MiniFold: Simple, Fast and Accurate Protein Structure Prediction
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: protein, structure prediction, efficiency, hardware-optimization
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Protein structure prediction has emerged as a powerful tool for biologists and drug makers. However, the computational toll associated with state-of-the-art models, such as AlphaFold2 or ESMFold, hinders their use in large-scale applications like virtual screening or mutational scanning, where a single experiment may involve processing millions of protein sequences. In an effort to develop a more efficient model, we aimed to understand which of the complex architectural choices proposed in AlphaFold2 were essential to achieve high performance, and which could be omitted without significantly compromising accuracy. This analysis culminated in a simple, yet highly expressive architecture for protein structure prediction. Our model, MiniFold, consists of a minimal Evoformer variant, a parameter-free coordinate recovery algorithm, and a custom hardware-optimized implementation composed of newly designed GPU kernels. When compared against ESMFold, MiniFold achieves over 100x speedup and shows improved scalability to long protein sequences while conserving over 95% of the original performance, making it a promising candidate for large-scale applications.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8942
Loading