Milestone moment: David Baker’s team used generative AI to redesign antibodies, and AI entered the US$100 billion antibody drug market
Recently, protein design pioneer David Baker used generative artificial intelligence to design a new antibody from scratch for the first time. This is a milestone moment, marking the beginning of AI's entry into the hundreds of billions of US dollars of antibody drug market by designing proteins from scratch.
?
Antibodies are the main type of protein therapy. Currently, more than 160 antibody drugs have been approved for marketing worldwide, and the market value is expected to reach US$445 billion in the next five years. Despite the interest of the pharmaceutical industry, the development of therapeutic antibodies still relies on animal immunization or screening of antibody libraries to identify candidate molecules that bind to the target of interest. These methods are time-consuming and labor-intensive and may not generate antibodies that interact with therapeutically relevant epitopes.
?
In recent years, artificial intelligence (AI) has made many breakthroughs in areas such as predicting protein structures and designing new proteins from scratch. However, using AI to design structurally precise antibodies from scratch remains elusive.
?
This research represents an important step in applying AI protein design tools to create entirely new antibodies. The research team stated that AI can simplify the current expensive, time-consuming and labor-intensive development of antibody drugs and democratize the ability to design antibodies. Ten years from now, AI will be how we design antibodies.
The diffusion model is a generative simulation method that has achieved considerable success in image and text generation modeling. The AI painting that has exploded in recent years is based on the diffusion model. Furthermore, diffusion models appear to be applicable to protein design as well. However, the success rate of diffusion models is not high when applied to protein modeling, and the generated sequences are basically unable to fold into the target structure. This may be due to the complexity of the relationship between protein backbone geometry and sequence structure.
?
On July 11, 2023, David Baker's team published a blockbuster research paper in the journal Nature titled: De novo design of protein structure and function with RFdiffusion [2].
?
They developed and described a deep learning method for de novo protein design, RFdiffusion, that generates a variety of functional proteins, including topologies never seen in natural proteins.
RFdiffusion is a comprehensive improvement over current protein design methods. It can design proteins with a total length of up to 600 amino acid residues from scratch and achieve unprecedented complexity and accuracy. More importantly, RFdiffusion can design proteins that bind to target proteins. However, this binding relies almost entirely on the interaction between regular secondary structures (helices or chains) and target protein epitopes, so these binding proteins cannot be counted as antibodies, and RFdiffusion cannot design antibodies from scratch.
?
To overcome this, David Baker's team made further modifications to the RFdiffusion model. They fine-tuned the RFdiffusion model by training it using thousands of experimentally determined structures of antibodies bound to target antigen proteins and other antibody-like molecular interactions.
领英推荐
?
Using the adjusted RFdiffusion model, the research team designed thousands of antibodies that can recognize and bind to specific regions of antigenic proteins of multiple bacteria and viruses, including F protein site III of respiratory syncytial virus (RSV), Hemagglutinin of influenza virus, S protein RBD of new coronavirus, TcdB of Clostridium difficile, and target proteins of several cancer therapeutic drugs (such as Her2, PD-L1, IL7Ra). They then synthesized some of these designed antibodies in the lab and further tested whether the antibodies could bind correctly to the target protein.
?The test results showed that about 1% of the de novo designed antibodies could successfully bind to and function on the target protein. The research team used cryo-electron microscopy technology to further determine the structure of one of the influenza virus antibodies. The design reached atomic level precision and found that It identifies the expected portion of the target protein.
Currently, a handful of companies (such as Generate: Biomedicines) have begun using generative artificial intelligence to help develop antibody drugs. David Baker hopes that the RFdiffusion model can help solve challenging drug targets, such as G protein-coupled receptors (GPCRs), currently the most important drug target class. However, it should be pointed out that the antibodies currently designed from scratch using the RFdiffusion model are still a long way from entering the clinic. The binding of these designed antibodies to their target proteins is not particularly strong.
?
Any antibody used for treatment needs to have its sequence modified so that it resembles a natural human antibody so as not to trigger an immune response. The antibodies the institute designed are so-called "nanobodies," similar to those found in alpacas and sharks, and not the vast majority of currently approved antibodies. Nanobodies naturally lack light chains and contain only one heavy chain variable region (VHH) and two conventional CH2 and CH3 regions. The structure is simpler and therefore easier to design and study.
?
Finally, the research team stated that this is a proof-of-principle work. The initial success of this work paves the way for de novo design of antibody drugs. This is like a milestone moment, confirming that it is feasible to use generative artificial intelligence to design new antibodies from scratch. , is expected to open a new era of structure-based artificial intelligence antibody design.
Reference:
Source: Biological World