Machine learning prediction of pediatric adverse drug reactions using consensus-derived scarce data

Abstract

Adverse drug reactions (ADRs) represent a significant cause of morbidity and mortality in children, who face distinct pharmacological vulnerabilities due to unique physiological development. Current pediatric drug safety research is hindered by limited clinical data and adult-focused studies, creating evidence gaps. We developed a comprehensive computational approach for pediatric pharmacovigilance, integrating consensus-driven signal detection, multi-level biological features, and interpretable machine learning. Using 1.4 million FDA Adverse Event Reporting System reports, we constructed the largest curated pediatric drug-ADR dataset. Severity-specific thresholds and voting across four algorithms (PRR, ROR, BCPNN, and EBGM) optimized ADR identification. Multi-level biological fingerprints spanning molecular, target, and network domains combined with XGBoost significantly improved predictive performance (ROC AUC: 0.7177), especially for imbalanced scenarios. Cross-domain analyses revealed that models trained on adult data exhibit poor generalization to pediatric contexts, confirming that adverse reactions in children cannot be reliably predicted using adult data. Our approach successfully identified established and novel pediatric-specific ADRs with strong literature support. Collectively, this work establishes methodological innovations for pediatric pharmacovigilance, bridges a critical evidence gap in pediatric drug safety, and delivers practical tools for clinical and regulatory decision-making.

Adverse drug reactions (ADRs) are a severe healthcare concern worldwide, however, accurate prediction of pediatric-specific ADRs remains challenging. Here, the authors present a machine learning approach integrating consensus-driven signal detection and multi-level biological features, significantly enhancing pediatric ADR prediction.

Introduction

Adverse drug reactions (ADRs) are a leading cause of morbidity and mortality worldwide¹, imposing substantial clinical and economic burdens on the healthcare system^2,3. Among all patients, children represent a particularly vulnerable population in drug therapy due to their distinct physiological characteristics and developmental dynamics^4,5. Immaturity of organ systems and age-dependent expression of drug-metabolizing enzymes significantly influence the pharmacokinetics (PK) and pharmacodynamics (PD) of medications, leading to altered absorption, distribution, metabolism, and excretion compared to adults^6–8. Adverse drug events (ADEs) account for up to 10% of pediatric hospitalizations, nearly half of which are classified as life-threatening^9,10. The incidence of ADRs in children ranges from 0.6% to 16.8%, with approximately 3.9% being severe or fatal^11–13. Furthermore, up to 90% of hospitalized children receive off-label prescriptions^14,15, exposing them to unevaluated risks from unapproved indications or dosages^16,17. Despite these risks, ethical challenges, and recruitment barriers have limited pediatric clinical trials, resulting in a scarcity of high-quality safety data^8,18–21. Consequently, pediatric drug safety monitoring remains insufficient, emphasizing the urgent need for dedicated research, robust pharmacovigilance, and individualized treatment guidelines to reduce the elevated risk of ADRs in children.

Given the shortage of pediatric clinical data, post-marketing surveillance utilizing real-world databases has become essential²². Electronic health records (EHRs) and administrative claims data (ACD) provide valuable drug exposure histories but often lack standardized coding for adverse events, limiting their effectiveness for systematic ADR detection²³. In contrast, spontaneous reporting systems (SRS) like the FDA Adverse Event Reporting System (FAERS) effectively capture detailed adverse event information using standardized terminologies such as MedDRA, facilitating large-scale ADR evaluations^6,23. While FAERS has identified significant pediatric safety signals, such as severe anxiety related to selective serotonin reuptake inhibitors²⁴, analyses focusing on children are still limited, often addressing isolated drugs or specific ADRs without comprehensive evaluations²⁵. This fragmentation leaves many potential pediatric ADRs undiscovered, hindering clinical decision-making.

Machine learning models have demonstrated promising progress in ADR prediction and offer valuable tools to enhance pharmacovigilance^26,27. Notable work in this area includes Galeano et al.‘s approach for predicting not only the occurrence but also the frequencies of drug side effects using machine learning techniques²⁸, which advanced quantitative assessment of adverse event likelihoods. Farnoush et al. recently showed that incorporating demographic features like age and gender as additional inputs to prediction models can improve overall performance²⁹; however, like most existing approaches, they utilized these characteristics primarily as performance-enhancing features without analyzing fundamental physiological differences between populations or accounting for the unique developmental characteristics of pediatric patients³⁰. Recently, Gao et al. proposed a heterogeneous graph neural network framework, PreciseADR, which advances patient-level ADR prediction by integrating demographic factors such as age and sex³¹. While this represents a significant step toward individualized risk assessment by modeling complex relationships among patients, drugs, and adverse events, reliance on spontaneous reports for training labels (which often lack stringent signal validation) raises concerns about data quality and may adversely affect predictive reliability.³².

