Vaccine Development Gets a Boost with AI

In the early 2000s, immunologist Bali Pulendran started his lab at Emory University to explore the mechanisms underlying vaccine-mediated immunity. To uncover these processes, he focused on the yellow fever vaccine because it induced long-term immunity, sometimes for decades or even a lifetime. When he started digging into what was known about the underlying protective mechanisms, he found that “the truthful answer to this question was not a whole lot.”

So, he and his team set out to uncover the immune mechanism behind the protective responses from the yellow fever vaccine. They started with finding out how the vaccine activated dendritic cells, since these sentinel cells are important bridges between innate and adaptive immune responses. Pulendran and his team found that the yellow fever vaccine activated several important receptors on dendritic cells that stimulated the production of proinflammatory molecules that led to robust immune responses.¹

Yet, Pulendran wanted to be able to delve deeper into the mechanisms of this immunity. “What we decided to do was to do an experiment in humans, and then to analyze the immune response to this Yellow Fever vaccine, but then to apply the tools of multiomics and systems biology,” he said.

The team vaccinated 15 individuals with the yellow fever vaccine and collected blood samples before and then at four points after the vaccination. They used high throughput technologies to assay various genes, molecular markers, and cell types activated by the vaccine shortly that gave rise to protective T and B cell responses.

As a result, the team collected a large amount of data. Finding specific signatures from this multifaceted information would have been challenging because of the data complexity and abundance, but at that time, computational tools had recently been applied to successfully predict chemotherapy responses and diagnose infections.^2,3 Pulendran’s team reasoned that these models could also help parse their immunology data.

Applying these tools, in 2008, Pulendran’s team identified cytokines and several genes involved in immune viral sensing and antiviral responses that the vaccine activated that also predicted later T cell and antibody responses; these genes, then, could be used as a signature to indicate protection after vaccination.⁴ In an experiment within this same study, this gene signature predicted immune responses from a separate group of individuals that Pulendran’s team vaccinated with the yellow fever vaccine. With this study, Pulendran’s groups became one of the first to pioneer a systems biology approach to vaccinology.

AI Parses Big Data to Yield Vaccine Insights

Today, Pulendran applies these tools to understand immune responses to vaccines, recently focusing on the seasonal influenza vaccine and a pre-pandemic avian influenza formulation.^5-7 “The data that’s been generated from these kind of studies have yielded a rich bonanza of new insights about the working of the human immune system, the mechanistic basis of human immunity,” Pulendran said.

These great datasets introduce a new problem, though. “We’re drowning in a sea of data, but thirsting for knowledge,” Pulendran quoted biologist Sydney Brenner to explain the outcomes of applying high-throughput techniques.

However, whereas Brenner believed that systems biology approaches yield minimal valuable scientific outputs, Pulendran disagreed. “Over the last decade there’s been tremendous new insights that have emerged from analyzing this big data,” he said. “The recent boom in AI is tremendously exciting, because what it offers is at least logarithmic scale improvements in our capacity to mine big data, to delineate patterns in big data, and then to shed new biological insight.”

Pulendran and his team have applied AI tools and systems biology to investigate shared vaccine responses across different immunizations to predict protection. In a collaborative study, Pulendran and other groups used blood transcriptional profiles from individuals before and after vaccination to identify gene signatures that predicted antibody responses across 13 vaccines.⁸ In a larger study, these profiles helped create an atlas of response kinetics across these vaccines.⁹

More recently, Pulendran’s team used machine learning to find a predictive signature based on platelet responses that could determine the longevity of antibody response after immunization with an avian influenza vaccine.⁷

A Vaccine Biotech Finding Novel Antigens with AI

AI tools for vaccine development are not relegated to the labs and offices of academia, though. In 2008, bioinformatician Andreas Holm Mattsson and Niels Iverson Møller launched a biotechnology company, Evaxion, with AI-powered insights at the heart of their scientific plan. Their goal was to train machine learning models on bacterial genomes so that they could predict antigens that were most likely to yield effective immune responses in the hopes of reducing the timelines, and by proxy the costs, of vaccine development.

“That was the ambition of developing a platform, so reducing cost, time, and also increasing the likelihood of actually generating a drug that would be approved in the end,” said Birgitte Rønø, chief scientific officer and chief operating officer of Evaxion.

Rønø joined Evaxion after working at a pharmaceutical company using traditional research and development approaches. “Everything is slow, and you build on the way that you’ve done this for many years,” she said. “[Evaxion’s approach] was new and very intriguing for me, that you could put the immune system into a computer, and then find the targets that would be the ones protecting patients in the end.”

Over the last 18 years, Evaxion scientists have improved their platform, called AI-Immunology, leveraging public and in-house developed data sets and expanding them beyond their original platform to include modules that identify, separately, T and B cell antigens. These different modules can be combined, like Lego blocks, to identify candidates with desired properties.

