221 Development of abatacept- and adalimumab-specific predictive models of response to therapy in rheumatoid arthritis using data from a head-to-head study

Abstract

Background: Highly effective, targeted DMARD therapies with different mechanisms of action are available for RA. Translating precision medicine into clinical practice requires treatment-specific predictive models, with a goal of individualised, targeted therapy. Therefore, we created separate predictive models for response to abatacept or adalimumab, using baseline biomarker data from the head-to-head AMPLE study. Methods: Predictive models were built using demographic data, baseline disease characteristics and several biomarkers, including RF, cyclic citrullinated peptide-2 (CCP2) and biomarkers from the multibiomarker disease activity test as predictor variables and 'polar' clinical responses as the response variables. The polar responses were defined as patients who, after one year of treatment, achieved an ACR70 response or failed to achieve an ACR20 response. The elastic net method was used to build separate predictive models for abatacept and adalimumab responders using their respective clinical data with 6-fold cross-validation (CV) repeated 10 times. Parameter tuning for model selection was based on a fixed alpha of 0.95 and varying levels of lambda. The final model was selected based on the lambda with maximum mean area under the curve (AUC) across all folds of CV. Glmnet and caret packages in R were used for modelbuilding purposes. Results: Based on the polar response criteria, there were 69 responders and 45 non-responders for the abatacept arm, and 64 responders and 41 non-responders for the adalimumab arm. Predictive models generated included 13 variables for abatacept and 11 for adalimumab. Of all the variables, resistin, vascular cell adhesion molecule-1, sex, CCP2 and patient-reported disease activity were unique to the abatacept predictor, whereas matrix metalloproteinase-1, physician-reported disease activity and disease duration were unique to the adalimumab predictor. The variables common to both models showed the same association (positive or negative) but differing magnitude with response. AUC by receiver operating characteristic curves were used to assess the performance of the predictive models. The performance (AUC) of the abatacept model on the abatacept and adalimumab arms was 0.855 and 0.530, respectively. The performance (AUC) of the adalimumab model on the adalimumab and abatacept arms was 0.860 and 0.631, respectively. This indicates that the models are very specific for their respective treatment and not for the other treatment. Conclusion: Response-to-treatment predictive models were generated using baseline data from AMPLE that were highly specific to their respective treatment. This suggests that treatment-specific response predictors could be developed and should be considered, and highlights the value of head-to-head studies in predictive biomarker generation. Further testing on validation datasets is warranted.