Background: Methotrexate (MTX) is the anchor drug in newly diagnosed rheumatoid arthritis (RA). Nevertheless, MTX treatment response remains highly variable, and a non-negligible number of patients fail to reach the treatment target, requiring other or additional therapy. Therefore, prediction of treatment response could help in early identification of MTX non-responders who should be offered other therapies up front.

Objectives: Using machine learning models, we assessed whether persistence to treatment with MTX in early RA could be predicted using data on demographics, clinical variables, medical history and genome-wide genetic variants, and measured the impact of genetic data on prediction quality.

Methods: Incident RA patients starting MTX as first-line treatment in DMARD monotherapy were identified from the Swedish Rheumatology Quality register (SRQ) and the Epidemiological Investigation of RA (EIRA) study. Data on demographics, clinical presentation and medical history five years prior to RA diagnosis were obtained via register linkages, with additional data on common genetic variants obtained through genotyping. Genetic data was used to create polygenic risk scores (PRSs), for traits thought to have an effect on persistence to MTX, and SNP-based principal components. Choice of traits for the former was based on the associations observed in a previous study from our group [1], whereas the latter corresponded to the first six principal components from a principal component analysis carried out in the set of SNPs remaining after quality control processing. Two regression models (logistic and elastic net) and two tree-based ensemble models (random forests and XGboost) were trained and assessed for their ability to predict persistence to treatment at one and three years, using nested cross-validation.

Results: We identified a cohort of 3128 RA patients initiating treatment with methotrexate as first-line treatment in DMARD monotherapy, of which 2034 (67%) were persistent (i.e., remaining on MTX in DMARD monotherapy) at one year, and 1335 (46%) were persistent at three years. Using data on a total of 99 variables on demographics, clinical variables and medical history, the highest area under the receiver-operating-characteristic curve (AUROC) was 0.62 (95%CI, 0.61-0.63) for predicting persistence at one year, and 0.62 (0.61-0.63) for predicting persistence at three years, both using elastic net. The addition of genetic data did not improve predictive performance, instead leading to a slight decrease in AUROC upon including data on PRSs and principal components, both for prediction of persistence at one (0.59, 0.58-0.60) and three years (0.61, 0.60-0.63). Of the four models, logistic regression generally performed worst.

Conclusion: The ability to predict persistence to treatment with MTX was weak overall. The addition of genetic data did not improve prediction quality neither in the form of polygenic risk scores for potentially associated traits, nor as SNP-based principal components of composite genetic ancestry.

REFERENCES: [1] Tidblad, L., et al., Comorbidities and chance of remission in patients with early rheumatoid arthritis receiving methotrexate as first-line therapy: a Swedish observational nationwide study. RMD Open, 2023. 9 (4).

Acknowledgements: The authors would like to thank the participants of the EIRA study, the SRQ biobank and STR, as well as deCODE genetics for making this study possible. The authors would also like to thank Jesper Gådin and Robert Karlsson for their help with the imputation of the data.

Disclosure of Interests: None declared.