A look at the next generation of AlphaFold

Progress update: Our latest AlphaFold model shows significantly improved accuracy and expands beyond proteins to other biological molecules, including ligands

Since its launch in 2020, AlphaFold has revolutionized the way we understand proteins and their interactions. Google DeepMind i Isomorphic laboratories together they worked to build the foundations of a more powerful AI model that expands beyond just proteins to the full range of biologically relevant molecules.

Today we are here sharing an update on progress towards the next generation of AlphaFold. Our latest model can now generate predictions for almost all molecules in Protein Data Bank (PDB), often achieving atomic precision.

It enables fresh understanding and significantly improves accuracy in many key classes of biomolecules, including ligands (diminutive molecules), proteins, nucleic acids (DNA and RNA), and those containing post-translational modifications (PTMs). These different types of structures and complexes are crucial to understanding biological mechanisms in the cell, and predicting them with high accuracy is challenging.

The model’s expanded capabilities and performance can aid accelerate biomedical breakthroughs and make the next era of “digital biology” a reality – providing fresh insights into how disease pathways function, genomics, biorenewable materials, plant immunity, potential therapeutic targets, drug design mechanisms, and fresh platforms enabling protein engineering and biology synthetic.

Above and beyond protein folding

AlphaFold was a major breakthrough in predicting single-chain proteins. AlphaFold multimer then extended to complexes with multiple protein chains, followed by AlphaFold 2.3, which improved performance and extended coverage to larger complexes.

In 2022, AlphaFold structure predictions for almost all cataloged proteins known to science were made freely available via AlphaFold protein structure databasein cooperation with the European Bioinformatics Institute (EMBL-EBI) EMBL.

To date, 1.4 million users in more than 190 countries have accessed the AlphaFold database, and scientists around the world have used AlphaFold’s predictions to aid advance research in everything from accelerating fresh malaria vaccines and advances in cancer drug discovery to developing plastic-eating enzymes to combat pollution.

Here, we demonstrate AlphaFold’s remarkable ability to predict correct structures beyond protein folding, generating highly correct structure predictions of ligands, proteins, nucleic acids, and post-translational modifications.

Accelerating drug discovery

Early analysis also shows that our model significantly outperforms AlphaFold 2.3 on some protein structure prediction problems that are significant for drug discovery, such as antibody binding. Additionally, correct prediction of protein-ligand structures is an extremely valuable tool in drug discovery because it can aid scientists identify and design fresh molecules that could become drugs.

The current industry standard is to exploit “docking methods” to determine interactions between ligands and proteins. These docking methods require a inflexible structure of the reference protein and a suggested position to which the ligand should bind.

Our latest model sets a fresh bar in protein-ligand structure prediction, outperforming the best-described docking methods, without the need for a reference protein structure or ligand pocket location – enabling the prediction of entirely fresh proteins that have not been previously structurally characterized.

It can also collaboratively model the positions of all atoms, enabling the full, inherent flexibility of proteins and nucleic acids to be represented as they interact with other molecules, which is not possible using docking methods.

Here, for example, are three recently published, therapeutically relevant cases in which the structures predicted by our latest model (shown in color) closely match the experimentally determined structures (shown in gray):

1. PORCN: A clinical-stage anticancer molecule bound to its target together with another protein.
2. KRAS: Ternary complicated with a covalent ligand (molecular glue) of an significant cancer target.
3. PI5P4Kγ: Selective allosteric lipid kinase inhibitor with numerous disease consequences, including cancer and immunological disorders.

Isomorphic Labs applies the next-generation AlphaFold model to therapeutic drug design, helping to rapidly and accurately characterize many types of macromolecular structures significant in treating diseases.

A fresh understanding of biology

By enabling the modeling of protein and ligand structures along with nucleic acids and those containing post-translational modifications, our model provides a faster and more correct tool for studying fundamental biology.

One example is structure CasLambda bound to crRNA and DNAHi CRISPR family. CasLambda provides the ability to edit the genome CRISPR-Cas9 system, commonly known as “genetic scissors,” which researchers can exploit to change the DNA of animals, plants and microorganisms. CasLambda’s smaller size may allow for more proficient exploit in genome editing.

The latest release of AlphaFold’s ability to model such complicated systems shows us that artificial intelligence can aid us better understand these types of mechanisms and accelerate their exploit in therapeutic applications. There are more examples available in our progress update.

Progress of scientific exploration

The dramatic leap in performance of our model shows that artificial intelligence can greatly improve scientific understanding of the molecular machines that make up the human body and the broader natural world.

AlphaFold has already become a catalyst for significant scientific advances around the world. Now the next generation of AlphaFold has the potential to aid advance scientific research at digital speed.

Our dedicated teams at Google DeepMind and Isomorphic Labs have made great progress on this crucial work, and we look forward to sharing our continued progress.