$28.6M ARPA-H Effort Will Train AI to Predict Harmful Protein Changes

A gap in what AI can predict

Neurodegenerative diseases such as Alzheimer's, Parkinson's, ALS and Huntington's are associated with proteins that change shape, stick together and eventually damage cells. By the time scientists can clearly see those changes, much of the damage may already be underway. The new effort aims to uncover what happens earlier and use that information to train AI.

The project targets a specific limit in current AI. While AI has transformed scientists' ability to predict the structures of many proteins, roughly one-third of proteins do not have a stable structure. These intrinsically disordered proteins, or IDPs, can shift among many shapes and become involved in harmful clumping, known as aggregation, that is associated with neurodegenerative diseases.

Even without a fixed structure, the behavior of IDPs is still influenced by their amino-acid sequence. That link is what the researchers want AI to learn to read.

What the Halfmann lab will do

The Halfmann lab at the Stowers Institute for Medical Research in Kansas City, Mo., will receive approximately $4.1 million over two years from the Advanced Research Projects Agency for Health, or ARPA-H. Its task is to generate large-scale experimental data on protein aggregation.

The lab will measure the aggregation tendencies of 50,000 proteins, each individually expressed in yeast cells under diverse conditions designed to mimic the things that go wrong in human cells as we age. Across more than 1 million samples, the team expects to produce more than 10 billion measurements. The work will use Distributed Amphifluoric FRET, or DAmFRET, a technology his team developed in 2018 that measures protein self-assembly inside individual living cells.

"Direct measurements of protein aggregation at cellular resolution have not previously been done at this scale," Halfmann said. According to the release, his lab's earlier studies showed that disease-relevant protein behavior seen in yeast can also inform how those proteins behave in human cells.

The project builds on earlier work. In 2023, the lab became the first to experimentally determine the structure of the initiating step in amyloid formation associated with Huntington's disease. It also builds on studies of TDP-43, a protein strongly associated with ALS and FTLD. "In those studies, we looked at hundreds of protein sequences. Now we're going to do 50,000," Halfmann said.

The NATIVE-ID team and structure

The project, called NATIVE-ID, is led by the Innovative Genomics Institute at the University of California, Berkeley, and is part of ARPA-H's BIOGAMI program, led by ARPA-H Program Manager Shannon Greene, Ph.D. Besides Stowers and UC Berkeley, the team includes scientists from Brown University, Emory University, Johns Hopkins University, Parallel Squared Technology Institute and Texas A&M University.

Team members will contribute complementary capabilities: generating human neurons from different genetic backgrounds, measuring protein interactions in neurons, developing advanced deep-learning frameworks, determining protein structures in test tubes, and running large-scale computational simulations of protein behavior.

Alejandro Sánchez Alvarado, Ph.D., Stowers President and Chief Scientific Officer, said the Berkeley colleagues "will bring the human neurons in which the model's predictions must ultimately hold."

The group will first focus on frontotemporal lobar degeneration, or FTLD, a disease that shares significant genetic and biological features with ALS. The larger goal is approaches that can apply across other diseases involving protein misfolding.

What comes next

The current award funds an initial two-year phase. BIOGAMI is structured in two 24-month phases: Phase 1 delivers the foundational datasets and models, and Phase 2 focuses on validating potential therapeutics and ways to detect protein dysfunction earlier.

Halfmann said better prediction of disease probabilities and onset ages "could allow many more people to seek preventive or early-stage treatments or enroll in clinical trials." He also said that decoding IDP interactions could allow researchers to create therapeutic IDPs to "intercept" problematic interactions that drive diseases like Alzheimer's. These are goals of the research, and the release does not report results yet.