research
A new AI system designs proteins that successfully bind to target molecules, which has great potential for advancing drug design, understanding of disease, and more.
Every biological process in the body, from cell growth to immune response, depends on interactions between molecules called proteins. Like a key in a lock, one protein can bind to another to help regulate important cellular processes. While protein structure prediction tools like AlphaFold have already provided tremendous insight into how proteins interact with each other to perform their functions, these tools cannot create new proteins that directly manipulate these interactions.
However, scientists can create new proteins that successfully bind to target molecules. These binders can help researchers accelerate progress in a wide range of research areas, including drug development, cell and tissue imaging, disease understanding and diagnosis, and even crop resistance to pests. Recent machine learning approaches to protein design have made great strides, but the process is still arduous and requires extensive experimental testing.
Today, we introduce AlphaProteo, the first AI system for designing novel high-strength protein binders that serve as building blocks for biology and health research. This technology has the potential to accelerate our understanding of biological processes and aid in drug discovery, biosensor development, and more.
AlphaProteo can generate novel protein binders for a variety of target proteins, including VEGF-A, which is associated with cancer and diabetes complications. This is the first time that an AI tool has been able to design a successful protein binder for VEGF-A.
AlphaProteo achieved higher experimental success rates and 3- to 300-fold higher binding affinities than the best existing methods for seven target proteins we tested.
Learn the complex ways proteins bind to each other
Protein binders that can bind tightly to target proteins are difficult to design. Traditional methods are time-consuming and require extensive laboratory work. After the binder is created, additional experimental processes are performed to optimize binding affinity so that it binds tightly enough to be useful.
Trained on the vast protein data from the Protein Data Bank (PDB) and over 100 million predicted structures from AlphaFold, AlphaProteo has learned the countless ways in which molecules bind to each other. Given the structure of a target molecule and a set of preferred binding sites on that molecule, AlphaProteo generates candidate proteins that bind to the target at those sites.
Illustration of predicted protein binder structures interacting with target proteins. In blue is the predicted protein binder structure generated by AlphaProteo that is designed to bind to the target protein. In yellow is the target protein, specifically the SARS-CoV-2 spike receptor binding domain.
Demonstrating success for important protein binding targets
To test AlphaProteo, we designed binders for a variety of target proteins, including two viral proteins involved in infection, BHRF1 and the SARS-CoV-2 spike protein receptor binding domain, SC2RBD, and five proteins involved in cancer, inflammation, and autoimmune diseases, IL-7RÉ‘, PD-L1, TrkA, IL-17A, and VEGF-A.
Our system has highly competitive binding success rates and best-in-class binding strengths. For seven targets, AlphaProteo generated candidate proteins in-silico that bound strongly to the intended protein when experimentally tested.
A grid of predicted structures for seven target proteins for which AlphaProteo generated successful binders. Blue is an example of a binder tested in the wet lab, yellow is the protein target, and dark yellow highlights the intended binding region.
For a specific target, the viral protein BHRF1, 88% of the candidate molecules successfully bound when tested in Google DeepMind Wet Lab. AlphaProteo binders also bind, on average, 10x stronger than the best existing design methods, based on the tested targets.
For another target, TrkA, our binder, after multiple rounds of experimental optimization, is more potent than the best binders previously designed for this target.
Bar graph comparing experimental in vitro success rates for AlphaProteo’s output for each of the seven target proteins to other design methods. A higher success rate means fewer designs need to be tested to find a successful binder.
Bar graph comparing the best affinity for AlphaProteo’s design without experimental optimization for each of the seven target proteins to other design methods. Lower affinity indicates that the binder protein binds more tightly to the target protein. Note the logarithmic scale of the vertical axis.
Verification of results
Beyond that In silico To validate and test AlphaProteo in the wet lab, we engaged the research groups of Peter Cherepanov, Katie Bentley, and David LV Bauer at the Francis Crick Institute to validate the protein binders. Through a variety of experiments, they dug deeper into some of our strong SC2RBD and VEGF-A binders. The research group confirmed that the binding interactions of these binders were indeed similar to what AlphaProteo predicted. The group also confirmed that the binders had useful biological functions. For example, some of our SC2RBD binders have been shown to block SARS-CoV-2 and some of its variants from infecting cells.
AlphaProteo’s performance demonstrates that it can significantly reduce the time required for initial experiments involving protein binders for a wide range of applications. However, we recognize that AI systems have limitations, as we were unable to design a successful binder for our eighth target, TNFɑ, a protein associated with autoimmune diseases such as rheumatoid arthritis. We chose TNFɑ as a strong challenge for AlphaProteo because our computational analysis showed that designing a binder would be extremely difficult. We will continue to improve and expand AlphaProteo’s capabilities, ultimately aiming to address these challenging targets.
Achieving strong binding is usually only the first step in designing a protein that could be practically useful; there are many more biotechnological hurdles to overcome during the research and development process.
Towards Responsible Development of Protein Design
Protein design is a rapidly evolving technology with great potential to advance science on everything from understanding what causes disease to accelerating the development of diagnostic tests for viral outbreaks, enabling more sustainable manufacturing processes, and even removing pollutants from the environment.
Given the potential risks to biosecurity, and building on our longstanding approach to responsibility and safety, we are working with leading external experts to share this work and inform a phased approach that feeds into community efforts to develop best practices, including through the Nuclear Threat Initiative’s new AI Bio Forum.
Going forward, we will work with the scientific community to leverage AlphaProteo for impactful biology problems and understand its limitations. We have also been exploring drug design applications at Isomorphic Labs and are excited to see what the future holds.
At the same time, we are continually improving the success rate and affinity of the AlphaProteo algorithm, expanding the range of design problems it can solve, and collaborating with researchers in machine learning, structural biology, biochemistry, and other fields to develop responsible and more inclusive protein design proposals for the community.