Our study presents several methodological innovations for pediatric pharmacovigilance that address two critical challenges: the scarcity of reliable adverse drug reaction data in children and the inability to accurately predict pediatric-specific ADRs. First, we established a high-quality pediatric ADR dataset by applying customized signal detection and voting strategies tailored to different ADR severity categories—using more sensitive thresholds for critical reactions such as FDA Black Box Warnings—across multiple algorithms (PRR, ROR, BCPNN, and EBGM), which effectively balanced sensitivity and specificity for serious and common ADRs. Second, we utilized multi-level biological fingerprints, providing complementary biological context beyond traditional chemical structures, which notably improved both the stability and overall predictive power of our models (ROC AUC: 0.7177, PRC AUC: 0.5456, and Accuracy: 0.7259), especially for challenging, imbalanced tasks. Third, our cross-domain analyses demonstrated that ADR prediction models trained on adult data could not be reliably transferred to pediatric settings, confirming the need for dedicated pediatric modeling approaches due to children’s distinct physiological development. Supported by extensive case studies and literature validation, our approach achieved robust and accurate identification of pediatric ADRs, demonstrating strong practical utility in enhancing pediatric drug safety assessment where clinical trial data is limited.

Results and discussion

FAERS data analysis and ADR dataset construction

Using the FDA Adverse Event Reporting System (FAERS), we constructed a comprehensive pediatric pharmacovigilance dataset spanning two decades (2004–2024). Initial extraction yielded 10,394,978 adverse event reports, from which 1,458,076 reports involved children aged 0-14 years. Following signal detection and rigorous entity filtering, we established a dataset containing 696,028 unique drug–ADR pairs involving 3070 distinct pharmaceutical compounds and 12,334 ADR terms (Fig. 1A). Due to significant data sparsity and imbalance, which present substantial modeling challenges, we applied stringent filtering criteria (see Methods) to select 2,363 clinically important drugs and 230 ADRs for model development. Age distribution analysis revealed distinct reporting patterns with three notable peaks: the highest report counts were observed for neonates under one month old (Month 0), followed by a sharp drop in months 1-11. A second moderate peak occurs at Year 1, with relatively stable reporting from years 2-10, followed by a progressive increase in the older pediatric age groups, with children aged 13-14 years showing the third and highest overall reporting frequency (Fig. 1B). This distribution highlights notable variability in adverse event reporting across pediatric age groups, potentially reflecting age-specific physiological vulnerabilities, differences in medication exposure, and developmental stage-dependent reporting practices. Examination of System Organ Class (SOC) data highlighted broad variation in reporting completeness and label distribution: for example, nervous system disorders encompassed 25 ADR endpoints, with 47,008 missing (empty) labels, 8339 positive and 3728 negative labels; general disorders included 25 endpoints with 36,471 missing, 1057 positive, and 4527 negative labels; respiratory disorders contained 23 endpoints with 42,722 missing, 8560 positive, and 3067 negative labels. As illustrated in Fig. 1C, these patterns reveal a highly imbalanced and long-tailed distribution of known positive and negative drug–ADR associations across SOC categories. Complementary to this, Fig. 1D shows that the top 40% of ADR endpoints account for nearly 60% of all drug–ADR associations, underscoring the concentration of signals within a relatively small subset of ADRs. Together, these observations highlight key challenges for effective modeling due to data sparsity, imbalance, and long-tail characteristics. Pyrexia emerged as the most prevalent ADR, followed by diarrhea, hypotension, and anemia, representing signature pediatric adverse reactions which often differ from adult profiles (Fig. 1D).

Fig. 1. Pediatric pharmacovigilance dataset construction and characterization from FAERS (2004–2024).

A Data processing workflow. B Age distribution of adverse event reports, showing peaks in neonates (<1 month) and children aged 1-2 years. C SOC distribution illustrating reporting completeness across categories (missing, positive, and negative labels); SOC terms are shown as abbreviations with full terminology available in Supplementary Table 1. D Word cloud of prevalent pediatric ADRs with corresponding cumulative distribution curve showing that 40% of ADR endpoints account for ~60% of all drug–ADR associations. E Comparison of unique ADR endpoints across pediatric (0–14 years, n = 11,411), adolescent (14–18 years, n = 8317), and adult (SIDER database, n = 5728) populations, with 3881 shared endpoints between children and adolescents and only 440 overlapping with adults. F Comparison of unique drug–ADR pairs across pediatric (213,602), adolescent (151,381), and adult (132,442) populations, showing limited overlap between children and adolescents (8496 pairs) and between pediatric and adult populations (6415 pairs).