At the end of an analysis, the AI-Immunology platform provides the scientists with ranked lists of potential vaccine antigens based on the specific properties they are targeting. These ratings allow the Evaxion researchers to start with targets most likely to lead to effective vaccines. “There are always very strong hits in that list. We have even been able to show that there is a correlation between protection, and then also how the AI-Immunology model ranks these different targets,” Rønø said.

Sometimes, she noted, the models suggest antigens that external researchers have doubts regarding their efficacy as a vaccine. But, Rønø said, “We can definitely see that those targets are driving protection in our animal models, so novelty is definitely something that is key for AI-Immunology. Find something that no one else can find.”

Yet, these novel targets are not always immediately accepted by stakeholders. “One of the challenges is…how is [AI-predicted antigens] perceived by the scientific community, and also the community, and also potential partners. What is in that black box?” Rønø said, adding that many had questions about how targets are identified or why they are predicted to work. “We don’t always know. We know that the targets protect, but we don’t know how they do it. So understanding that and setting up the right experiments to understand that can be a bit challenging.”

AI Models Aid in Cancer Vaccine Designs

Today, Evaxion has several vaccines in the preclinical and clinical stages: three cancer vaccines and ones against Staphylococcus aureus, Neisseria gonorrhea, and cytomegalovirus. Their leading cancer vaccine candidate, a personalized peptide-based vaccine against individuals’ tumors, is in Phase 2 clinical trials. Reflecting on watching a vaccine go from a predicted target to human studies, Rønø said, “That journey has been really, really exciting to see that we’re continuously building on data and on experience.”

So far, the group has reported an overall 75 percent response rate, with 92 percent of trial participants still demonstrating protective responses after two years. “So that’s very encouraging, and if we compare to a standard-of-care monotherapy, this is definitely very favorable data,” Rønø said, adding that the trial is validating the use of AI models to predict protective vaccine targets.

Currently, to produce a cancer vaccine, scientists sequence either the whole genome, whole exome, or RNA from a patient’s tumor to identify mutations. Then, they predict which of these mutations are most likely related to the tumor, since not all sequence variations cause deleterious effects. Alongside this, researchers must consider the immunogenicity—or how well a particular candidate elicits an immune response—safety, interactions, and many other factors of potential candidates.

This process, while important, is not fast. Guangyuan Li, a computational biologist at University of Notre Dame, noted that can be a huge detriment since many patients exploring cancer vaccines already have tumors in advanced stages. “They don’t have time to wait for another four months or five months for you to validate the validity of every single antigen,” he said.

Li uses AI to help identify novel tumor antigens to improve cancer immunotherapy. But even to target a short peptide, for instance one with only nine amino acids, the potential combinations can lead to astronomical possibilities. “So, we really need computational tools to help us to narrow down to a short list so that we don’t have to find a needle in a haystack,” Li said.

This is where machine learning and AI tools can provide huge benefits. As a graduate student, he created DeepImmuno, a tool that ranks antigens based on their immunogenicity, to expedite the process of selecting effective tumor antigens.¹⁰ As opposed to traditional screens of hundreds of peptides over a month or more, the software runs in a matter of seconds, according to Li. “That’s the power or realistic example about how those computational models [are] helping or facilitating the development process,” he said.

However, Li said that one problem that can arise in some models is false positives in antigens. To help address this, he recently developed a second model, ImmunoVerse, to screen antigens across cancers to find tumor-specific mutations.¹¹ While he developed this with CAR T cell therapy in mind, much like DeepImmuno, they can aid with the identification of vaccine candidates as well.

Despite these many opportunities, challenges remain in incorporating AI successfully into vaccine development pipelines. One problem, universal across machine learning applications, remains obtaining high-quality data. The Evaxion team overcame this issue by using a combination of public data and their own proprietary set of information. However, Li said that for more models to become more valuable, they will need access to reliable training data. One concern he shared was that individuals or companies may seek opportunities to produce and then charge potential users for these types of resources.

“I don’t know how things will play out, but unfortunately the sentiment I’m seeing so far is people start to realize how powerful those large language models are, so they started to close the door,” he said. He said that this could lead to issues where researchers can’t access models trained on the most valuable data.

However, Li remains optimistic that enough researchers are coming into the field of AI that there will remain sufficient resources for new models. “I think part of my future research will be focusing on the process in which those data will be generated and hopefully that kind of practice will be adopted by those powerful players in the field,” he said.

AI Boosts Clinical Trials, but Access Challenges Risks Inequity

While AI benefits vaccine research and development, it also raises certain concerns. “If anything, AI might deepen inequity when it comes to [vaccine] access,” said Shabir Madhi, a physician at the University of the Witwatersrand, Johannesburg. For instance, AI could allow vaccine makers to pursue previously challenging vaccines of interest in high income countries as opposed to focusing on vaccines needed by low- and middle-income countries.

