A new stage of stanford The RNA 3D Folding Challenge, currently live on Kaggle, invites researchers and machine learning teams to develop models that can predict the three-dimensional structure of RNA molecules using only sequence. of competitionn builds on the first phase, which was a milestone for the field, and raises the bar with more complex goals and more rigorous evaluations.
This challenge lies at the intersection of AI research, life sciences, and skills development. Despite its importance in understanding disease, therapeutic development, and basic biology, RNA structure prediction is far less mature than protein folding.
RNA plays a central role in how cells function, but it remains difficult to predict how RNA folds into functional three-dimensional structures. Unlike protein structure prediction, where AI systems have made great strides in recent years, RNA modeling has been limited by sparse data and the inherent complexity of RNA folding.
The first Stanford RNA 3D Folding Challenge demonstrated for the first time that fully automated machine learning models can compete with human experts. In Part 2, we build on our results by introducing more difficult targets, such as RNA molecules for which there are no structural templates available, and applying a revised evaluation framework designed to reward higher precision.
The competition is organized through a global collaboration involving experimental RNA structural biologists, Stanford University School of Medicine, and the Howard Hughes Medical Institute’s AI@HHMI initiative. This is timed to unveil new approaches ahead of the 17th Structural Prediction Critical Assessment scheduled for April 2025.
Kaggle hosts research-grade contests and assessments
This challenge is hosted on Kaggle, a platform widely used in the machine learning community for research competitions, benchmarking, and skill development. Kaggle provides the infrastructure for codebase submissions, leaderboards, and ratings, positioning contests as both research exercises and public testing grounds for new modeling approaches.
Participants must submit a notebook that generates five predicted structures for each RNA sequence in the test set. Submissions are scored using the TM score, a standard structural similarity metric that compares predicted structures to experimentally determined reference structures. To avoid shortcuts, only correctly aligned residues are rewarded by numbering, and scores are averaged across the best of the five predictions for each target.
Organizers say this scoring approach aims to bring the model closer to exact structural accuracy, rather than partial alignment or template reuse.
Stake includes prizes and scientific copyrights
The contest will run on a fixed timeline, with entries accepted until March 18, 2026, with final submissions due by March 25. Private leaderboards will be released the following week after additional evaluation.
The total prize pool is $75,000, with the top team receiving $50,000. In addition to prize money, top-performing participants will be invited to submit their code and model descriptions to a peer-reviewed scientific paper summarizing the results of the competition. This is a great incentive for academic and research-focused teams.
The return of the Stanford RNA 3D Folding Challenge highlights the role of public competitions in advancing difficult scientific problems while developing applied AI skills. Unlike benchmark-only tasks, this challenge emphasizes end-to-end modeling, reproducibility, and consistency with experimental data.