Comparative analysis among pediatric (0–14 years), adolescent (14–18 years), and adult (SIDER database³³) populations revealed marked age-specific differences. Pediatric patients exhibited 11,411 unique ADR endpoints, exceeding adolescents (8317) and adults (5728). Overlap analysis showed 3881 shared endpoints between children and adolescents, yet only 440 endpoints overlapped with adults (Fig. 1E). Drug–ADR pair comparisons showed an even stronger stratification: 213,602 unique pediatric pairs compared to 151,381 adolescent and 132,442 adult pairs. The child-adolescent overlap was 8496 pairs, whereas pediatric–adult overlap was limited to 6415 pairs (Fig. 1F). This pronounced population stratification underlines the developmental specificity of drug safety profiles and challenges the common practice of extrapolating adult pharmacovigilance data to pediatric contexts.

Signal detection analysis

To systematically identify genuine drug–ADR associations from spontaneous reporting data, we employed four complementary disproportionality analysis methods: Reporting Odds Ratio (ROR)³⁴, Proportional Reporting Ratio (PRR)³⁵, Empirical Bayesian Geometric Mean (EBGM)³⁶, and Bayesian Confidence Propagation Neural Network (BCPNN)³⁷. Figure 2 illustrates the distribution characteristics of these methods when applied to our pediatric FAERS dataset. All four approaches demonstrated clear discrimination between positive signals (red) and negative/non-signals (blue), with threshold values (ROR = 1, PRR = 1, EBGM = 2, and IC = 0) effectively separating the distributions. Notably, FDA Black Box Warning ADRs (circled points) exhibited significantly higher signal scores across all methods, validating the sensitivity of our signal detection framework for capturing clinically critical safety concerns. For regular ADRs, we implemented a minimum case count threshold (n ≥3) to reduce false positives from spurious associations³⁸, while for Black Box Warning signals, we intentionally removed this reporting frequency constraint to maximize detection sensitivity for these critical safety concerns³⁹. This stratified approach significantly enhanced our ability to detect severe ADRs, as evidenced by the numerous Black Box Warning signals captured in the low-frequency region (reporting count = 1) of the ROR (Fig. 2A) and PRR (Fig. 2B) distributions that would have been missed under uniform thresholding rules.

Fig. 2. Disproportionality analysis methods for pediatric adverse drug reaction signal detection.

Scatter plots showing the distribution of signal scores versus reporting frequency for four complementary methods: (A) ROR, (B) PRR, (C) EBGM, and (D) BCPNN. In each panel, positive signals are shown in red and negative/non-signals in blue, with threshold lines (ROR = 1, PRR = 1, EBGM = 2, IC = 0) separating the distributions. Circled points represent FDA Black Box Warning ADRs, which consistently demonstrate higher signal scores across all methods. Regular ADRs were subjected to a minimum case count threshold (n ≥ 3), while this constraint was removed for Black Box Warning signals to maximize detection of critical safety concerns. All metrics are displayed using logarithmic transformations (log₁₀ for ROR, PRR, and EBGM; IC value for BCPNN) to facilitate visualization of the distribution patterns.

The logarithmic transformations applied to each method (log₁₀(ROR), log₁₀(EBGM), log₁₀(PRR), and IC value for BCPNN) revealed comparable overall distribution patterns, suggesting methodological robustness, yet with subtle differences in signal strength quantification. BCPNN (Fig. 2D) demonstrated particularly distinctive behavior through its implementation of severity-adapted signal thresholds: a more conservative threshold (IC-2SD ≥3) for regular ADRs to minimize false positives, and a more sensitive threshold (IC-2SD > 0) for Black Box Warning ADRs to ensure comprehensive detection of severe reactions. This adaptive thresholding strategy optimized the trade-off between sensitivity and specificity across the severity spectrum of pediatric adverse events. Conversely, ROR exhibited the widest dynamic range but with greater variability in the low-frequency region, potentially indicating higher sensitivity but lower specificity for rare ADRs. EBGM (Fig. 2C) and PRR (Fig. 2B) distributions occupied intermediate positions between these extremes. We observed that a substantial majority of Black Box Warning signals were consistently identified by all four methods, while only approximately half of regular ADR signals achieved such consensus, highlighting the increased detection challenge for less severe but potentially still clinically relevant ADRs.

