GUIDEing Drug Development

The Generative Unconstrained Intelligent Drug Engineering (GUIDE) program accelerates development of medical countermeasure candidates to redefine biological defense.  

Antibodies are the future of therapeutics. Conventional drug design has a common problem: Drugs not only bind to their intended target but to additional, unwanted targets as well, leading to side effects. On the other hand, antibodies—proteins designed to recognize and fight pathogens containing specific antigens—are specific to a particular target. Antigens, foreign substances that enter the body, can be harmful themselves or as part of a pathogen, a larger, harmful whole. Each antibody couples and binds to a specific antigen to block, for example, a virus from entering human cells. When operating properly, the immune system produces its own antibodies. Artificial antibodies, known as monoclonal antibodies, mimic the immune system, making them extremely safe as a treatment.

In the past two decades, antibodies have gained in popularity as a treatment relative to conventional, small-molecule drugs. However, antibody development can take years. Laboratory-based development methods most commonly require scientists to wait for a patient infected with a disease to recover, after which the patient’s blood can be analyzed for antibodies. Scientists either use these antibodies directly or mimic their design. These approaches are not only time consuming but also constrained by humans’ natural immune responses. A patient’s blood may be difficult to study, or the infection may not leave a survivor at all. Such limitations have made antibodies a nonideal method for managing urgent health needs such as the world saw during the COVID-19 pandemic. “We get what we get,” says Dan Faissol, a computational engineer at Lawrence Livermore. “Our study is limited to the patient and the response.” 

Faissol is Livermore’s principal investigator for the Generative Unconstrained Intelligent Drug Engineering (GUIDE) program, a Department of Defense (DOD) effort executed on behalf of the Chemical and Biological Defense Program (CBDP) by the Joint Program Executive Office for Chemical, Biological, Radiological, and Nuclear Defense’s (CBRND’s) Joint Project Lead for CBRND Enabling Biotechnologies. With funding from the GUIDE program, interagency teams integrate computational approaches—including molecular simulations, machine learning (ML), deep learning, and language models—with experimental approaches to develop antibody designs as medical countermeasures, both preemptively and in response to ongoing biological threats. “Our immune systems make antibodies already. GUIDE is trying to achieve the same goal computationally and validate antibodies in the laboratory,” explains Faissol. 

The primary goal of GUIDE is to develop the predictive tools and experimental capabilities required to enable a rapid response to new pathogens and accelerate countermeasures to biological and medical emergencies. This unprecedented use of high-performance computing (HPC) for biologics will greatly accelerate the drug development cycle. Although experiments and computational simulations each come with their own advantages and drawbacks—for example, the time expense needed for experiments and the lack of biological tests in computational simulations—GUIDE can help leverage the strengths of each component while mitigating their limitations. Data on emerging pathogens is expected to be limited, but theory-driven computational methods, such as molecular dynamics simulations that can predict how proteins bind, are a viable alternative to data-driven approaches for generating accurate predictions.

Leading the Search for Antibodies

The GUIDE platform integrates existing data from related experiments, biological models, and molecular dynamics simulations to provide an initial set of conditions and predictions of what a potential antibody for a specific antigen may look like. ML and optimization methods are then used to iterate off these designs. The process can be broken down into three primary phases: First, researchers formulate the problem they need to solve by identifying the target antigens and a known related antibody for a similar antigen. Then, team members design a computational approach for considering antibody properties, such as binding affinity to a target antigen, and use this information to generate an initial genetic sequence by proposing changes to the initial antibody to address the desired properties. Finally, the experimental team verifies the predicted antibody design’s binding properties, suitability for human use, and thermal stability. (See the box below.)

“Part of the risk in making predictions is not just the set of molecular simulation approaches and algorithms we use but the input structural antibody–antigen models that we prepare for processing,” says GUIDE structural bioinformatics expert Adam Zemla. “The approach can be fantastic and work well in most cases, but, still, the wrong input yields the wrong output.” This risk, along with the fact that simulations of interactions between molecules are computationally expensive, means that simulating every possible variant of an antibody design is not computationally feasible or productive. ML, therefore, plays an important role in determining what to simulate. ML directs the selection of simulations by proposing modifications to the initial parameters to make predictions for improved antibody design. Then, other computational tools estimate how well the proposed antibody binds to the desired antigen to corroborate or reject these selections. 

These computational steps work back into a feedback loop, and the proposed antibody modifications are then incorporated back into the ML model for the next iteration of predictions. Together, ML identifies potential antibody sequences, and high-fidelity simulations test how well these predictions will work. Much of this data is not pre-existing, so the datasets GUIDE develops for these models are the first of their kind. 

