ASSESSMENT OF CLINICAL FEATURES AS PREDICTORS OF INADEQUATE RESPONSE TO TNFI THERAPIES—FURTHER ANALYSIS OF THE MOLECULAR SIGNATURE RESPONSE CLASSIFIER

V. Strand, T. Mellors, A. Jones, L. Zhang, J. Withers, S. Asgarian, V. Akmaev, D. GhiassianStanford University, Division of Immunology/Rheumatology, Palo Alto, United States of America Scipher Medicine Corporation, Data Science, Waltham, United States of America Scipher Medicine Corporation, Medical, Clinical & Regulatory, Waltham, United States of America  Background Clinical and molecular features have been used to predict how RA patients will respond to distinct treatments. Clinical datasets are more readily available, more easily analyzed, and less costly to obtain than are molecular datasets . But predictions based on clinical features alone are usually insufficiently reliable for deciding among treatment options . Although molecular features are increasingly valuable for predicting treatment responses , physicians are often uncertain whether it is necessary to consider molecular measures . And health care providers must determine whether the benefits of using molecular datasets justify their added costs. The molecular signature response classifier (MSRC) uses clinical features and molecular (genomic and serologic) features to identify RA patients unlikely to adequately respond to TNFi therapies . But the relative contribution of clinical and molecular features to MSRC performance has yet to be fully addressed. This information is invaluable because most RA patients respond inadequately to their initial treatments and response-prediction classifiers offer much promise for improving patient outcomes. Objectives Our objective was to evaluate how well clinical features alone predict inadequate TNFi responses in the patient cohort used for developing the MSRC. Methods Clinical features (sex, body mass index, and patient global assessment) and molecular features (expression levels of 8 genes, 10 single nucleotide polymorphisms, anti-CCP, and CRP) for RA patients in the CERTAIN study were described . ACR50 at 6 months was the primary outcome defining response. The cohort was divided into a training set (n = 143) and a validation set (n = 175). Predictive models were trained through repeated cross-validation by using a random forest machine learning algorithm on 80% of the training data (n = 114); models were optimized by validation on the remaining 20% (n = 29). The final model was built on the entire training data set. Performance was then evaluated on the validation set (n = 175) and reported as area under the receiver operating characteristic curve (AUC). All statistical analyses were performed by using Python 3.7.6. Results Classifier performance when using only molecular features (AUC, 0.65–0.68) was better than its performance when using only clinical ones (AUC, 0.41–0.60) (Table 1). Classifier performance was best when using combined molecular and clinical features (AUC, 0.65–0.68). Conclusion These results confirm the importance of including molecular features in TNFi response–prediction models. Identifying new molecular biomarkers of treatment response can help address prediction challenges caused by the heterogeneity of RA. And considering clinical and molecular features together, with other assessments increases the predictive performance of precision medicine tests such as the MSRC. References Koo BS, et al. Arthritis Res Ther 2021, 23(1) 178. Cohen S, et al. Rheumatol Ther 2022, doi.org/10.1007/s40744-022-00506-0. Tao W, et al. Arthritis Rheumatol 2021, 73(2) 212-222. Guan Y, et al. Arthritis Rheumatol 2019, 71(12) 1987-1996. Mellors T, et al. Network and Systems Medicine 2020, 3.1 91-104. Image/graph: Acknowledgements: NIL. Disclosure of Interests Vibeke Strand Consultant of: Scipher Medicine Corporation, Ted Mellors Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation, Alex Jones Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation, Lixia Zhang Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation, Johanna Withers Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation, Sam Asgarian Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation, Viatcheslav Akmaev Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation, Dina Ghiassian Shareholder of: Scipher Medicine Corporation, Employee of: Scipher Medicine Corporation. Keywords: Biomarkers, Rheumatoid arthritis, Artificial intelligence DOI: 10.1136/annrheumdis-2023-eular.5874Citation: , volume 82, supplement 1, year 2023, page 485Session: Rheumatoid arthritis - prognosis, predictors and outcome (Poster View)