Model performance evaluation

We conducted a comprehensive evaluation of multiple molecular descriptor types in combination with XGBoost classification for pediatric ADR prediction. As illustrated in Fig. 3A, biological fingerprints (BioFeat) demonstrated superior performance characteristics compared to Morgan fingerprints, RDKit descriptors, and hybrid approaches. When analyzing ROC AUC versus sample size distributions, BioFeat consistently achieved higher ROC AUC values across diverse endpoint tasks, particularly maintaining robust performance even for endpoints with limited samples (200–400 drug–ADR pairs). More importantly, BioFeat showed marked improvement in precision-recall metrics for highly imbalanced datasets (balance ratio >0.5), where other descriptors typically faltered. This advantage is critical for pediatric pharmacovigilance applications where class imbalance is pervasive.

Fig. 3. Performance evaluation of molecular representation for pediatric ADR prediction.

A ROC AUC versus sample size for different molecular representations (BioFeat⁴⁰, RDKit+Morgan, RDKit, and Morgan) using XGBoost. Points colored by balance ratio (darker red = greater imbalance). BioFeat shows superior performance, especially with limited samples. B System Organ Class stratification showing ROC AUC (blue), PRC AUC (orange), and Accuracy (green). Performance varies across physiological systems, with highest ROC AUC in vascular and skin disorders and lowest in endocrine and pregnancy conditions.

The performance stratification by SOC revealed clear variability in model accuracy across physiological systems (Fig. 3B). Our XGBoost+BioFeat model achieved the highest mean ROC AUC in vascular (0.76), skin (0.75), neoplasms (0.75), psychiatric (0.75), and eye disorders (0.75). In contrast, lower mean ROC AUCs were observed in congenital (0.65), pregnancy (0.63), and especially endocrine disorders (0.60), all of which were characterized by few endpoints and limited labeled data (e.g., endocrine: 1 endpoint, only 246 negative and 110 positive labels, with 2,007 missing). As illustrated in Fig. 1C, SOC categories with more endpoints and richer labeled datasets, such as nervous, general, and gastrointestinal disorders, tended to exhibit more stable and robust predictive performance, as reflected by lower standard deviation in ROC AUC and related metrics. Conversely, sparsely represented categories with fewer endpoints or a high proportion of missing labels showed reduced accuracy and greater variance. These findings underscore the critical influence of both endpoint diversity and label availability on the reliability of ADR prediction, highlighting the need for more well-annotated data for rare pediatric safety events.

Table 1 presents a comprehensive comparison of our XGBoost+BioFeat model against alternative machine learning approaches. Our model achieved the highest discriminative performance as measured by ROC AUC (0.7177) and PRC AUC (0.5456), demonstrating superior ability to distinguish between adverse and non-adverse outcomes. While some models exhibited marginally better performance in specific metrics, such as TabPFN+BioFeat (Fine-tuned) for accuracy (0.7408) and Logistic Regression+BioFeat for F1 score (0.4751), the XGBoost+BioFeat model maintained strong and balanced performance across all evaluation criteria. The ablated version of our model showed slightly improved MCC (0.2683 vs 0.2651), suggesting potential for further optimization. Notably, across all descriptor types tested (BioFeat, RDKit2D, and Morgan fingerprints), the biologically-informed features consistently enabled better model performance, highlighting the value of domain-specific molecular representations for pediatric pharmacovigilance applications. In contrast, more complex deep learning models like D-MPNN demonstrated poor performance, with a low ROC AUC (0.6192) and PRC AUC (0.4041), indicating that complex models are unsuitable for this challenging dataset characterized by small and imbalanced data.

Table 1.

Performance comparison of machine learning models and descriptor types for pediatric ADR prediction

