tetano
Editor, Senior Moderator
EBioMedicine
. 2025 May 26:116:105744.
doi: 10.1016/j.ebiom.2025.105744. Online ahead of print. Leveraging pre-vaccination antibody titres across multiple influenza H3N2 variants to forecast the post-vaccination response
Hannah Stacey[SUP] 1 [/SUP], Michael A Carlock[SUP] 2 [/SUP], James D Allen[SUP] 2 [/SUP], Hannah B Hanley[SUP] 3 [/SUP], Shane Crotty[SUP] 4 [/SUP], Ted M Ross[SUP] 5 [/SUP], Tal Einav[SUP] 6 [/SUP]
Affiliations
Background: Despite decades of research on the influenza virus, we still lack a predictive understanding of how vaccination reshapes each person's antibody response, which impedes efforts to design better vaccines. Models using pre-vaccination antibody haemagglutination inhibition (HAI) titres against the vaccine strain alone poorly predict post-vaccination responses.
Methods: We combined fifteen prior H3N2 influenza vaccine studies from 1997 to 2021, collectively containing 20,000 data points, and develop of a machine learning model that uses pre-vaccination HAI titres against multiple influenza variants to predict post-vaccination responses. To further test the model, four new vaccine studies were conducted in 2022-2023 spanning two geographic locations and three influenza vaccine types.
Findings: The most predictive pre-vaccination features were HAI titres against the vaccine strain and against historical influenza variants, with smaller predictive power derived from age, sex, vaccine dose, and geographic location. The resulting model predicted future responses even when the vaccine strain or vaccine formulation changed. A pre-vaccination feature-the time between peak HAI across recent variants-distinguished large versus small post-vaccination responses with 73% accuracy. Model predictions against prior vaccine studies had 2.4-fold error (95% CI: 2.34-2.40x, no large outliers with >4-fold error), yielding more accurate and robust predictions than a null model with 3.2-fold error (95% CI: 3.12-3.21x, 12% large outliers). The four new vaccine studies presented here were predicted with comparable accuracy to the intrinsic 2-fold error of the experimental assay.
Interpretation: A person's pre-vaccination influenza HAI titres using multiple variants are highly predictive of their post-vaccination response. Many individuals exhibited little-to-no vaccine response, as exhibited by the null model's accuracy, yet the machine learning model identified and accurately predicted both weak and strong responses with statistical superiority. Taken together, this approach paves the way to better utilise current influenza vaccines, especially for individuals that exhibit the weakest responses.
Funding: NIAID, UCSD PREPARE Institute, LJI & Kyowa Kirin, Inc. (KKNA-Kyowa Kirin North America), UGA, Cleveland Clinic, the Georgia Research Alliance, and the Bodman family.
Keywords: Antibody–virus interactions; Haemagglutination inhibition; Influenza; Machine learning; Prediction; Vaccine response.
. 2025 May 26:116:105744.
doi: 10.1016/j.ebiom.2025.105744. Online ahead of print. Leveraging pre-vaccination antibody titres across multiple influenza H3N2 variants to forecast the post-vaccination response
Hannah Stacey[SUP] 1 [/SUP], Michael A Carlock[SUP] 2 [/SUP], James D Allen[SUP] 2 [/SUP], Hannah B Hanley[SUP] 3 [/SUP], Shane Crotty[SUP] 4 [/SUP], Ted M Ross[SUP] 5 [/SUP], Tal Einav[SUP] 6 [/SUP]
Affiliations
- PMID: 40424667
- DOI: 10.1016/j.ebiom.2025.105744
Background: Despite decades of research on the influenza virus, we still lack a predictive understanding of how vaccination reshapes each person's antibody response, which impedes efforts to design better vaccines. Models using pre-vaccination antibody haemagglutination inhibition (HAI) titres against the vaccine strain alone poorly predict post-vaccination responses.
Methods: We combined fifteen prior H3N2 influenza vaccine studies from 1997 to 2021, collectively containing 20,000 data points, and develop of a machine learning model that uses pre-vaccination HAI titres against multiple influenza variants to predict post-vaccination responses. To further test the model, four new vaccine studies were conducted in 2022-2023 spanning two geographic locations and three influenza vaccine types.
Findings: The most predictive pre-vaccination features were HAI titres against the vaccine strain and against historical influenza variants, with smaller predictive power derived from age, sex, vaccine dose, and geographic location. The resulting model predicted future responses even when the vaccine strain or vaccine formulation changed. A pre-vaccination feature-the time between peak HAI across recent variants-distinguished large versus small post-vaccination responses with 73% accuracy. Model predictions against prior vaccine studies had 2.4-fold error (95% CI: 2.34-2.40x, no large outliers with >4-fold error), yielding more accurate and robust predictions than a null model with 3.2-fold error (95% CI: 3.12-3.21x, 12% large outliers). The four new vaccine studies presented here were predicted with comparable accuracy to the intrinsic 2-fold error of the experimental assay.
Interpretation: A person's pre-vaccination influenza HAI titres using multiple variants are highly predictive of their post-vaccination response. Many individuals exhibited little-to-no vaccine response, as exhibited by the null model's accuracy, yet the machine learning model identified and accurately predicted both weak and strong responses with statistical superiority. Taken together, this approach paves the way to better utilise current influenza vaccines, especially for individuals that exhibit the weakest responses.
Funding: NIAID, UCSD PREPARE Institute, LJI & Kyowa Kirin, Inc. (KKNA-Kyowa Kirin North America), UGA, Cleveland Clinic, the Georgia Research Alliance, and the Bodman family.
Keywords: Antibody–virus interactions; Haemagglutination inhibition; Influenza; Machine learning; Prediction; Vaccine response.