When antibodies and antigens interact, they press against each other, creating an interface. The majority of the simulations the GUIDE computational team conducts are binding free-energy simulations, which quantitatively measure the favorability of these interface interactions. “Since the interaction between these molecules is mediated by forces that happen primarily at the antigen−antibody interface, changing one or the other of those molecules changes the favorability of the interaction,” explains Thomas Desautels, who led the computational component of GUIDE through February 2024. When a virus mutates, as has been the case throughout the COVID-19 pandemic, the interaction between the antibody and antigen is affected, and the antibody needs to be adjusted accordingly. “We’re trying to study a change in the favorability of the interaction after the antigen is modified. Then, we strive to make compensatory changes in the antibody to get back to a strong interaction,” says Desautels.

Although these ML-based predictions will ultimately require experimental validation, they save a significant number of experimental steps by eliminating the need for constant experimental feedback between multiple iterations (experimental evaluation only occurs once, at the end of the computational evaluation). While experiment-driven drug development can take years to complete, the GUIDE process cuts the drug discovery process to a mere four or five weeks using ML. 

The computational team provides the experimentalists with antibody candidates in the form of a digital amino acid sequence, which is essentially a long string of letters. Once the biologists filter the candidates provided by the computational team, they determine the DNA sequence of the antibodies from the digital sequence, which they clone into a plasmid vector to produce the protein. They then purify and assay these proteins for experimental testing. Using x-ray crystallography or cryogenic electron microscopy, the researchers observe structural properties of the antibody proteins, including how well they interact with the target antigen.

Prior to the onset of the COVID-19 pandemic, GUIDE had performed this work in the opposite direction. Rather than modeling antibody properties, the GUIDE team used their experimental and computational capabilities to study antigen properties—specifically, for Meningitis B. In contrast to antibodies, this approach to vaccine development requires introducing a weak version of the antigen to the body to train its immune response. Despite the flipped approach, the basis of the work is, in essence, the same. “At the core, what we’re modeling is an antibody−antigen interaction. The question is which protein we’re modifying. We use the same kind of simulation to do the prediction,” explains GUIDE structural biologist Fangqiang Zhu. Desautels adds, “The idea is that we’re learning something about the physical interaction of the two proteins—the antibody and its antigen—from the simulations.”

Early Successes

In early 2020, shortly after the emergence of SARS-CoV-2, the virus that causes COVID-19, researchers at Lawrence Livermore—along with scientists all over the world—shifted their research focus toward this lethal threat. At the Laboratory, scientists involved in GUIDE investigated the application of HPC to screen variations of the antibodies known to bind and neutralize SARS-CoV-1, the original SARS virus, which caused an outbreak in the early 2000s. Their goal was to propose modifications to its amino acid sequences—the building blocks of proteins, including antibodies—that could allow the antibodies to bind the SARS-CoV-2 spike protein. By February 2020, in just 22 days, the team had accurately predicted the protein structure of the SARS-CoV-2 receptor binding domain, the key antigen target, and suggested 20 computationally identified antibodies predicted to target it. 

Even before the onset of the SARS-CoV-2 Omicron variant, scientists understood the importance of future-proofing vaccines. “We were aware that changes in SARS-CoV-2, ways the virus could mutate, would render this antibody ineffective,” explains Zhu. These predictions ultimately came true. With the Omicron variant, which the World Health Organization designated a “variant of concern,” drugs began to lose their potency. 

During this period, the GUIDE team engaged in what they call the “Omicron sprint” as U.S. government officials requested its help addressing the emerging variants. “All of a sudden, we were called to respond in real time, and we needed to pivot,” says Kathryn Twigg Arrildt, a GUIDE virologist who led the biology–experimental component of GUIDE through September 2024.  

Livermore researchers worked with the biopharmaceutical company AstraZeneca to computationally redesign one of their SARS-CoV-2 antibodies to recover its potency against the Omicron variant and its subvariants. For immunocompromised individuals, AstraZeneca offers a COVID-19 prophylactic antibody drug product called Evusheld, comprised of a cocktail of two long-acting antibodies, tixagevimab and cilgavimab. The goal of the Omicron sprint was to identify a replacement antibody for the cilgavimab component, which binds to a specific region of the SARS-CoV-2 spike protein to prevent it from attaching to human cells and causing a COVID-19 infection.

Starting with a version of the cilgavimab antibody known to work against earlier SARS-CoV-2 variants, the Livermore team used the GUIDE platform—backed by ML, biological modeling, HPC molecular simulations, and experimental data—to produce a list of potential antibody candidates. With only a few key amino acid substitutions, the researchers restored the original antibody’s effectiveness against the Omicron variant while maintaining its efficacy against the Delta variant, which was still the dominant variant at that point in late November 2021. 

