\n
Overview
\n
Welcome to the Stanford Ribonanza RNA Folding Challenge! In this competition, your goal is to predict the chemical reactivity at each nucleotide position within an RNA molecule. These reactivity signals are highly dependent on the RNA's three-dimensional structure. A model that can accurately forecast reactivity must therefore capture complex structural features—effectively \"understanding\" RNA folding patterns.
\n\n
Such predictive models hold significant potential in biotechnology. They could pave the way for precise RNA structure design, contributing to breakthroughs in programmable therapeutics. This includes progress toward treatments for diseases like Alzheimer's and pancreatic cancer, as well as the development of next-generation antibiotics and climate-related biotechnological tools.
\n\n
This notebook introduces a basic starting point that you can build upon. Potential directions for improving performance include:
\n\n
\n - Designing a more appropriate loss function that accounts for samples with
SN_filter = 0 and penalizes incorrect reactivity predictions. \n - Enhancing the model architecture and incorporating additional data sources, such as the
Ribonanza_bpp_files. \n - Considering sequence length distribution differences between the public and private leaderboard sets. While the public set includes shorter sequences (115–206 bases), the private set focuses on longer ones (207–457 bases). Ensuring that your model performs well across this full range will help avoid leaderboard surprises and improve generalization.
\n
\n