Latest Machine Learning Research From Baidu Proposes A Unique Approach That Speeds Up The Training For AlphaFold2 By 38.67%

A DeepMind AI system called AlphaFold predicts the three-dimensional structure of proteins based on their amino acid composition. Since they were made available to the public for users to explore and research their protein of interest, the AlphaFold software and “AlphaFold Protein Structure Database” have been around for a year.

The AlphaFold2 protein structure prediction model has transformed structural biology research by offering quicker and more effective approaches to identifying links between protein structure and function. However, due to its complicated structure, the AlphaFold2 model’s wider use in biocomputing and life science research is constrained by the training process’s high memory and time requirements.

Without compromising performance, the unique approach increases AlphaFold2 training speed by up to 38.67 percent.

The research team upgrades AlphaFold2’s Evoformer to a Parallel Evoformer, eliminating the MSA(multiple sequence alignment) and pair representation’s computational reliance. 

Experiments demonstrate that this does not reduce accuracy.

To increase training efficiency, they suggest Branch Parallelism for Parallel Evoformer, which divides various computing branches over several devices in parallel. This eliminates the official AlphaFold2 implementation’s restriction on data parallelism.

The training performance of AlphaFold2 is enhanced by 38.67% and 36.93%, respectively, as they shorten the end-to-end training period to 4.18 days on UniFold and 4.88 days on HelixFold. AlphaFold2 training is effective, saving money on R&D expenses for biocomputing research.

AlphaFold2 is a protein estimate system that can directly display the 3D coordinates of every atom in a given protein. While AlphaFold2 has reached unparalleled accuracy in protein structure prediction, it is incredibly time- and compute-intensive, taking 11 days to train a model from the start on 128 TPUv3 cores.

The group suggests the following two optimization methods to boost AlphaFold2 training effectiveness: 

•  A Parallel Evoformer that converts the existing Evoformer block’s two serial computing branches into a parallel computing structure; and 
•  A Branch Parallelism (BP) technique for the Parallel Evoformer that accelerates computation by scaling to new devices using data parallelism.

On two AlphaFold2 models implemented in deep learning frameworks—UniFold in PyTorch and HelixFold in PaddlePaddle—the team conducted extensive tests to assess their methodology. According to the findings, the Parallel Evoformer and Branch Parallelism method can reduce training time on both systems to under five days while increasing training effectiveness by 38.67% in UniFold and 36.93% in HelixFold.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Efficient AlphaFold2 Training using Parallel Evoformer and Branch Parallelism'. All Credit For This Research Goes To Researchers on This Project. Check out the paper and code.

Please Don't Forget To Join Our ML Subreddit