Roughly one-third of all human proteins have no stable shape, and that structural disorder is precisely what makes them so difficult for AI to read and so dangerous when they malfunction. The Halfmann lab at the Stowers Institute for Medical Research will receive approximately $4.1 million of a $28.6 million ARPA-H BIOGAMI program award to generate what its investigators describe as the largest aggregation dataset ever attempted in biomedical research: more than 10 billion measurements across 50,000 proteins, each captured one living cell at a time.
The scale matters because current AI protein tools were built on structure prediction, which works well for proteins with a fixed shape. Intrinsically disordered proteins, or IDPs, shift among many conformations and can clump in ways that current models cannot anticipate. That clumping, known as aggregation, sits at the center of Alzheimer’s, Parkinson’s, ALS, and Huntington’s disease. Despite decades of research, approved therapies for most of these conditions remain narrow in scope: lecanemab and aducanumab represent the FDA-approved disease-modifying options for Alzheimer’s, and for ALS the field has worked largely with riluzole and edaravone since the 1990s. Earlier, more precise prediction of when proteins go wrong is one plausible route to changing that picture.
Halfmann’s lab will use DAmFRET, a fluorescence-based technology it developed in 2018, to measure protein self-assembly inside individual yeast cells under conditions designed to mimic what happens in aging human cells. The group’s prior work showed that disease-relevant aggregation behavior observed in yeast translates to human cells, which is the methodological bet the entire dataset rests on. The NATIVE-ID project, led by the Innovative Genomics Institute at UC Berkeley, will pair that yeast-derived data with human neurons, and teams at Brown, Emory, Johns Hopkins, Texas A&M, and Parallel Squared Technology Institute will contribute deep-learning frameworks and structural analysis. The collaboration’s initial target is frontotemporal lobar degeneration, a disease that shares genetic features with ALS, with the intent to build methods that generalize across protein-misfolding conditions.
The number to watch as this project runs is not the dollar figure but the measurement count: whether 10 billion aggregation datapoints actually improve a model’s ability to predict IDP behavior in human neurons is the experimental question the next two years of ARPA-H funding will either answer or reframe.
Jon Napitupulu is Director of Media Relations at The Clinical Trial Vanguard. Jon, a computer data scientist, focuses on the latest clinical trial industry news and trends.