Desautels describes the combinations of amino acids that make up antibody protein structures as beads on a string, with the string folding up for the protein to function. “When we understand how the protein is interacting with something, we know that certain beads are the ones that are actually touching the object that we want it to interact with,” says Desautels. By knowing which amino acid “beads” are likely to be most important, the team can configure the GUIDE engine to allow changes at these sites in any combination. The problem then becomes how to select among many possible designs. In redesigning cilgavimab to counter Omicron, the team applied the GUIDE engine to explore and simulate 126,000 of 1017 total allowable antibody sequences and selected 376 candidate antibodies to evaluate experimentally.  

The 376 selected antibodies were then distributed to partner institutions for evaluation support, where biologists synthesized, produced, purified, screened, and characterized each proposed antibody drug to determine whether its ability to bind to the virus was improved. In addition to checking an antibody’s binding affinity to new variants, downstream assays were also used to ensure the antibody efficiently neutralizes both new variants and old strains, and is thermally stable, safe, and easy to manufacture. Moreover, researchers needed to be certain the new antibodies did not create new liabilities—that is, viral variants that would escape the new antibody but would not have escaped the old one. Out of the antibodies tested, partner institution Vanderbilt University discovered the ideal option, which met all the required tests with high efficacy. 

Livermore and its partner institutions confirmed the new antibody’s ability to bind to different targets including types of Omicron subvariants as well as Delta variants through tests on mice. Unfortunately, in the short time between the work and its publication, COVID had moved on to a new subvariant, and the discovered antibody never made it to market, although it will continue to be used to study COVID as well as other pathogens. Still, this exercise has proven the platform’s value in identifying therapeutics in mere weeks, rather than years, as needed in the past.

Emerging Tools and Facilities

 “Lawrence Livermore has some of the fastest computers in the world, and they have attracted some of the best computational scientists in the world—scientists interested in computational solutions to biological problems,” says Arrildt. Integrating these computational capabilities with experimental science and empowering researchers to work together is as important in GUIDE as the Laboratory’s unique resources. The Laboratory’s ability to foster collaboration among diverse scientists and with external partners is invaluable to the success of programs such as GUIDE.

Among these resources is the Laboratory’s Center for Predictive Bioresilience (CPB). Co-funded by the National Nuclear Security Administration, the DOD CBDP, and Lawrence Livermore, CPB is built to accelerate biological research by uniquely integrating the Laboratory’s HPC, high-throughput experimentation, and imaging capabilities. The center provides resources and guides programs in utilizing these resources, and any developments under CPB—for example, software tools for data analysis or ML developed by a CPB member—are available to all. “GUIDE is both a user and a motivator to have the resources in the CPB because it demonstrates that when we have access to resources, we at Livermore can do something amazing,” says Barry Chen, GUIDE’s machine learning for de novo design lead. “We need it all: We need the hardware, the software, and the people.” 

At the start of GUIDE, the researchers depended on help from external companies specializing in high-throughput protein production for some of their production and testing needs. The Rapid Response Lab was born out of GUIDE’s need for a faster and more flexible platform for testing drug designs. Providing the computational team with dedicated partners that can generate and evaluate the drugs helps streamline the process into a pipeline approach. This enables better and more frequent feedback on the designs, and multiple iterations otherwise not possible. 

“Bringing that process internally lets us play with things a little, so we can tailor our process to exactly what we need,” says Edwin Saada, who leads the Rapid Response Lab. For example, some drug development may only require a small modification to an existing drug. In such cases, the GUIDE team may seek only to redesign a specific property of a drug—such as its affinity from the Delta variant to the Omicron variant in the case of their SARS-CoV-2 work. Conducting the production process at the Laboratory rather than with external partners allows the GUIDE team to scale down and accelerate the entire workflow. 

The resources used and created by GUIDE are not limited to the biological and experimental side. As a byproduct of their protein modeling work for drug design, GUIDE computational scientists developed a generative large language model. “Our models allow us to ‘fill in the blank’ in a plausible way when we want to redesign an antibody but don’t know which substitutions are more admissible or humanlike,” says Desautels. Instead of letters from the alphabet being used to build words, the model uses amino acids to build and optimize antibodies.

Adapting to Future Threats

GUIDE’s rapid response capabilities are important for multiple facets of national security, including both public health threats and defense applications. Recent improvements in the availability and accessibility of molecular and biological tools come with pitfalls: Biological threats are easier to manufacture than ever before, increasing the United States’ vulnerability to future attacks with biological weapons. “No one knows what threat will arrive next, so we’re getting ready to respond quickly when that threat does arrive,” says Faissol. 

