University of Missouri releases PSBench protein model benchmark for AI-driven drug development

University of Missouri researchers have released PSBench, a database of 1.4 million annotated protein structure models, a new resource that could accelerate drug development for diseases such as Alzheimer’s and cancer. The models were verified by independent experts and are intended to help scientists build more accurate artificial intelligence systems for assessing the quality of protein structure models. Recent advances in AI, including tools such as AlphaFold, have dramatically improved protein structure prediction, but no single AI tool is consistently accurate for every type of protein.

The database gives scientists information to assess whether predicted protein models can be trusted, which is critical for developing future medical treatments. Proteins drive every biological process in the human body, and their three-dimensional shapes determine how they function. Even small structural changes can lead to disease.

The PSBench resource was built leveraging in-house and community-wide resources generated in the Critical Assessment of protein Structure Prediction (CASP), widely recognized as the international gold standard for assessing computational methods for protein prediction. The biennial competition was created to independently test computer models that predict how protein chains fold into the 3D shapes they need to function.

These advances reflect a broader shift in biologic drug discovery from a slow, experimental process into a data-driven engineering discipline. Advances in deep learning, from protein language models that interpret sequence grammar to structure predictors and next-generation generative models, are enabling researchers to decode, predict, and create complex biologic molecules with unprecedented precision. By learning from vast biological datasets, these systems can uncover patterns linking sequence, structure, and function, shifting the field away from serendipitous discovery toward rational, design-led innovation.

AI can design new proteins, antibodies, peptides, and nucleic acids with tailored functions while optimizing properties such as binding affinity, stability, and manufacturability. Generative approaches, including diffusion models and autoregressive architectures, allow scientists to explore vast regions of molecular design space that would be inaccessible through conventional screening or directed evolution. In parallel, machine learning is improving delivery strategies such as lipid nanoparticles, viral vectors, and antibody-drug conjugates by predicting performance, guiding formulation, and proposing novel components.

Early AI-designed biologics, including peptide therapeutics, antibodies, and mRNA-based candidates, are entering clinical evaluation. At the same time, AI-guided optimization is accelerating affinity maturation and stability engineering, reducing the need for extensive experimental screening and shortening development timelines.

Current models often excel at predicting molecular structure but struggle to capture the complexity of biological systems, leaving a gap between in silico predictions and in vivo outcomes. Immunogenicity, pharmacokinetics, and cellular context remain difficult to model accurately, and progress is constrained by limited access to high-quality, task-specific datasets and the difficulty of optimizing multiple drug properties simultaneously without trade-offs. The authors of a recent review said tighter integration between computation and experiment, particularly through closed-loop, AI-driven workflows in which automated experiments continuously generate data to refine models, could help bridge the gap between prediction and performance.