Creating effective vaccines against rapidly evolving viruses like influenza and HIV has presented one of medicine's most persistent challenges. These pathogens continuously adapt through "viral escape," a process that allows them to evade immune detection by altering their surface proteins, rendering many vaccines ineffective over time.
In a revolutionary breakthrough, MIT researchers have developed an innovative computational approach to predict viral escape patterns. By adapting artificial intelligence frameworks originally designed for language processing, this cutting-edge technology can accurately identify which regions of viral surface proteins are susceptible to mutation-enabled escape. Crucially, it also pinpoints stable protein segments that represent promising targets for next-generation vaccine development.
"Viral escape remains one of the most significant obstacles in virology today," states Bonnie Berger, the Simons Professor of Mathematics and head of the Computation and Biology group at MIT's Computer Science and Artificial Intelligence Laboratory. "The continuous evolution of surface proteins in influenza and HIV explains why we still lack universal vaccines for these diseases, which collectively claim hundreds of thousands of lives annually."
In their groundbreaking study published in Science, Berger and her team identified potential vaccine targets for influenza, HIV, and SARS-CoV-2. Since the paper's acceptance, the researchers have applied their predictive model to analyze emerging SARS-CoV-2 variants from the United Kingdom and South Africa. This ongoing analysis, currently undergoing peer review, has flagged specific viral genetic sequences that require urgent investigation for their potential to compromise existing vaccine effectiveness.
Berger and Bryan Bryson, an assistant professor of biological engineering at MIT and member of the Ragon Institute of MGH, MIT, and Harvard, served as senior authors of the paper, with MIT graduate student Brian Hie as lead author.
Applying Language Models to Protein Sequences
Different virus families accumulate genetic mutations at varying rates, with HIV and influenza among the most rapidly evolving. For mutations to facilitate viral escape, they must modify surface protein structures enough to prevent antibody binding while preserving the proteins' essential functionality.
The MIT team implemented language models from natural language processing (NLP) to simulate these biological constraints. Originally designed to analyze linguistic patterns and predict appropriate word completions based on grammatical rules and context, these models find surprising parallels in genetic sequences. In this application, "grammatical correctness" corresponds to functional protein structures, while "semantic meaning" relates to a protein's ability to adopt new shapes that evade immune detection.
"For a virus to successfully escape immune surveillance, it must mutate without sacrificing its ability to replicate," explains Hie. "It needs to maintain biological fitness while effectively disguising itself from the immune system."
To implement this approach, the researchers trained an NLP model to recognize patterns in genetic sequences, enabling it to predict new sequences with novel functions that still adhere to biological constraints. A key advantage of this methodology is its reliance solely on sequence information—far more accessible than detailed protein structures. Remarkably, the model achieved robust results with relatively modest training datasets: 60,000 HIV sequences, 45,000 influenza sequences, and 4,000 coronavirus sequences.
"Language models possess remarkable power because they can extract complex distributional patterns and derive functional insights from sequence variation alone," Hie notes. "By analyzing extensive viral sequence databases, our model learns the properties of amino acid relationships and variations across different positions."
Blocking Viral Escape Mechanisms
Once trained, the model successfully predicted sequences within the coronavirus spike protein, HIV envelope protein, and influenza hemagglutinin (HA) protein that were more or less likely to generate escape mutations.
For influenza, the model identified the HA protein stalk as the region least susceptible to escape mutations. This finding aligns with recent research demonstrating that antibodies targeting the HA stalk—which most flu patients or vaccinated individuals don't develop—can provide broad protection against diverse flu strains.
The coronavirus analysis highlighted the S2 subunit of the spike protein as relatively stable against escape mutations. Questions remain about SARS-CoV-2's mutation rate and the long-term durability of current COVID-19 vaccines. While preliminary evidence suggests the virus mutates more slowly than influenza or HIV, researchers have recently identified concerning variants in Singapore, South Africa, and Malaysia that merit investigation for potential escape properties.
In HIV research, the model confirmed that the V1-V2 hypervariable region contains numerous potential escape mutations, consistent with previous studies, while also identifying sequences with lower escape probability.
The team is now collaborating with other research groups to apply their model in identifying targets for cancer vaccines that activate the body's immune system against tumors. They believe the technology could also guide the development of resistance-resistant drugs for diseases like tuberculosis.
"The potential applications are extensive, and the most exciting aspect is that we only need sequence data, which is relatively straightforward to generate," Bryson observes.
The research received funding from a National Defense Science and Engineering Graduate Fellowship from the Department of Defense and a National Science Foundation Graduate Research Fellowship.
