High-dimension Mechanistic Model Building using LASSO Approaches : Application to Ebola Vaccination
Auriane Gabaut (1), Mélanie Prague (1)
(1) Université de Bordeaux, Inria, Inserm, Bordeaux Population Health Research Center, SISTM Team ; Vaccine Research Institute, Créteil, France
Introduction
Constructing non-linear mixed-effects models (NLMEM) can significantly enhance our understanding of biological processes. In the population approach, NLMEM combine fixed effects, which represent population-level relationships with explanatory covariates, and random effects, which describe remaining heterogeneity between individuals. The estimation of NLMEM typically employs maximum likelihood methods, such as the Stochastic Approximation Expectation-Maximization (SAEM) algorithm [1], which are computationally demanding. Specifically, when selecting covariates to define individual-level parameters, it is impractical to compare all possible models. This issue is compounded by the presence of high-dimensional explanatory covariate data, such as that found in -omics research.
Method
To optimize the construction of NLMEM, traditional methods often use modified stepwise approaches [2,3]. An alternative, the Stochastic Approximation for Model Building Algorithm (SAMBA), constructs the covariate model by iteratively evaluating the posterior parameter distributions [4]. SAMBA is efficient and quick within low-dimensional contexts, aiming to minimize an information criterion. However, adapting SAMBA to high-dimensional data is essential. SAMBA initially uses a stepwise AIC algorithm for selecting covariates. Our proposed method substitutes this with a lasso selection technique. We also integrate a stability selection algorithm to address the lasso estimator's instability issues and prevent the inclusion of non-significant covariates [5]. This new algorithm also minimizes the chosen information criterion, here corrected BIC, and will enhance the control over the false positive discoveries. Implementation in R connected with Monolix is available.
Results
We conducted simulations to replicate the humoral immune response to an Ebola vaccine, linking it with transcriptomics data from initial time points [6]. These simulations were performed 100 times with 100 individuals each, using parameters for humoral response from the EBOVAC trial [7]. We simulated high-dimensional transcriptomic data mimicking the Prevac-UP trial data [8] and related some with mechanistic model parameters. Then, we compared the SAMBA and the lasso-SAMBA algorithm for building the covariate model in these simulations. Our method notably reduced false discovery rates, maintaining zero false negatives. The false discovery rate decreased from 30-40% to 0%. The F1-score of our method exceeded 98%, outperforming the original method's median F1-score of 75-80%.
Our methodology was then applied to an immunological sub study nested in the Prevac/Prevac-UP trial data, which compares two approved Ebola vaccines in Africa. We concentrated on 30 adults who received the two-dose heterologous vaccine Ad26.ZEBOV/MVA-BN-Filo regimen. Following the vaccine's second dose, we modeled antibody levels from 7 days post second dose and analyzed its link with the baseline genetic expression of over 28,000 genes. Our strategy was to select predictive genes for the humoral immune response parameters, from the 1000 most variable genes measured at baseline. We identified several genes associated with the humoral response, including some distinctly recognized within the immune response pathway.
Conclusion
The findings show improved control of the False Positive Rate, affirming our method's effectiveness in correctly identifying pertinent covariates and enhancing model accuracy in high-dimensional settings. This advancement is particularly pertinent given the growing access to high-dimensional data, with significant implications for pharmacogenomics.
References
[1] Estelle Kuhn and Marc Lavielle. Coupling a stochastic approximation version of EM with an MCMC procedure. ESAIM : Probability and Statistics, 8 :115–131, 2004
[2] Svensson, R. J. and Jonsson, E. N. (2022). Efficient and relevant stepwise covariate model building for pharmacometrics. CPT: Pharmacometrics & Systems Pharmacology, 11(9):1210–1222
[3] Ayral, G., Si Abdallah, J.-F., Magnard, C., and Chauvin, J. (2021). A novel method based on unbiased correlations tests for covariate selection in nonlinear mixed effects models: The cossac approach. CPT: Pharmacometrics & Systems Pharmacology, 10(4):318–329.
[4] Prague, M. and Lavielle, M. (2022). Samba: A novel method for fast automatic model building in nonlinear mixed-effects models. CPT: Pharmacometrics & Systems Pharmacology, 11(2):161–172.
[5] Meinshausen, N. and Bühlmann, P. (2010). Stability selection. Journal of the Royal Statistical Society, Series B, 72:417–473
[6] Pasin, C., Balelli, I., Van Effelterre, T., Bockstal, V., Solforosi, L., Prague, M., Douoguih, M., and Thiébaut, R. (2019). Dynamics of the Humoral Immune Response to a Prime-Boost Ebola Vaccine: Quantification and Sources of Variation. Journal of Virology, 93(18):e00579–19
[7] Alexandre, M., Prague, M., McLean, C., Bockstal, V., Douoguih, M., and Thiébaut, R. (2023). Evaluation and prediction of the long-term humoral immune response induved by the two-dose heterologous Ad26.ZEBOC,MVA-BN-Filo vaccine regimen against ebola. Nature Vaccine 8 (1) 174
[8] Badio, M., and all. (2021). Partnership for research on ebola vaccination (prevac): protocol of a randomized, double-blind, placebo-controlled phase 2 clinical trial evaluating three vaccine strategies against ebola in healthy volunteers in four west african countries. Trials, 22