At the center of the experimental effort is Randal Halfmann, Ph.D., an Investigator at the Stowers Institute for Medical Research in Kansas City, whose laboratory has been selected to generate the project’s large-scale data on protein aggregation. His lab will receive approximately $4.1 million over two years as part of the initiative, which is called NATIVE-ID and is led by the Innovative Genomics Institute at the University of California, Berkeley. The work falls under ARPA-H’s BIOGAMI program, a broader effort to understand and ultimately control harmful protein aggregation, led by ARPA-H Program Manager Shannon Greene, Ph.D. Halfmann’s role is to supply the raw experimental fuel that the rest of the team will convert into predictive models.

The reason such a massive data-generating effort is needed lies in a fundamental blind spot in modern computational biology. Artificial intelligence has transformed scientists’ ability to predict the three-dimensional structures of many proteins, but roughly one-third of all proteins lack a stable structure altogether. These molecules, known as intrinsically disordered proteins, or IDPs, do not hold a single fixed shape. Instead, they shift among many conformations, and in doing so they can become entangled in harmful clumping, a process called aggregation that is closely associated with neurodegenerative disease. For AI systems trained largely on well-ordered structures, reading the behavior of these shape-shifting proteins remains an enormous challenge.

Halfmann believes the payoff for cracking that challenge could be transformative. Treatments for neurodegenerative diseases remain extremely limited, and most interventions arrive only after symptoms appear. If researchers could predict the probabilities and onset ages of disease from a person’s protein sequences, he argues, far more people could seek preventive or early-stage treatments or enroll in clinical trials before irreversible damage occurs. He has even proposed a more futuristic application: designing therapeutic IDPs that could intercept the problematic interactions driving diseases like Alzheimer’s, essentially fighting disordered proteins with engineered ones.

The key insight underlying the project is that even without a fixed structure, the behavior of an IDP is still governed by its amino-acid sequence. Small changes in that sequence can tip a protein between remaining soluble and beginning to clump. Halfmann’s lab will conduct large-scale experiments in yeast cells to test how sequence variations affect this balance, building on prior studies demonstrating that disease-relevant protein behavior observed in yeast can reliably inform how those same proteins behave in human cells. The workhorse technology is Distributed Amphifluoric FRET, or DAmFRET, a method his team developed in 2018 that measures protein self-assembly inside individual living cells.

The scale of the planned measurements is what sets this effort apart. Halfmann’s team will measure the aggregation tendencies of 50,000 proteins, each expressed individually in yeast cells under diverse conditions designed to mimic the cellular mishaps that accumulate as humans age. Across more than one million samples, the team expects to produce more than 10 billion individual measurements of protein aggregation. Direct measurements of protein aggregation at cellular resolution have not previously been done at anything approaching this scale, according to Halfmann, who noted that his laboratory has been laying the foundation for this moment since he joined Stowers in 2015.

That foundation includes a landmark achievement. In 2023, Halfmann’s lab became the first to experimentally determine the structure of the initiating step in amyloid formation associated with Huntington’s disease, work that involved the polyglutamine proteins implicated in that disorder. If scientists can figure out exactly how aggregation starts, Halfmann has argued, they can potentially stop the forest fire before it ignites. His lab has also studied TDP-43, a protein strongly associated with ALS and frontotemporal lobar degeneration. Those earlier studies, which examined hundreds of protein sequences, served as pilots for the new undertaking, which will scale the approach to 50,000 sequences.

The initial focus of NATIVE-ID is frontotemporal lobar degeneration, or FTLD, a neurodegenerative disease that shares significant genetic and biological features with ALS. The larger ambition, however, is to develop approaches that generalize across other diseases involving protein misfolding. The team’s other members bring complementary capabilities: generating human neurons from different genetic backgrounds, measuring protein interactions inside those neurons, developing advanced deep-learning frameworks, determining protein structures in test tubes, and running large-scale computational simulations of protein behavior. In addition to Halfmann, the collaboration includes scientists from UC Berkeley, Brown University, Emory University, Johns Hopkins University, Parallel Squared Technology Institute and Texas A&M University, with the Innovative Genomics Institute effort led by Executive Director Brad Ringeisen and Hanqin Li, head of the institute’s Advanced Translational Genetics lab.

