OA03 Serum proteome profiling reveals candidate biomarkers for paediatric chronic nonbacterial osteomyelitis

Background/Aims Chronic non-bacterial osteomyelitis (CNO) is an autoinflammatory bone disease characterised by pain, bone deformity and/or fractures. The pathophysiology of CNO is complex and may vary between patients. Timely diagnosis is paramount to allow effective treatment and prevent complications. However, in the absence of diagnostic criteria or biomarkers, CNO remains a diagnosis of exclusion. Aim: To identify serum proteins separating CNO from key differential diagnoses. Methods Sera from children with CNO, infectious osteomyelitis, malignancies, CNO-associated Crohn’s disease and healthy controls were collected towards independent discovery (N = 30/group) and validation (N = 10/group) cohorts. Proteins were quantified using a quadrupole time-of-flight mass spectrometer (timsTOF_HT, BrukerDaltonics) set to data-independent acquisition (DIA). Proteins detected in > 80% samples were retained, normalised and remaining missing values imputed (four methods). Discovery data were split into train and test sets. Univariate filtration was performed on the train set using CNO-specific differential abundance and remaining proteins used as input to Least Absolute Shrinkage and Selection Operator (LASSO) regression for 100 models/fold. Candidates were chosen based on consistent selection across imputation methods, univariate thresholds, folds and LASSO models. Random forest models were built to assess predictive performance on test data. Candidates were assessed for univariate test significance and multivariate prediction power in the independent validation dataset. Results Normalised abundance of 479 proteins was taken for selection (discovery cohort). Multivariate feature selection across a grid-search of filtration parameters and imputation methods delivered a panel of 30 proteins (Table 1) with excellent predictive capacity for CNO (specificity=1, sensitivity=0.97, Matthew’s correlation coefficient/MCC=0.96) across 10-fold data splits. Performance on test data was comparable. Twenty-one of 30 proteins were detected in the validation cohort with comparable fold-changes. Predictive performance for CNO across validation samples was remarkable (specificity=1, sensitivity=0.67, MCC=0.79). Individual CNO samples were mis-predicted as healthy (N = 2) or Crohn disease (N = 1). Conclusion Protein signatures can distinguish paediatric CNO from differential diagnoses. Misclassification of one CNO patient as having Crohn’s disease may be explained by gastrointestinal inflammation in a subset of CNO patients. Because CNO presents with bone pain and pathological imaging, high specificity (100%) alongside moderate sensitivity (67%) promise significant potential for this panel in clinical practice. Disclosure E. McDonnell: None. E.K. Roberts: None. J. Blair: None. J. Cassat: None. C. Eyers: None. H.J. Girschick: None. J. Hawkes: None. R. Knoefler: None. M.W. Laass: None. C. Leuck: None. R. Maher: None. L. Mee: None. H. Morbach: None. E. Caamano Gutierrez: None. C. Hedrich: Other; receives unrestricted funding from Merck (MISP) to study lupus nephritis.