Modeling protective meningococcal antibody responses and factors influencing antibody persistence following vaccination with MenAfriVac using machine learning

Abstract

Meningococcal meningitis poses a significant public health burden in the meningitis belt region of sub-Saharan Africa. The introduction of the meningococcal PsA-TT vaccine (MenAfriVac®) has successfully eliminated Neisseria meningitidis serogroup A (NmA) cases in the region. However, the duration of post-vaccination immunity and the need for booster doses remain uncertain. To address this knowledge gap, we developed computational models using machine learning techniques to improve the effectiveness of modeling in guiding vaccination strategies for the African meningitis belt. Using serologic data from previous clinical trials of PsA-TT, we proposed a short-term and a long-term model that integrated demographic and medical variables (such as age, height and weight) with previous antibody titer levels and vaccination information to predict NmA antibody titer levels following vaccination. In the short-term model, we found moderately high performance (R-squared = 0.59) for out-of-training-data subjects and even better performance (R squared = 0.83) in the long-term evaluation. Our models estimated the half-life of the vaccine to be 13.9 years for the study population overall, similar to previously reported estimates. Machine learning techniques offer several advantages over previous approaches, as they do not require multiple readings from the same subject, can be rigorously validated using a subset of subject data not used for training. The proposed approach also facilitates the interpretation of the relationship between input variables and antibody levels at a population level. By incorporating subject-specific demographic and medical variables, our models could potentially be used to tailor vaccination schedules to at-risk populations.

Fig 1. a-b. Schematic representation of the long-term and short-term modeling, respectively.

a) Long-term modeling approach includes the baseline antibody level ${Ab}_0$ only and no further readings, b) Short-term modeling approach includes the baseline antibody level ${Ab}_{n-1}$ only and no prior readings.

Fig 2. a-b: Distribution of errors (residuals) on test set with long-term model and EBM.

a) Long-term model, b) Short-term model.

Fig 3. a-k: Sample trajectories of the predicted and true reference rSBA titer levels for different individual subjects.

The vertical lines represent the vaccination times and the stand-alone (diamond-shaped) points represent the initial baseline for the rSBA titer levels. The type of the vaccines given (no vaccine or a placebo was given in some cases, indicated by ‘none’) is mentioned under each subfigure. a) 1st: PsA-TT, 2nd: Hib-TT, b) 1st: PsA-TT, 2nd: 1/5th dose of PsACWY, c) 1st: PsA-TT, 2nd: 1/5th dose of PsACWY, d) 1st: PsA-TT, 2nd: none, 3rd: PsA-TT, e) 1st: none, 2nd: none, 3rd: PsA-TT, f) 1st: none, 2nd: none, 3rd: PsA-TT, g) 1st: none, 2nd: none, 3rd: PsA-TT, h) 1st: PsA-TT, 2nd: Hib-TT, 3rd: PsA-TT, i) 1st: PsA-TT, 2nd: Hib-TT, 3rd: PsA-TT, j) 1st: PsA-TT, 2nd: Hib-TT, 3rd: PsA-TT, k) 1st: PsA-TT, 2nd: Hib-TT, 3rd: PsA-TT.

Fig 4. a-b: Relative importance (as weights) of top 15 features for long-term and short-term model with EBM.

a) Long-term model, b) Short-term model.

Fig 5. Impact of the time* elapsed since first vaccination.

*We should note that certain ranges (78.6–129 and 483–736 days) did not have enough sample to obtain a reliable estimate of the impact, while the drop in immunity around 280 days, 430 days and 800 days were obvious as there are more samples around these timepoints.