echemi logo
  • Product
  • Supplier
  • Inquiry
    Home > Active Ingredient News > Antitumor Therapy > AlphaFold has met a strong competitor, and Facebook's Meta AI is more efficient and intelligent

    AlphaFold has met a strong competitor, and Facebook's Meta AI is more efficient and intelligent

    • Last Update: 2022-11-14
    • Source: Internet
    • Author: User
    Search more information of high quality chemicals, good prices and reliable suppliers, visit


    "We're going to see protein structure prediction become more efficient, simpler, and cheaper, which will open the door
    to new things.

    — Burkhard Rost, a computational biologist at the Technical University of Munich in Germany, who compared
    Meta AI and AlphaFold.

    What is Meta AI? It was developed by Meta and, like AlphaFold, is a system
    that uses artificial intelligence (AI) to predict protein structure.
    The predecessor of Meta is the Facebook we are familiar with

    Today, Meta AI is moving into the field
    of AI-predicted proteins with the 600 million protein structures it predicts.

    What will Meta's foray into the field of protein prediction bring? Compared to AlphaFold, what are the advantages and disadvantages of the new Meta AI?

    Two or three months ago, do you remember AlphaFold, which once again caused a sensation in the scientific community? The company that built it, DeepMind, has said that AlphaFold has successfully predicted the structure
    of more than 200 million proteins on Earth.

    In fact, since AI entered the field of protein structure prediction, more and more companies have invested in it, and systems that use AI to predict protein structure have emerged in an endless

    Today, our protagonist is the most recent one – Meta AI, which, surprisingly, was designed by Facebook


    Meta AI: "autocomplete" protein structure

    What is Meta AI that Rost, a computational biologist at the Technical University of Munich in Germany, praised as "more efficient, simpler and cheaper"?

    In July 2022, the news that AlphaFold predicted about 220 million protein structures still seems to haunt
    us today.

    AlphaFold predicts protein structure

    It covers almost all the proteins
    of known organisms in the DNA database.
    Now, another tech giant, Meta, is filling in an unknown part
    of the protein universe.

    On October 31, 2022, the Meta team published a paper titled "Evolutionary-scale prediction of atomic level protein structure with a language model" on the biorxiv preprint platform, giving a detailed introduction
    to Meta AI.

    Related papers posted in preprints, screenshot from biorxiv.

    According to the paper, Meta AI predicted about 600 million protein structures from bacteria, viruses and other microorganisms
    that have not yet been characterized.

    Yes, you heard that right, it's
    600 million.

    In order to predict such large-scale protein structure data, one of the first important problems the research team needs to overcome is to break through the speed limit

    To this end, they trained a "large language model" (LLM), which can train a huge number of models with very large numbers of parameters through large-scale data and predict accurate protein structures
    directly from protein sequences.

    How to understand it, like giving it a few letters or words, it can predict
    text paragraphs through calculation.

    Previously, LLM was often used for text prediction, but Meta researchers used it for protein structure prediction and taught it to "autocomplete"
    proteins when amino acid ratios are not clear.

    As a result, Meta AI's predictions are vastly faster — "not only 60 times faster than current state-of-the-art technology, but also accurate
    ," the Meta team says.

    They named the algorithmic model using LLM ESMFold, applied it to a metagenomic DNA database, and rolled out the results in the form of an "ESM metagenomic map" containing more than 600 million protein structures

    Metagenomic database, screenshot from genome.

    These structures not only provide a new perspective on the breadth and diversity of nature, but will also accelerate protein discovery
    in areas such as medicine, green chemistry, environmental applications, and renewable energy.

    Alexander Rives, head of research at the Meta AI team, said: "This is a structure that we did not know before, and these mysterious proteins have made it possible
    for us to explore biology in depth.
    " ”

    So, what kind of charm does Meta AI have? What are its advantages and disadvantages compared to AlphaFold?


    617 million protein structures, 2 weeks to complete

    The Meta team bluntly states that although it is not as accurate as AlphaFold, its advanced algorithms allow it to predict structures much faster than AlphaFold

    What concept? AlphaFold took about a few minutes to generate a prediction during the prediction process, while Meta AI predicted 617 million proteins, which took only 2 weeks

    ——Meta AI can predict more than 30,000 protein structures per minute!

    Although the predictions are not accurate enough, overall, the results of this algorithmic model out of these 617 million predictions show that more than one-third of the predictions are highly accurate, and in some cases, their recognition accuracy can reach the atomic level

    In addition, the Meta team is publishing a rapid protein folding model for creating a database, as well as an API program interface

    The model and interface will be made available to all researchers to help them identify previously unstudied protein structures, explore human evolution, and develop new proteins
    that can be used in medicine and other fields.

    The Meta team said, "With 15 billion parameters, our new language model is the largest protein language model
    to date.

    Of course, in addition to Meta AI and AlphaFold, there are many companies that are also investing in the field of AI+ protein prediction today.


    Endless AI+ protein prediction systems

    Companies and teams researching protein prediction at home and abroad are emerging
    In addition to the Meta AI and AlphaFold we talked about above, there are also RoseTTAFold, Hermite, etc

    Designed by David Baker et al.
    at the University of Washington's Institute for Protein Design, RoseTTAFold is based on a "three-track" neural network model capable of simultaneously dealing with patterns in protein sequences, how a protein's amino acids interact, and the possible three-dimensional structure of

    Screenshot of RoseTTAFold's paper from science.

    The Hermite system is designed by the well-known mathematician of Peking University, Ao Weinan, Zhang Linfeng, Sun Weijie and others, which is a new generation of drug computing platform, which can complete protein structure prediction, compound screening, lead prediction and other work

    These are just a few examples
    of how AI predicts protein structure.
    At present, there
    are countless companies and research teams that are truly committed to this field.

    More and more companies, talents, and funds are invested in it, what does this mean for protein research?

    This new way of protein research is getting closer and closer, which can not only be used to develop new drugs, develop vaccines, and discover the treatment mechanism of complex diseases, but also closely related to
    our environmental detection and environmental protection.

    Perhaps, we will soon be able to feel the new changes
    brought about by AI and protein structure prediction.


    ESM Metagenomic Structure Atlas by Meta AI (esmatlas.

    AlphaFold’s new rival? Meta AI predicts shape of 600 million proteins (nature.

    Evolutionary-scale prediction of atomic level protein structure with a language model | bioRxiv

    ESM Metagenomic Atlas: The first view of the ‘dark matter’ of the protein universe (facebook.

    Metagenomics (genome.

    RoseTTAFold: Accurate protein structure prediction accessible to all – Institute for Protein Design (uw.

    Accurate prediction of protein structures and interactions using a three-track neural network | Science

    This article is an English version of an article which is originally in the Chinese language on and is provided for information purposes only. This website makes no representation or warranty of any kind, either expressed or implied, as to the accuracy, completeness ownership or reliability of the article or any translations thereof. If you have any concerns or complaints relating to the article, please send an email, providing a detailed description of the concern or complaint, to A staff member will contact you within 5 working days. Once verified, infringing content will be removed immediately.

    Contact Us

    The source of this page with content of products and services is from Internet, which doesn't represent ECHEMI's opinion. If you have any queries, please write to It will be replied within 5 days.

    Moreover, if you find any instances of plagiarism from the page, please send email to with relevant evidence.