Developing vaccines against pathogens that are predominantly of concerns in these lower resource settings and administering vaccines to these communities has been a problem well before AI.^12,13 Although non-governmental organizations, governmental and academic institutions, and partnerships like Gavi, The Vaccine Alliance have worked to close these gaps, shortcomings in vaccine access remain.¹⁴ While pathogens like tuberculosis, HIV, and malaria receive attention, Klebsiella pneumonia, which accounts for many infant infections, and respiratory syncytial virus (RSV), another early life challenge, are far less explored by external researchers for vaccine development.¹⁵

This introduces another potential issue of current AI models as they are applied to vaccine development: bias. “The sort of clinical scenarios that are fed into the tools are not necessarily sort of inclusive of the type of disease pathology that exists in Africa,” Madhi said. This could lead to inaccurate modeling in safety and efficacy studies. Previous studies have identified several sources of unintended bias in AI models designed for healthcare settings that impact quality of care.¹⁶ Addressing these will be important to reduce potential, unintended harm in vaccine development and testing.

Although AI has the potential to strain already existing challenges in vaccine access, Madhi thinks it could also help researchers address some of the historic challenges in vaccine development. One idea is to use AI in immunobridging studies, in which an investigational vaccine is compared to an approved vaccine as opposed to a placebo. This could help introduce new, hopefully cheaper RSV vaccines while still being able to identify safety and efficacy. And, he said, “We’re in a completely different ballpark when it comes to the efficiency of AI to do an analysis plan.”

In his group, Madhi is tackling issues for two pathogens with the help of AI: K. pneumonia and RSV. Currently, there is no vaccine against K. pneumonia. His group is using machine learning models to find targets against this pathogen that could be effective across strains and serotypes of the pathogen.

I do feel very optimistic that AI is going to become the best intellectual and cognitive companion of the human scientist.

—Bali Pulendran, Stanford University

Separately, while the immunization for RSV protects infants from viral infection, it remains too expensive for many countries in need, and some reports pointed to a risk of preterm birth associated with the RSV vaccine during pregnancy in women from low- and middle-income countries. ^17-19 This prompted a call to conduct studies focused on these populations to evaluate this risk.

Madhi and his team have spoken to companies about AI tools that could improve participant recruitment, identify important populations, and help monitor study patients in real-time, such as for follow up studies for gestational RSV vaccination. These could cut down on clinical trial costs and give researchers a way to identify safety signals faster. But Madhi admitted, “We are almost in embryonic stages of the use of AI in terms of its potential for clinical trials.”

However, Madhi said that incorporating these tools requires electronic health records, which many African countries do not have. “Without that sort of digital health infrastructure, you can’t leverage on the potential of AI for surveillance in terms of detecting outbreaks, in terms of understanding safety of vaccines or any of that,” he said.

Despite these challenges, Madhi believes that AI will be “absolutely transformative” for vaccine development and testing, from identifying antigens to decreasing timelines to improving clinical trial decisions to make them more efficient and safer. But, he added, “For low-and middle- income countries to benefit, what is really going to be required is sort of the seeding of digital health infrastructure…it provides an opportunity to leapfrog into where we should be. But unless we make that investment in digital infrastructure, we risk, again for especially African countries, lagging behind and missing out on another opportunity.”

AI As the Next Research Assistant

As with any new technology, Pulendran said, “There’s a great deal of hype, and AI is not the answer to everything. We can’t switch off our brains and expect AI to do everything. So, we need to be realistic in what AI offers us but also be very critical and skeptical of what it offers us.”

While there are still limitations in the capacity, data availability, and overall access to efficient AI models, researchers anticipate that machine learning tools will be important components in vaccine development. Pulendran cited how tools like AlphaFold have already demonstrated the potential for AI, but added, “AI has so much more to offer.”

Rønø agreed, “AI can be applied in many places in the whole drug development value chain, and further reducing time and costs, and maybe even improving quality of the vaccines.” But her biggest hope for AI-predicted vaccines is the ability to bring the technology into the hospital to streamline the process from diagnosis to treatment.

Meanwhile, Pulendran hopes to see AI go beyond parsing datasets to suggest promising antigens for vaccines. Recently, he and his team evaluated the ability of existing AI models to complete tasks that utilized different cognitive abilities.²⁰ Pulendran found that when these models were given sequences previously determined to be important to vaccine responses, many could successfully review the literature for information about the genes.

In investigating their ability to propose experiments to test hypotheses, the answers reminded Pulendran of “enthusiastic high school students”, in that they could provide superficial answers that didn’t reflect much depth in the subject matter. Similarly, the models could not efficiently consider their findings in the broader context of immunology and biology.

Yet, Pulendran feels certain that future models will overcome these current limitations. “I do feel very optimistic that AI is going to become the best intellectual and cognitive companion of the human scientist,” Pulendran said. “The field is looking extremely exciting and bright.”