Leadership at Stowers emphasizes that the experimental and computational halves of the project are inseparable. Alejandro Sánchez Alvarado, Ph.D., President and Chief Scientific Officer of the institute, noted that a deep learning model capable of decoding the language of disordered proteins needs rigorously obtained data at scale, something that has yet to exist. Halfmann’s laboratory, he said, will provide what has been missing: more than 10 billion measurements of aggregation, taken one living cell at a time, across 50,000 proteins, while colleagues at the Innovative Genomics Institute supply the human neurons in which the model’s predictions must ultimately hold true. Neither half of the project, he argued, can succeed without the other.

The award funds an initial two-year phase of work, structured within BIOGAMI as the first of two 24-month phases. Phase 1 is dedicated to delivering foundational datasets and models; Phase 2 will focus on validating potential therapeutics and on new ways to detect protein dysfunction earlier in the course of disease. For Halfmann, the timing is critical. Artificial intelligence now stands a real chance of decoding the language of intrinsically disordered proteins, he said, but only if it is provided sufficient high-quality data, and DAmFRET has finally matured enough to generate that data. If the collaboration succeeds, the earliest molecular whispers of Alzheimer’s, ALS and related diseases may one day be readable, and interceptable, years before the first symptom appears.

Subject of Research: AI prediction of intrinsically disordered protein aggregation in neurodegenerative disease

Article Title: Stowers scientist selected for $28.6 million effort to predict protein changes behind neurodegenerative disease

Article References: Stowers scientist selected for $28.6 million effort to predict protein changes behind neurodegenerative disease. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: protein aggregation, intrinsically disordered proteins, neurodegenerative disease, artificial intelligence, ARPA-H, DAmFRET, ALS, Huntington's disease, frontotemporal lobar degeneration, deep learning, Stowers Institute, protein misfolding

The proteins at the center of the study belong to a family first identified by Stowers Investigator and Howard Hughes Medical Institute Investigator David Stern, Ph.D., in 2021. He named them BICYCLE proteins because of a repeating cysteine motif that runs through their sequences. These molecules are secreted by aphids into the plants they attack, where they help reprogram plant tissue into galls: living structures made of plant cells but built to an insect’s specifications, providing both shelter and food for the aphid’s offspring. Hundreds of these proteins can be deployed in a single interaction, making them a defining weapon in the ongoing conflict between aphids and their hosts.

From the moment they were discovered, BICYCLE proteins posed a puzzle. Biologists typically identify an unknown protein by comparing its amino acid sequence against the millions catalogued in public databases, since a close match to a characterized protein offers strong clues about function. When BICYCLE proteins were run through those comparisons, the results came back empty. The family evolves so quickly, generation after generation, that their sequences have been rewritten past the point where computational tools can recognize any family resemblance. “We could tell immediately that they didn’t look like any other proteins that you might find in a database,” Stern said. That extraordinary rate of change is itself a signature of conflict, because proteins on the front line between a parasite and its host tend to evolve fastest, each side under relentless pressure to counter the other’s latest move.

To break through the wall of sequence divergence, the team turned to experimental structural biology. In collaboration with the lab of Angela Gronenborn at the University of Pittsburgh, researchers spent years crystallizing two BICYCLE proteins and solving their three-dimensional structures by X-ray diffraction. The structures matched nothing on record, but they revealed a version of a structural motif seen elsewhere in biology, known as a saposin-like fold. By an extraordinary coincidence, that same week a new deep learning system called AlphaFold2 was released, promising scientists highly accurate structure predictions from sequence alone. Stern immediately fed the BICYCLE sequences into the new tool, and it failed.

