(BSPOT31) Real-world Research on Drug Safety Based on Quantitative Structure-Activity Relationship: A Case Study on Pediatric Drug-Induced Immune Thrombocytopenia

Background: Drug-induced immune thrombocytopenia (DITP) is a frequently overlooked ADR in children. Failure to timely identification of DITP cases can lead to severe bleeding complications and even death. Traditional signal detection methods require relatively time-consuming population-based signal validation. Quantitative structure-activity relationship (QSAR) models evaluated drug toxicity based on drug physicochemical properties and molecular fingerprints, offering a promising tool for ADR signal identification and post-market verification.

Objectives: To predict the DITP toxicity of drugs based on QSAR model and to assess the accuracy of signals which discovered by disproportionality and regression methods in the real world.

Methods: We first used EMR data from Beijing Children's Hospital from 2009 to 2020 and proposed a two-stage (disproportionality and regression) method to construct the DITP signal mining algorithm. Subsequently, we summarized information from the published drug instructions and literature to build a DITP benchmark database (DITPst) which contains the high-risk drugs for DITP and their molecular descriptors. Finally, based on the DITPst database, we used the XGBoost algorithm and drug molecular descriptors to construct QSAR models to predict potential DITP toxicity. We applied this constructed model to assess the accuracy of new pediatric DITP signals discovered in real-world data previously.

Results: We found griseofulvin and cefoperazone/sulbactam are two new potential DITP positive signals in both children and adults based on EMR data, as well as positive signals in sex drugs (imipenem, telavancin, fosfomycin, cefoperazone, cefotetan, and cefpirome) previously implicated in DITP only in adults. The DITPst database containing 1,319 drugs was constructed by integrating data from multiple sources. Based on this database, the QSAR model constructed using the XGBoost algorithm and ECFP4 descriptors for drugs performed well (training set: AUC of 0.88, sensitivity of 0.95; validation set: AUC of 0.75, sensitivity of 0.89). Finally, we validated the model performance in the 8 drugs with positive DITP detected by EMR data above and successfully identified 6 (75%) among them.

Conclusions: We developed a two-stage DITP safety signal detection algorithm and identified 8 potential new signals in pediatrics. Combining the QSAR model with real-world signal mining methods can further improve the methodology for evaluating the clinical safety of pediatric drug use and drug surveillance systems.