Model	ROC AUC	PRC AUC	Accuracy	F1 Score	MCC
Ours	0.7177	0.5456	0.7259	0.4261	0.2651
XGBoost+BioFeat (Ablated)	0.7130	0.5455	0.7275	0.4306	0.2683
XGBoost+RDKit2D	0.6671	0.4935	0.6930	0.4115	0.2087
XGBoost+Morgan	0.6511	0.4738	0.6571	0.4348	0.2012
SVM+BioFeat	0.6621	0.4783	0.7345	0.1043	0.1180
SVM+RDKit2D	0.6584	0.4798	0.7370	0.1790	0.1363
SVM+Morgan	0.6069	0.4303	0.7326	0.1067	0.1103
Logistic Regression+BioFeat	0.6840	0.5101	0.6755	0.4751	0.2553
Logistic Regression+Morgan	0.6398	0.4635	0.6766	0.4134	0.1929
Logistic Regression+RDKit2D	0.6170	0.4334	0.6179	0.4289	0.1689
TabPFN+BioFeat (Fine-tuned)	0.6772	0.5054	0.7408	0.2096	0.1636
TabPFN+RDKit2D	0.6065	0.4329	0.7192	0.1944	0.1017
TabPFN+Morgan	0.5980	0.4173	0.7248	0.0812	0.0537
MLP+BioFeat	0.6743	0.4985	0.7178	0.3704	0.2099
MLP+Morgan	0.6281	0.4541	0.6958	0.2979	0.1479
MLP+RDKit2D	0.6128	0.4308	0.7224	0.2285	0.1311
D-MPNN	0.6192	0.4041	0.7171	0.1269	0.1022

The best performance for each metric is highlighted in bold.

Cross-domain model generalization analysis

To evaluate the transferability of ADR prediction models across age domains, we conducted a comprehensive cross-domain analysis comparing four distinct training-testing paradigms: Child-to-Child, Child-to-Adult, Adult-to-Adult, and Adult-to-Child. Figure 4 illustrates the dramatic performance disparities between in-domain and cross-domain predictions. In-domain models (Child-to-Child and Adult-to-Adult) demonstrated robust predictive capacity, with median ROC AUC values of 0.7005 and 0.6913, respectively. However, cross-domain applications exhibited substantial performance degradation, with Child-to-Adult models achieving a median ROC AUC of 0.4463 and Adult-to-Child models performing similarly poorly at 0.4315. These differences were highly statistically significant (p < 0.001) across all cross-domain comparisons.

Fig. 4. Transferability assessment of ADR prediction models across age domains.

Probability density distributions of (A) ROC AUC and (B) PRC AUC for four training-testing paradigms: Child-to-Child (orange), Child-to-Adult (blue), Adult-to-Adult (yellow), and Adult-to-Child (purple). In-domain models (Child-to-Child, Adult-to-Adult) demonstrated robust predictive capacity, while cross-domain applications exhibited significant performance degradation (p < 0.001). Results indicate pediatric and adult ADR mechanisms represent distinct pharmacological domains that cannot be reliably bridged through simple model transfer.

The precision-recall metrics revealed a corresponding pattern, with in-domain Child-to-Child and Adult-to-Adult models achieving median PRC AUC values of 0.5104 and 0.5858, respectively, while cross-domain models decreased markedly to 0.2200 for Child-to-Adult and 0.2607 for Adult-to-Child predictions. Notably, although precision-recall performance is challenged by class imbalance in the pediatric dataset, the Child-to-Child median PRC AUC remains reasonably high, indicating maintained precision-recall capability despite the imbalance.

These findings provide compelling quantitative evidence that pediatric and adult ADR mechanisms represent fundamentally distinct pharmacological domains that cannot be reliably bridged through simple model transfer. The profound performance collapse observed in cross-domain applications underscores the critical importance of developing dedicated age-specific pharmacovigilance models rather than extrapolating from adult data.

This age-domain specificity likely reflects the complex interplay of developmental pharmacokinetics, immature organ systems, and pediatric-specific drug metabolism pathways that substantially alter the drug-safety relationship landscape in children compared to adults. Our results thus provide powerful empirical support for regulatory imperatives mandating pediatric-specific drug safety evaluations.

Model ablation analysis

To address potential data leakage concerns and evaluate the independence of our prediction framework, we conducted a targeted ablation study focusing on the E3 signature from the Chemical Checker (CC) framework⁴⁰. Within the hierarchical organization of CC signatures, E3 specifically encodes adverse effect information of small-molecule drugs derived from the SIDER database, representing one of the five clinical-level (E-level) bioactivity spaces in the 25-dimensional signature framework. Given that E3 signatures were trained on adult adverse event data, this ablation analysis was critical to establish whether our pediatric models were inadvertently leveraging adult ADR patterns rather than learning pediatric-specific relationships. Figure 5A presents a comprehensive comparison between complete models and those with E3 signatures ablated across all ADR endpoints.

Fig. 5. Model ablation analysis and feature importance evaluation.

A Performance comparison between complete models and those with E3 signatures ablated across all ADR endpoints, showing high stability around the no-change diagonal. Points are colored by evaluation metric. B Heatmap showing endpoint-specific sensitivity to E3 ablation, with differential impacts across various ADRs. C Left: SHAP value distributions across hierarchical biological fingerprint spaces (A: Molecular, B: Target, C: Biological Network, D: Cellular, E: Clinical) for neonatal jaundice prediction; Right: Pie chart showing relative contribution of feature levels to model predictions, with clinical-level fingerprints (28.0%) providing the highest predictive value, followed by network (21.4%), cellular (19.5%), target (17.4%), and molecular (13.6%) levels.