The team continues to improve the GUIDE platform and add additional relevant calculations. Strengthening GUIDE’s generalized capabilities and expertise will allow them to respond to any threat that surfaces. “If a threat emerges—whether it’s something intentionally engineered or a pandemic we didn’t see coming—we can respond more quickly with new therapeutics and mitigate the threat,” says Saada. 

In terms of broadening the scope of GUIDE, one focus area is incorporating different types of medical countermeasures in addition to monoclonal antibodies. For example, an increased emphasis on vaccines and small molecules yields larger applicability to both biological and chemical threats. Vaccine development follows a process similar to GUIDE’s but taken on in reverse. As with the team’s past work on Meningitis B, vaccine development requires modifying antigens instead of antibodies. 

Along with DOD, Lawrence Livermore is one of two U.S. members of the Organisation for the Prohibition of Chemical Weapons (OPCW). This self-regulating international community designs regular tests and challenges its members to detect different types of potential biological and chemical weapons. By demonstrating their detection capabilities, the OPCW hopes to deter adversaries from utilizing such weapons. 

Important to this defense and national security application is GUIDE’s recent focus on small molecule design. Chemical weapons are strong binders of targets that cause immediate damage compared to viruses, which take hours or days to cause symptoms. Small molecule therapeutics can be targeted at exactly the location in the body where a chemical weapon is inhibiting normal function and break the chemical apart to reverse or detoxify its effects. These can be modified to bind to many different areas in the human body, and therefore, also have medical applications.

In addition to these military applications, GUIDE continues to make pharmaceutical breakthroughs and is working with AstraZeneca on a different COVID-19-targeting antibody that is in preclinical development. Aside from COVID-19, the team plans to eventually focus on pandemic-influenza, which is highly mutable and results in a slightly different seasonal flu every year. “Such work will usher in a new generation of therapeutics with an impact so high on human health that we can hardly imagine where it can lead,” says Faissol. Looking further into the future is the ability to design antibodies from scratch, rather than modifying known antibodies for similar illnesses: GUIDE’s ultimate grand challenge. 

As a large, multi-institutional and multidisciplinary research program, GUIDE has required Cooperative Research and Development Agreements, Non-Disclosure Agreements, and Material or Data Transfer Agreements with entities involved. Livermore’s Innovation and Partnerships Office (IPO), working closely with the Laboratory’s Office of General Counsel, has negotiated and executed these complex agreements for the GUIDE project. IPO also manages the capture of new intellectual property (IP) generated during the collaborative research by assessing inventions and then filing and prosecuting patent applications. “We look forward to licensing valuable IP being generated through GUIDE to the private sector for commercialization,” says Yash Vaishnav, Lawrence Livermore IPO business development executive who supports GUIDE.

GUIDE plays a major role in advancing the experimental element of biology research. In a biosecurity first, Tuolumne, the Laboratory’s unclassified sibling machine to its upcoming El Capitan supercomputer, will offer dedicated servers to GUIDE and other CPB projects. “We believe Tuolumne will be the largest dedicated computer for biology in the world,” says Chen. “Biology comes with a lot of data, and the time is right to integrate the data and the simulation with the experimental component.” ML is a major growth area for biological research, so access to these new machines is critical for achieving this integration even more efficiently. As Tuolumne, El Capitan, and El Capitan’s other counterparts begin to run, GUIDE’s applications will help showcase the Laboratory’s state-of-the-art capabilities. 

Even as recently as five years ago, the type of work GUIDE has achieved by integrating experimental and HPC approaches would have been impossible. With its cost-effective, accelerated approach to drug development processes, GUIDE will continue to be prepared to respond quickly to national and international emergencies including pandemics and other biological dangers. 

The team emphasizes that GUIDE is not just about the computing capabilities, the hardware, and the biological experiments, but also about the people guiding discoveries including collaborators from Sandia, Los Alamos, and Lawrence Berkeley national laboratories as well as Vanderbilt University, Washington University in St. Louis, and others. Desautels says, “Together, we’re sinking our teeth into big and important problems.”

—Anashe Bandari

For further information contact Dan Faissol (925) 423-2544 (faissol1 [at] llnl.gov (faissol1[at]llnl[dot]gov)).

The GUIDE program is executed by the Joint Program Executive Office for Chemical, Biological, Radiological, and Nuclear Defense (JPEO-CBRND) Joint Project Lead for CBRND Enabling Biotechnologies (JPL CBRND EB) on behalf of the Department of Defense’s Chemical and Biological Defense Program. The views expressed in this article reflect the views of the authors and do not necessarily reflect the position of the Department of the Army, Department of Defense, nor the United States Government. References to non-federal entities do not constitute nor imply Department of Defense or Army endorsement of any company or organization.