Putting computational models of immunity to the test - an invited challenge to predict B. pertussis vaccination outcomes

Abstract

Systems vaccinology studies have been used to build computational models that predict individual vaccine responses and identify the factors contributing to differences in outcome. Comparing such models is challenging due to variability in study designs. To address this, we established a community resource to compare models predicting B. pertussis booster responses and generate experimental data for the explicit purpose of model evaluation. We here describe our second computational prediction challenge using this resource, where we benchmarked 49 algorithms from 53 scientists. We found that the most successful models stood out in their handling of nonlinearities, reducing large feature sets to representative subsets, and advanced data preprocessing. In contrast, we found that models adopted from literature that were developed to predict vaccine antibody responses in other settings performed poorly, reinforcing the need for purpose-built models. Overall, this demonstrates the value of purpose-generated datasets for rigorous and open model evaluations to identify features that improve the reliability and applicability of computational models in vaccine response prediction.

Figure 1.. Generation of multi-omics datasets for 117 study participants.

Contestants were provided with training datasets containing two cohorts (datasets 2020 and 2021), while the prediction dataset contained a newly generated cohort (dataset 2022). The training datasets contain pre-vaccination and post-vaccination immune response data, whereas the prediction dataset for 21 participants only contains pre-vaccination immune response data. Post-vaccination data was released after the challenge ended and used to evaluate submitted models.

Figure 2. Evaluation of the prediction models submitted for the invited CMI-PB challenge.

a) control models and models submitted by contestants b) models from systems vaccinology literature. Model evaluation was performed using Spearman’s rank correlation coefficient between predicted ranks by a contestant and actual rank for each of (1.1 and 1.2) antibody titers, (2.1 and 2.2) immune cell frequencies, and (3.1 and 3.2) transcriptomics tasks. The number denotes Spearman rank correlation coefficient, while crosses represent any correlations that are not significant using p ≥ 0.05.

Figure 3.. The method implemented by the winning team.

Schematic overview of the data processing, feature selection, and prediction modeling workflow. (a) The workflow begins with raw experimental data, including training and challenge datasets from plasma antibody levels, PBMC gene expression, PBMC cell frequency, and plasma cytokine concentration assays. The common features across these datasets are identified, followed by batch-effect correction and timepoint-wise imputation. (b) Feature selection was performed using various dimension reduction techniques, including LASSO, Ridge, PLS, PCA, and Multiple Co-inertia Analysis (MCIA). MCIA outperformed the other models and was selected for further analysis. MCIA integrates different data types (e.g., X1, X2, X3, X4) and their associated weights (A1, A2, A3, A4) to produce MCIA factors (G) that represent the combined data structure. (c) These MCIA factors were then used in a Linear Mixed Effects (LME) model to predict the outcome. The model was trained on 80% of the data (train set) using 5-fold cross-validation and evaluated on the remaining 20% (test set). The trained model was then applied to the challenge baseline data to generate predictions, which were used to rank subjects according to their predicted outcomes.