An interesting way to look for better biological foundation models.
The core bet is simple: biological data is expensive. Text and image models often improve by consuming more data, but protein structure data is far more constrained, far harder to generate, and far more sensitive to leakage. If we want better biological foundation models, we need architectures and training methods that make stronger use of the data we already have.
Full details are on GitHub https://github.com/ChrisHayduk/nanoFold-Competition