The results demonstrate remarkable performance stability despite E3 signature removal, with the majority of endpoints clustering tightly around the no-change diagonal. Statistical analysis using paired t tests confirmed no significant difference in overall model performance across all evaluation metrics. Specifically, for ROC AUC, the paired t test yielded a t statistic of 1.7009 and a p-value of 0.0903 (mean difference: 0.0047; Cohen’s d: 0.0685). For PRC AUC, the t statistic was 0.0382 with a p-value of 0.9696 (mean difference: 0.0001; Cohen’s d: 0.0012). For accuracy, the t statistic was -0.4066 with a p-value of 0.6847 (mean difference: −0.0016; Cohen’s d: −0.0177). In all cases, the p-values are well above the threshold for statistical significance, definitively confirming that our models do not rely on potentially leaked adverse effect patterns from SIDER, but instead capture independent structure-ADR relationships specific to pediatric patients.

Endpoint-specific analysis revealed differential sensitivity to E3 ablation (Fig. 5B), with cardiovascular endpoints showing modest performance reductions (muscle spasms: ΔPRC AUC = -0.20), while certain endpoints exhibited slight improvements with E3 removal. These variations likely reflect the complex and often dissimilar manifestation of ADRs between pediatric and adult populations rather than data leakage effects. The minimal impact of removing E3 signatures aligns with our earlier overlap analysis demonstrating substantial differences between adult SIDER data and pediatric ADR profiles, providing dual confirmation that our pediatric models capture age-specific safety signals rather than recapitulating adult adverse effect patterns.

Model interpretability analysis

Analysis of feature importance for neonatal jaundice prediction using SHAP (SHapley Additive exPlanations)⁴¹ revealed distinctive contributions across the hierarchical biological fingerprint spaces. Clinical-level fingerprints emerged as the dominant contributors, accounting for 28.0% of the total feature impact. The broad distribution of both positive and negative SHAP values at this level suggests that specific clinical phenotypes may either predispose to or protect against jaundice development. Network-level fingerprints ranked second (21.4%), underscoring the role of pathway-level mechanisms, particularly those involved in bilirubin metabolism and hepatic processing.

Cellular-level and target-related fingerprints contributed 19.5% and 17.4% respectively, while molecular descriptors exhibited the lowest contribution at 13.6%. This hierarchical pattern suggests that higher-level biological representations capture more predictive information for neonatal jaundice than structural properties alone, reflecting the multifactorial complexity of drug-induced hyperbilirubinemia in neonates⁴².

Symmetric SHAP value distributions around zero across all fingerprint levels indicate that features within each biological space can both increase and decrease jaundice risk. For example, clinical-level fingerprints likely capture known associations with hyperbilirubinemia, while target-level fingerprints may include interactions with albumin or bilirubin-processing enzymes such as UGTs. These bidirectional effects highlight the mechanistic complexity whereby drug influences on protein binding, hepatic clearance, and erythrocyte stability jointly modulate bilirubin levels through multiple biological pathways⁴³.

Neonatal adverse drug reactions and maternal medication exposure

Neonates represent a uniquely vulnerable population, with immature physiological systems and heightened susceptibility to pharmacological interventions⁵. Pediatric drug administration inherently presents multiple challenges: limited dosing guidelines, sparse pharmacokinetic data, ethical constraints on clinical trials, and significant inter-individual variability in drug metabolism^22,44. Moreover, maternal medication exposure introduces additional uncertainties, potentially subjecting infants to unforeseen pharmacological risks through transplacental transfer and breast milk transmission⁴⁵. Our model successfully identified a substantial number of neonatal adverse drug reactions, with all predicted associations substantiated by comprehensive literature evidence, demonstrating the model’s robust predictive performance and computational validation capabilities (Table 2).

Table 2.

