PSBench: 1.4 million expert-verified protein models to build trust in AI predictions
University of Missouri researchers have released PSBench - the largest collection of protein structure models with quality assessment to date. It includes 1.4 million annotated models, each verified by independent experts. The goal: give scientists a reliable benchmark to train and evaluate AI systems that judge the quality of predicted protein structures, a critical step for drug discovery in areas like Alzheimer's and cancer.
Proteins do the work. Their 3D shapes determine how they function, and small errors can derail experiments. AI methods such as AlphaFold have pushed accuracy forward, but no single tool is consistently right across every protein family or fold. PSBench gives you the ground truth to know when to trust a model - and when to keep looking.
"With PSBench, scientists can develop AI methods to assess the quality of predicted protein models and decide if they can be trusted," said Jianlin "Jack" Cheng, Curators' Distinguished Professor and Paul K. and Diane Shumaker Professor in Bioinformatics. "Our work represents a significant step toward applying protein models to understanding diseases and developing new treatments."
Why PSBench matters for your research
- Trust at scale: Access a massive, expert-validated set of protein models to benchmark and calibrate your QA pipelines.
- Model selection you can defend: Choose predicted structures for downstream experiments with evidence, not guesswork.
- Better generalization: Compare methods across diverse protein types and difficulty levels, reducing overfitting to any single dataset.
- Faster iteration: Shorten the loop between in silico screening and wet-lab validation by filtering poor models early.
What's inside (high level)
- Scale: 1.4 million protein structure models with quality annotations.
- Independent verification: All entries are checked by external experts for reliable assessment signals.
- Community backbone: Built using in-house resources and community data shaped by CASP - the biennial, international benchmark for protein structure prediction.
Practical ways to use PSBench now
- Train QA models: Build regression/classification models that score predicted structures and flag unreliable regions.
- Benchmark fairly: Compare QA methods across standardized, expert-vetted data to identify real gains vs. noise.
- Triage predictions: Prioritize structures for MD refinement, docking, or wet-lab validation based on assessed quality.
- Ensemble strategies: Blend multiple QA signals and calibrate them against PSBench to improve decision thresholds.
- Method transfer: Test how your QA approach holds up across folds, domains, and protein sizes before deploying in production workflows.
How we got here
For decades, protein folding was a stubborn problem. At the 2012 CASP competition, Cheng's group was the first to show deep learning could move the needle. That early push paved the way for today's top predictors and set the stage for stronger QA systems.
The PSBench study was presented at NeurIPS 2025 in San Diego - a fitting venue given its history of debuting foundational AI methods. As Cheng put it, "With PSBench, Mizzou is contributing a powerful new tool to the global scientific community. We're helping lead the next chapter in AI-driven biomedical discovery."
Where this helps most
- Target discovery: Filter structure candidates before committing assay time.
- Structure-guided design: Use higher-confidence regions for pocket analysis and ligand placement.
- Cryo-EM/X-ray workflows: Decide when predicted models are good enough to guide map interpretation or model building.
- Cross-lab reproducibility: Standardize QA criteria across teams using a shared benchmark.
Keep building
Whether you're advancing structure prediction, QA modeling, or downstream design, PSBench gives you a wide, stable foundation to test ideas and ship tools that biologists can trust.
If you're formalizing your AI skill stack for research, explore the AI Learning Path for Research Scientists.
Developed at Mizzou's College of Engineering.
Your membership also unlocks:
