ESMFold, Meta's Rival To AlphaFold, Gets New Public Releases
ESMFold, Meta AI's horse in the protein folding race, gets a new public model set release that gets comparable results at significantly faster speeds.
Created on August 22|Last edited on August 22
Comment
Researchers on the Protein Team at Meta AI have worked for several years to develop a lineup of models for protein structure prediction called Evolutionary Scale Modeling, or ESM for short. The project has released a number of models in the past, but the newest and most impressive iteration has just got it's public release.
Details on this model's creation and the research that went into it is available in this research paper released late last month.
What is ESMFold?
ESMFold is a protein structure prediction model which relies on transformer models to understand and encode protein sequences. Unlike many other models, it does not rely on multiple sequence analysis, greatly increasing speed.
This model incorporates new transformer models developed alongside it, ranging from 8 million parameters to 15 billion parameters. This lineup of models is called ESM2, and prove significantly more capable than their predecessors.
While other models like AlphaFold2 use multiple sequence analysis (MSA), a process which incorporates an external database of protein sequences deemed related to the input sequence, to increase it's prediction effectiveness, ESMFold does not. Using only the single input sequence, ESMFold greatly speeds up inference time by only looking at the single protein sequence.
ESMFold's comparatively small model size and analysis method which does not include MSA, it proves itself significantly faster than the larger models which do use MSA. Despite this focus on speed and efficiency, ESMFold maintains high quality predictions. This small architecture, plus the fact it's open source, means researchers can easily use it for their own projects without breaking the bank on expensive hardware or compute.
But it doesn't stop there - researchers can also choose to use the ESM project's own MSA model in conjunction with ESMFold if they wish to, increasing the accuracy of predictions.
Get access to ESMFold
Everything easy available at the GitHub repository, where you can find pre-trained models and the datasets used to train them, code notebooks, and more.
Find out more
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.