Bond LSC researchers create model to shorten flu vaccine development and bolster efficiency

By Sarah Gassel | Bond LSC

Flu vaccines could be getting a booster of their own with the help of machine learning.

MAIVeSS — the Machine-learning Assisted Influenza VaccinE Strain Selection framework — has the potential to reduce the time it takes to choose flu virus strains used in annual vaccines from months to mere days.

Researchers at the University of Missouri partnered with Mississippi State University to develop a computational model that combines algorithms to select optimal flu candidate vaccine viruses when provided with a specific flu strain. Their work on MAIVeSS published in the journal Nature Communications Feb. 6, 2024.

“That’s our goal; we really want to help people speed up the procedure and make effective products,” said Cheng Gao, lead author and Ph.D. student at MU’s Bond Life Sciences Center. Gao spearheaded the project to build the machine learning model. He is a member of the lab of Henry Wan, a Bond LSC principal investigator and director of the NextGen Center for Influenza and Emerging Infectious Diseases created in 2022. 

This new algorithm improves on a tried-and-true vaccine creation method. The current process to identify flu virus candidates for development of the annual flu vaccines is both time and resource-intensive. If the wrong variants are chosen, a vaccine will be ineffective against dominant strains causing illness.

Researchers choose the virus strains several months ahead of the flu season in order to leave enough time to manufacture the vaccines. Though they use the dominant strains at the time to guide their selection, they also must account for how the naturally circulating virus could mutate and choose a strain candidate that produces a broader immunity.

The MAIVeSS model combines two learning algorithms which each respectively identify the antigenic and yield properties of a virus strain. In choosing these learning algorithms, Gao used historical virological data on the currently circulating flu strain A(H1N1)pdm09 to test the predictive capabilities and accuracy of each.

Antigenicity tops a long list of ideal qualities for a viral candidate. It determines whether antibodies produced by the body in response to the vaccine can block the proteins on the surface of a virus from aiding in infecting cells. An effective virus for a vaccine matches how a naturally circulating virus stimulates the immune system.

High yield is also paramount, but strains need to maintain the same surface proteins throughout production without destroying the chicken egg or cell culture where they are grown.

As with any reliable product, the model had to go through a process of fine-tuning after being built.

“I had a lot of discussion with my advisor because we needed to refine the model; we needed to discuss every detail of the model,” Gao said.

For machine learning algorithms to make accurate predictions, they require large sets of data. While the historical virological data was a start, Gao wanted to diversify the model’s predictive capabilities as much as possible.

Feng Wen at Mississippi State University, a former lab member and Ph.D. graduate from Wan Lab, partnered to build a library of specific virus variants for computational model development. 

HA proteins are one type of surface protein on the envelope of a flu virus. These proteins attach a virus to a cell, and, therefore, are typically focused on most when developing a vaccine.

After developing mutant strains, Wen analyzed them and gave Gao the data to train the model to predict what strains blocked surface proteins best with highest yield based on the molecular features of a flu virus’ genomic sequence.

“Without the data from him, the model would be impossible,” said Gao.

After further refining the model, he used the virological data from mutants not previously used in the training to test it. The results showed the model accurately selected ideal candidate vaccine strains in consideration of these features for a given flu virus.

Though this model is only trained and specialized for A(H1N1) flu viruses — one of the four main groups of strains — Gao believes the model could also be trained for the other flu viruses. This could even further shorten vaccine production by eliminating the need to repeat the process for each of the main groups.

“One cool thing is that our model can be generalized to different applications,” Gao said. “We would need to change the training data set and the testing data set.”

The success of the model could reduce the time it takes to choose the candidate strains and prepare a good vaccine seed to a few weeks. To have a vaccine strain for next flu season, currently, this process requires surveillance of the virus across the entire flu season by the World Health Organization’s (WHO) Global Influenza Surveillance and Response System (GISRS). It tests naturally circulating flu virus samples received from patients at influenza centers to examine how virus strains mutate. A main aim of these efforts is to determine whether the current season’s vaccine is still effective by measuring differences in genetic sequences, which can indicate pre-existing antibodies will not be effective and a new vaccine strain is required.

After the vaccine strain is selected, it may take several more weeks to ensure that it can yield high production levels and meet manufacturing standards. By that time dominant strains could have already changed.

“Nowadays, artificial intelligence can help people to speed up for a lot of the process,” Gao said.

MAIVeSS could not only offer a timely solution to the problems raised by the current process but could also greatly reduce the amount of resources required in the trial period.

Gao hopes to expand his model to predict characteristics of neuraminidase proteins, another surface protein on the flu virus.

Gao welcomes people to test their own data sets on the model and give feedback at https://github.com/FluSysBio/MAIVeSS.

Several other lab members from Wan Lab, including Minhui Guan, Bijaya Hatuwal, Beatriz Praena, and Cynthia Tang, contributed to this study. Additionally, the research was a joint effort with colleagues from Mississippi State University, Rice University, Georgia State University, Clemson University, St. Jude Children’s Research Hospital, and the Food and Drug Administration.

The paper “MAIVeSS: Streamlined selection of antigenically matched, high-yield viruses for seasonal influenza vaccine production” by Cheng Gao, Feng Wen, Minhui Guan, Bijaya Hatuwal, Lei Li, Beatriz Praena, Cynthia Y. Tang, Jieze Zhang, Feng Luo, Hang Xie, Richard Webby, Yizhi Jane Tao, and Xiu-Feng Wan was published in the Nature Communications journal on Feb. 6, 2024.

This work was supported by the US National Institutes of Health grant 1R01AI116744, R01AI147640, R21AI144433, F30AI172230 and Welch Foundation grant C-1565 to YJT.

Article originally published on Decoding Science.