Abstract
Vaccines are the main pharmaceutical intervention used against the global public health threat posed by influenza viruses. Timely selection of optimal seed viruses with matched antigenicity between vaccine antigen and circulating viruses and with high yield underscore vaccine efficacy and supply, respectively. Current methods for selecting influenza seed vaccines are labor intensive and time-consuming. Here, we report the Machine-learning Assisted Influenza VaccinE Strain Selection framework, MAIVeSS, that enables streamlined selection of naturally circulating, antigenically matched, and high-yield influenza vaccine strains directly from clinical samples by using molecular signatures of antigenicity and yield to support optimal candidate vaccine virus selection. We apply our framework on publicly available sequences to select A(H1N1)pdm09 vaccine candidates and experimentally confirm that these candidates have optimal antigenicity and growth in cells and eggs. Our framework can potentially reduce the optimal vaccine candidate selection time from months to days and thus facilitate timely supply of seasonal vaccines.
Introduction: The Challenge of Seasonal Influenza
Every year, the global scientific community faces a high-stakes race to update seasonal influenza vaccines. The goal is to identify a "seed virus" that satisfies two critical, yet often conflicting, criteria: an antigenic match to circulating strains to ensure efficacy, and a high production yield in substrates like embryonated chicken eggs or mammalian cells to ensure adequate supply.
Traditionally, this selection process is a major bottleneck. It involves labor-intensive laboratory passaging and genetic reassortment, which can take up to six months. Furthermore, the process of adapting a virus for better growth in eggs can inadvertently change its antigenic profile, leading to a vaccine mismatch. The MAIVeSS framework aims to bypass these hurdles by using predictive modeling to identify naturally occurring viruses that already possess both essential traits.
The MAIVeSS Framework
The Machine-learning Assisted Influenza VaccinE Strain Selection (MAIVeSS) framework streamlines the identification of vaccine candidates directly from clinical sequence data. By identifying molecular signatures for both antigenicity and yield, the system reduces the selection timeline from months to just a few days.
Methods and Pipeline
The researchers developed MAIVeSS using a combination of experimental phenotyping and advanced sparse learning algorithms. The process began with the creation of a mutant library consisting of 189 A(H1N1)pdm09 viruses. These mutants were meticulously phenotyped for growth in cells and eggs, antigenic profiles via HAI assays, and glycan-binding preferences.
To handle the high-dimensional genetic data, the team employed two specific machine learning models:
- MTL-GGSL (Multi-task learning group-guided sparse learning): Used to predict antigenicity and glycan binding by identifying synergistic groups of mutations.
- GHSM (Generalized hierarchical sparse model): Used to assess growth and yield, specifically looking for interactions between multiple mutations.
Key Insights: Yield is Encoded in the Sequence
A central finding of the study is that high-yield traits are linked to specific molecular signatures that can be predicted from sequence data alone. The researchers identified 38 specific residues associated with virus yield.
Crucially, the study revealed that high-yield variants are often already circulating in the human population. Instead of engineering these traits in a lab, MAIVeSS allows researchers to find them in nature. Furthermore, the team discovered that certain mutations can boost yield by 100 to 800-fold without altering the virus's antigenic signature. This "decoupling" of match and yield ensures that production volume does not have to come at the expense of vaccine efficacy.
Results and Experimental Validation
When applied to over 11,000 A(H1N1)pdm09 sequences from the GISAID database, MAIVeSS identified 761 candidates predicted to have high yields in both eggs and cell cultures. The framework tracks these candidates across time and geography, ensuring the selection remains relevant to current outbreaks.
Experimental validation confirmed the model's predictive power. Four identified candidates were synthesized and tested in the lab; all four reached titers at least 100-fold higher than the standard wild-type virus used for comparison. The model also accurately categorized viruses into their respective antigenic clusters, ensuring a perfect match for circulating strains.
Future Outlook
While this study focused on A(H1N1)pdm09, the MAIVeSS framework is designed to be adaptable. Future work will involve expanding the model to include other influenza subtypes like H3N2 and Influenza B. Additionally, incorporating human serological data could further refine the model's antigenic predictions, moving beyond the current gold standard of ferret antisera. By slashing the time required for seed virus selection, MAIVeSS provides a vital tool for more agile and effective global influenza response.
Summary updated