Identifying Pediatric and Developmental Screening Targets With Biologically Informed Molecular Representations

More than 75% of Food and Drug Administration approved drugs lack pediatric dosing, efficacy, or safety data, and roughly 40% of medications prescribed to children are used off-label. For most therapies children receive, dose, formulation, or indication has never been formally evaluated in anyone under 18. In this work, we build on our previous model, BioMADE, to identify toxicity screening targets whose relevance is specific to defined developmental stages. We analyzed 3.8 million drug-condition records from KidSIDES across 7 developmental stages and 6 system organ classes. Univariate analysis identified biologically coherent genes associated with pediatric conditions, including known genes like GABRB1 (pathogenic variants cause developmental and epileptic encephalopathy), SLC6A1 (associated with developmental encephalopathies), and GRIN2A (defining gene of epilepsy-aphasia spectrum). Signals were the strongest in early life, spanning multiple organ systems. The highest univariate model was gene CTSS, area under the receiver operating characteristic curve of 0.776; 95% confidence interval 0.62-0.90. Multivariate models surpassed 0.8 area under receiving operator curve in 13 datasets (0.81-0.91), indicating that multi-gene drug interaction profiles carry predictive signal beyond single-target associations. We further identify secondary targets that may influence pediatric development and contribute to pediatric-specific adverse drug events, many of which are not included in current safety screening panels. Our findings support the development of novel pediatric-specific safety screening panels, with a particular emphasis on respiratory and congenital disorders in early life. Prioritizing relevant drug targets for pediatric screening can potentially reduce the challenges of pediatric clinical trials and advance safer, more precise pediatric drug therapies.