Clinically significant neonatal adverse drug reactions identified by the model with supporting literature evidence

Index	Drug/drug class	Adverse drug reaction		Evidence
1	Hydroxyamphetamine; Levomethadone; Benzyl alcohol Antidepressant	Neonatal Disorder	49,51,86,87
2	Chloramphenicol	Neonatal Jaundice, Respiratory Distress Syndrome	46
3	Antidepressant; Opioids Lithium Citrate	Respiratory Disorder Neonatal	52,53,56,58,88–91
4	Brimonidine	Respiratory Depression	92
5	Antithyroid Drugs	Hypothyroidism	93
6	Lidocaine	Cyanosis, Hypoxia, Oxygen Saturation Decreased	55
7	NSAID; ACEI; ARB	Acute Kidney Injury	94,95
8	Tazobactam/Piperacillin	Hemoglobin Decreased	57
9	Opioids	Deafness	96
10	Aminoglycosides	Acute Kidney Injury	97
11	Valproate	Congenital Anomaly	48
12	Amphotericin B	Electrolyte Imbalance	98
13	Ardeparin	Pyrexia	99
14	Carbamazepine	Hepatic Enzyme Increased	47

Chloramphenicol emerged as a prominent historical case of neonatal drug toxicity correctly identified by our model. This antibiotic is notorious for causing “Gray Baby Syndrome” in neonates, characterized by progressive cyanosis, vascular collapse, and often death, due to the immature hepatic glucuronidation pathways that lead to drug accumulation. Additionally, our model predicted neonatal disorders associated with benzyl alcohol, a common pharmaceutical preservative linked to the “gasping syndrome” in premature infants, where undeveloped metabolic pathways result in toxic benzoic acid accumulation, metabolic acidosis, and neurological complications⁴⁶. The successful identification of these well-documented adverse effects validates our model’s capacity to detect established pharmacological risks in the neonatal population.

Our analysis revealed specific safety concerns with medications commonly used in pediatric neurological conditions. Carbamazepine was specifically identified with hepatic damage and cholestasis, while anticonvulsants as a drug class were associated with bleeding complications⁴⁷. Valproate showed associations with congenital gingival hyperplasia⁴⁸. Notably, our model identified the potential link between anticonvulsant medications and autism spectrum disorders, corroborating emerging clinical evidence⁴⁹. In the realm of antimicrobial therapy, aminoglycoside antibiotics including gentamicin and tobramycin demonstrated associations with nephrotoxicity, highlighting important considerations for neonatal antimicrobial stewardship.

Maternal mental health presents another critical domain of pharmacological concern. With depression rates significantly elevated during pregnancy, psychotropic medication use has become increasingly prevalent. The opioid epidemic in the United States has further complicated maternal medication exposure, introducing complex pharmacological challenges for neonatal health⁵⁰. While opioids such as buprenorphine and methadone are often associated with medication-assisted treatment for substance use disorders, they demonstrate significant associations with neonatal respiratory disorders and abstinence syndromes^51,52. Concurrently, some of antidepressants reveal intricate connections to neonatal respiratory, neurological, and developmental complications⁵³. These findings underscore the profound neurobiological vulnerabilities of neonates to maternal pharmacological exposures, highlighting the critical need for comprehensive risk assessment and targeted intervention strategies⁵⁴.

Beyond the most prominent findings, our model uncovered several less frequently recognized but clinically significant adverse drug reactions. Lidocaine’s potential for methemoglobinemia⁵⁵, lithium citrate’s respiratory distress associations⁵⁶, and tazobactam/piperacillin’s hemolytic effects represent nuanced interactions that, while less common, warrant careful clinical attention⁵⁷. Particularly noteworthy are drugs like natural and synthetic opioids, which demonstrate significant potential for long-term neurodevelopmental consequences when exposed during critical developmental periods⁵⁸.

Collectively, these findings illustrate the model’s capacity to integrate multi-level biological data effectively, enabling comprehensive identification of both classical and emerging neonatal ADRs.

Analysis of pediatric and adult ADRs reveals developmental pharmacological disparities

The model successfully predicted numerous potential adverse drug reactions (ADRs) in pediatric populations. A comprehensive drug–ADR network was constructed to visualize these predicted results and compare pediatric and adult ADR profiles (Fig. 6). Importantly, all pediatric ADRs identified in the network have supporting evidence in the literature, as detailed in Supplementary Table 2. The network illustrates both ADRs common to children and adults, as well as those uniquely observed in pediatric patients. These findings emphasize profound disparities in medication responses between pediatric and adult patients, fundamentally rooted in the developmental complexity of children’s physiological systems^5,54,59.

Fig. 6. Pediatric–adult drug–ADR network.

Network visualization of drugs and predicted adverse drug reactions (ADRs) in pediatric versus adult populations. Drug nodes are colored by ATC category; ADR nodes are shaped by type (black-box or regular). Solid edges denote known associations; dashed edges indicate model predictions, with blue/red edges showing adult/pediatric relations. The network highlights metabolic immaturity, developmental neuroplasticity, and organ-specific vulnerabilities as key factors in pediatric drug safety, all supported by literature evidence (Supplementary Table 2).

