myself explaining my molecular research to an elderly person over tea and cookies. As I watched her eyes glaze over at terms like "protein folding" and "molecular dynamics," I re...