PSBench at the University of Missouri: A New Trust Layer for AI-Driven Protein Discovery

Artificial intelligence has solved one of biology’s most stubborn mysteries: how proteins fold into their intricate three-dimensional shapes. But as the field shifts from prediction to application, a new question is becoming more urgent than ever:

When can we trust the model?

Researchers at the University of Missouri believe they’ve taken a major step toward answering that question. The university has announced the release of PSBench, a massive new benchmark dataset containing 1.4 million protein structure models with annotated quality assessments. Led by Jianlin ‘Jack’ Cheng, a Curators’ Distinguished Professor in Bioinformatics, the project is designed not to generate new structures—but to evaluate them.

That distinction could prove critical for the future of AI-driven medicine.

The New Bottleneck in Protein AI

The protein-folding problem stood unsolved for more than half a century. That changed dramatically when AlphaFold from Google DeepMind demonstrated near-experimental accuracy in predicting many protein structures. The breakthrough was so transformative that AI-powered protein prediction was recognized with a share of the 2024 Nobel Prize in Chemistry.

Since then, prediction systems have expanded beyond single proteins to complexes, interfaces, and biomolecular interactions. The AlphaFold Protein Structure Database now contains hundreds of millions of predicted structures, turning what was once scarce into something almost abundant.

But abundance introduces a new challenge.

A predicted protein model can look convincing, even elegant. Yet subtle errors—especially at binding interfaces or flexible regions—can make the difference between a viable drug target and a costly dead end. Internal confidence metrics such as pLDDT and predicted aligned error provide helpful guidance, but they remain model-generated signals. They estimate uncertainty from within.

PSBench approaches the problem from the outside.

What Makes PSBench Different

Rather than building another predictive engine, PSBench functions as a large-scale evaluation platform. The database compiles 1.4 million structural models drawn from community-wide efforts such as the Critical Assessment of protein Structure Prediction (CASP), the long-running gold standard for blind protein modeling experiments. These models are paired with accuracy labels that allow researchers to train and test independent AI systems capable of estimating structural reliability.

In essence, PSBench enables AI models that score other AI models.

That capability is becoming increasingly important as the field shifts from asking “Can we predict a structure?” to asking “Is this structure reliable enough to guide experiments?”

Cheng’s team has deep roots in this evolution. Back in 2012, during an earlier CASP competition, his group was among the first to demonstrate that deep learning could meaningfully improve protein structure modeling. Over a decade later, PSBench reflects the next phase of that journey: refining how predictions are judged, not just generated.

The work was recently presented at NeurIPS 2025, underscoring how tightly machine learning research and structural biology are now intertwined.

AlphaFold in 2026: From Folding to Interactions

Meanwhile, the broader ecosystem continues to advance. The latest generation of AlphaFold models extends beyond folding individual chains to modeling interactions between proteins, DNA, RNA, and small molecules. Databases have grown to unprecedented scale, and community contributions are accelerating coverage across microbial, viral, and human proteomes.

As these tools mature, researchers increasingly treat predicted structures as starting points for hypothesis generation. Experimental validation still matters deeply, but AI now sets the agenda for what gets tested first.

That is precisely why quality assessment matters so much.

If predictive AI systems are generating more structural hypotheses than laboratories can possibly validate, then the ability to triage those hypotheses—accurately and objectively—becomes foundational infrastructure.

Implications for Drug Discovery

Proteins are the functional engines of biology. Their three-dimensional shapes determine how they interact, signal, and regulate life’s processes. When structures are misinterpreted, especially in therapeutic contexts, the consequences can cascade through years of development.

By improving the training and benchmarking of model quality assessment systems, PSBench could help reduce false confidence in flawed predictions. More reliable structural scoring means better prioritization of targets, more efficient use of laboratory resources, and potentially faster paths to therapies for complex diseases such as Alzheimer’s and cancer.

Importantly, PSBench does not replace predictive tools like AlphaFold. Instead, it complements them—adding a trust layer to an ecosystem that is rapidly expanding in power and scale.

The Rise of the Scientific Trust Layer

AI in biology has entered a new phase. The first era was about solving prediction. The second was about scaling access. The emerging third era is about validation, benchmarking, and governance.

PSBench represents that shift.

As AI systems become central to biomedical discovery, the ability to evaluate their outputs with rigor will determine how confidently researchers can build upon them. In a domain where angstrom-level precision can influence billion-dollar decisions, trust is not optional.

If AlphaFold helped unlock the structure of life at scale, PSBench may help ensure that what we unlock is solid enough to stand on.