AlphaFold2 did not simply deliver low-confidence predictions; it returned the wrong answer. That failure became the study’s most productive result, because it exposed how the AI actually operates. AlphaFold2’s power does not come from machine learning alone. The program leans heavily on evolutionary data, drawing on multiple sequence alignments of related proteins to infer which parts of a structure are constrained and which are flexible. For BICYCLE proteins, that evolutionary record simply did not exist in the databases the program consults. “When it searched the database for similar proteins, it couldn’t find any,” Stern explained. “That part of the AlphaFold2 program was empty. It was actually an empty box.”

The fix, remarkably, required leaving the computer entirely. The team collected aphids across Virginia and West Virginia, traveled to Japan to obtain one critical species, and sequenced the genomes of closely related aphid species. By supplying AlphaFold2 with this missing evolutionary context, the researchers effectively reconstructed the input the AI needed. The result was dramatic: the program’s predictions now reproduced the crystal structures the team had solved experimentally. “Lo and behold, AlphaFold2 gave us back the crystal structure that we had solved,” Stern said. The lesson is a striking one for the era of AI-driven biology. Deep learning can achieve extraordinary things, but only when evolution itself provides the raw material. Without the information encoded in related sequences, even the most powerful model is working with an empty box.

With the method validated, the team scaled it up. They generated roughly 2,400 high-confidence structure predictions spanning seven aphid species, and a clear pattern emerged. The same saposin-like fold kept reappearing, duplicated, reoriented, and rearranged in different combinations from one protein to the next. In other words, across thousands of proteins whose sequences are too divergent to be recognized as related, a single shared architectural plan persists. Evolution has remodeled that plan so extensively that sequence-based tools fail, yet the underlying blueprint remains unmistakable once structures are compared.

What the researchers expected to find next, however, eluded them. They anticipated some shared surface feature among the BICYCLE proteins, some conserved patch of chemistry that would hint at a common job inside the plant cell. “We didn’t find anything,” Stern said. “No conserved patches of positive charge or negative charge, or regions that hated water, or regions that really liked water. Nothing.” The proteins clustered neatly by structural similarity but spread into a near-continuous spectrum when their surface chemistries were compared. “We used every method we could think of to force them into different clusters, and we just couldn’t,” Stern said. “That suggests that these proteins are really exploiting many different mechanisms in the plant to take over the plant cell.”

The authors propose that BICYCLE proteins are evolving on two fronts simultaneously: reaching new molecular targets inside the plant while remaining unrecognizable to the plant’s immune surveillance system. This dual pressure would explain both the conserved fold, which may be essential for the proteins’ basic structural integrity, and the astonishing surface diversity, which allows aphids to strike many different plant targets while evading detection. The result is a highly adaptive molecular arsenal, forged over evolutionary time by an arms race between insect and plant, in which each new plant defense selects for aphid effectors that are both functionally versatile and immunologically invisible.

Beyond aphids, the study delivers a methodology that other laboratories can immediately apply. Thousands of rapidly evolving proteins across biology, including many involved in immunity, host-parasite interactions, and agricultural traits, have remained largely inaccessible to structure prediction because their sequences lack recognizable relatives in public databases. The Stowers team’s approach, pairing close genomic sampling of related species with AlphaFold2, offers a general solution: generate the missing evolutionary context experimentally, then let the AI do the rest. For Stern, the work is only a beginning. “This is really the beginning of our work on the BICYCLE proteins, the beginning of our work on how insects control plants,” he said. “I’m very excited about the future and what we’re going to be able to do here to really understand how these molecules control plant biology.” What began as a failure of artificial intelligence has ended as a demonstration that evolution, properly supplied, can teach the machine what it could not learn alone.

Subject of Research: Structural evolution of rapidly evolving aphid BICYCLE effector proteins and AI-based structure prediction

Article Title: Stowers scientists uncover a hidden blueprint in aphids, providing a way to predict what AI couldn't solve alone

Article References: Stowers scientists uncover a hidden blueprint in aphids, providing a way to predict what AI couldn't solve alone. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: aphids, BICYCLE proteins, AlphaFold2, protein structure prediction, saposin-like fold, evolutionary biology, plant-insect interactions, host-parasite conflict, X-ray crystallography, gall formation, agriculture, Stowers Institute