Known as the Critical Assessment of protein Structure Prediction (CASP), the 3-decade-old competition has run out of funding from the U.S. National Institutes of Health (NIH) and will exhaust emergency support from the University of California (UC) Davis, which oversees the grant, on 8 August. UC Davis has told the two researchers who run the program that their jobs will end in weeks.

NIH officials have offered no reassurance about the program’s future, despite repeated inquiries from CASP organizers, who submitted a request to renew their $800,000 grant last year. The agency did not respond to multiple requests for comment. John Moult, a CASP co-founder at the University of Maryland, says contest organizers are “scrambling” to find alternative funding from foundations and other countries

CASP nears end

Known as the Critical Assessment of protein Structure Prediction (CASP), the 3-decade-old competition has run out of funding from the U.S. National Institutes of Health (NIH) and will exhaust emergency support from the University of California (UC) Davis, which oversees the grant, on 8 August. UC Davis has told the two researchers who run the program that their jobs will end in weeks.

NIH officials have offered no reassurance about the program’s future, despite repeated inquiries from CASP organizers, who submitted a request to renew their $800,000 grant last year. The agency did not respond to multiple requests for comment. John Moult, a CASP co-founder at the University of Maryland, says contest organizers are “scrambling” to find alternative funding from foundations and other countries

We know who we can thank for this...

CASP co-founder John Moult told STAT that Google-owned DeepMind will make a one-time gift to support the Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction, better known as CASP. No amount was disclosed, but the gift will support CASP for approximately 12 months.

CASP17 will be held in May-August 2026.

Details of the experiment will be posted here soon.

Details of the experiment will be posted here soon.

Recent CASPs saw enormous jumps in the accuracy of computed structures, first in CASP14 (2020) for single proteins and domains, with many models competitive in accuracy with experiment, and second, in CASP15 (2022), with a large increase in the accuracy of protein complexes. These advances are primarily the result of the successful application of deep learning methods, particularly AlphaFold2 and other methods built around it. But results from CASP16 (2024) suggest a performance plateau.

The primary goal for the 2026 CASP17 experiment is to catalyze breakthroughs in areas where deep learning has yet to deliver and particularly where success has major practical implications. CASP17 modeling categories are chosen accordingly.


Modeling categories

Immune Complexes: This is an area of major practical importance, and also one that in CASP is a major failure area for current deep learning methods. However, two promising approaches have been seen in recent CASPs. To nurture progress, we are introducing immune complexes as a specific category in CASP17, aiming to provide a rich and varied set of non-homologous targets spanning antibody-antigen, nanobody-antigen complexes, and T-cell receptor complexes.

Organic Ligand-Protein Complexes: These complexes have obvious importance for the development of new small-molecule drugs. The most recent CASP16 (2024) showed that current deep learning results often fall short of the experimental accuracy, especially where there is little homologous or analogous data available. Thus, we will again include this as a specific target area in CASP17, aiming to obtain targets representing a wide range of realistic conditions.

Nucleic Acids and Complexes: Despite claims that deep learning methods have solved the problem of computing nucleic acid structures, CASP16 and a subsequent Kaggle challenge showed that these methods are usually no better than classical approaches and that both fail badly in the absence of homologous structural information. But new deep learning methods are now appearing at a fast pace. To assess these, CASP17 will include a diversity of non-homologous RNA/DNA structures and protein-nucleic acid complexes. As before, this category will be in collaboration with RNA Puzzles, and will be co-ordinated with RNA Kaggle challenges. Because of the speed with which some experimental structures may be obtained we plan to shorten the prediction window for this kind of target.

Conformational Ensembles: Testing methods for computing ensembles of structures is a major expansion area for CASP. CASP17 will again include two main types of ensemble target: First, those where there are two or a few discrete conformations, each with a well-determined experimental structure, that can be assessed in the conventional way. Second, targets where there are multiple lower resolution experimental datasets available, such as cryo-tomography; SAXS; NMR (RDC, chemical shifts and other data); FRET; and cross-linking. For these, assessment will be based on agreement between the low-resolution experimental data and corresponding values calculated from submitted structural ensembles.

Difficult Protein Structures and Complexes: Recent CASPs have shown that in many cases, current deep learning methods deliver high-accuracy structures for single proteins and complexes. But there are critical weaknesses. To help address these, CASP17 will focus on performance in the following areas:
Membrane proteins.
Proteins and complexes with weak evolutionary information such as those with viral or parasite origin, "shallow" sequence alignments and recently evolved interfaces.
Large proteins and complexes with complicated stoichiometry or arrangement of subunits (>1,000 amino acids).
For protein complexes we will again work in close collaboration with our CAPRI partners.

Note: CASP17 protein targets will be initially released without stoichiometry information. All multimeric targets will be re-released in a second modeling stage, with the experimental stoichiometry data provided. For this second stage, the MassiveFold team will also provide large sets of AF-based models after the server deadline (typically 3 days after the second-stage target release).

Accuracy Estimation: A structural model that is not accompanied by reliable and detailed accuracy methods is of very limited value. Estimates for protein structures were a success story in CASP even before the introduction of deep learning methods. Now CASP results have shown that accuracy estimates provided by model builders are consistently reliable. For third-party methods (those that estimate accuracy for models produced by others) we will no longer include single proteins, but will include protein complex interfaces from the MassiveFold and CASP stage-2 models. New this CASP is self-assessment of accuracy for nucleic acid structures and complexes, and for protein-ligand complexes.