The analysis identifies three critical factors distinguishing pediatric drug safety from adult pharmacology: metabolic immaturity, developmental neuroplasticity, and organ-specific vulnerability. Children’s hepatic enzyme systems and renal clearance mechanisms operate with fundamentally different efficiencies compared to adults, leading to altered drug metabolism and excretion⁵⁹. For instance, drugs like theophylline and fluoxetine can cause ADRs such as neonatal jaundice⁶⁰, brain edema⁶¹, and respiratory depression⁶², which reflect the immature metabolic pathways that make children more susceptible to toxic effects.

Developmental neuroplasticity introduces additional complexity to pediatric pharmacology⁵⁴. Nervous system drugs, including aripiprazole and fluoxetine, demonstrate markedly different ADR profiles in children. These profiles are characterized by heightened sensitivity to muscular rigidity⁶³, cognitive disruptions⁶⁴, and neurological disturbances⁶⁵. Critically, these differences extend beyond simple dosage variations, stemming from the unique neural network development occurring during childhood. Pharmacological interventions during this period can significantly influence neurodevelopmental trajectories, creating a dynamic and sensitive physiological environment⁵⁴.

Organ-specific vulnerabilities further underscore the pediatric–adult pharmacological divide. Pediatric cardiovascular and respiratory systems are rapidly maturing, presenting distinct safety challenges compared to adult physiological systems⁶⁶. Drugs like losartan and pravastatin exemplify these differences, with ADR spectra ranging from cardiovascular changes in adults to metabolic disturbances in neonates^67,68. These variations reflect the intricate developmental sensitivities unique to children’s physiological systems.

The adult-pediatric ADR network covers widely used drugs across multiple ATC categories, revealing that some commonly prescribed medications carry unexpected risks in children—highlighting the potential of computational models to advance pediatric drug safety monitoring⁶⁹.

Conclusion

Our study presents the first comprehensive pediatric pharmacovigilance framework that combines consensus-driven signal detection for dataset construction with machine learning modeling based on hierarchical biological fingerprint features. By curating a large pediatric-specific ADR dataset from FAERS through a robust voting mechanism and systematically integrating multi-level features, we provide a practical reference for future research on drug safety in vulnerable populations. Models incorporating clinical, molecular, and network-level information offered benefits over those using traditional chemical descriptors alone, potentially providing insights into the pharmacodynamic and pharmacokinetic characteristics of children.

Our results reveal substantial differences between pediatric and adult ADR profiles, emphasizing that drug safety in children requires independent evaluation and cannot simply be extrapolated from adult data^5,70. This gap reflects the unique vulnerabilities of children, whose developing physiological systems and age-dependent drug metabolism create distinct susceptibility patterns^54,59,66. Through our computational framework, we successfully captured historically significant pediatric ADRs, including the chloramphenicol-induced gray baby syndrome and other well-documented cases unique to pediatric populations. Our case studies also highlight specific instances where developmental differences may contribute to variations in ADR risk between children and adults. These findings underscore the critical importance of dedicated pediatric pharmacovigilance systems.

It should be noted that, despite applying stringent data preprocessing, advanced signal detection, and a voting-based approach to enhance data quality, the use of spontaneous reporting data still entails inherent limitations, including under-reporting, reporting biases, and considerable clinical heterogeneity, issues which these measures can only partly address^71,72. Furthermore, defining the pediatric population as ≤14 years inevitably covers a range of developmental stages; finer age stratification and integration of longitudinal clinical data will be helpful to improve model specificity and applicability. Leveraging these multilevel bioactivity fingerprints, our model can generate predictions for previously unseen small-molecule drugs. Predictive performance is most reliable for compounds within the model’s applicability domain, as defined by the training distribution in these bioactivity spaces. Predictions, especially for newly approved pediatric medicines, should therefore be interpreted as safety alerts or prioritization signals for follow-up assessment rather than confirmatory evidence.

However, several limitations should be noted. First, the scope of biological fingerprint features depends on the completeness of current bioactivity databases, which may not fully capture pediatric-specific mechanisms not yet profiled in adult or preclinical assays. Second, because the training data are derived from U.S. FAERS reports, the model may not fully capture regional differences in drug use and reporting practices. Application to non-U.S. pediatric populations should thus be approached with caution, and adaptation with local data is recommended when available.

In summary, our work provides a new, generalizable approach for pediatric ADR prediction and sets a valuable benchmark for pharmacovigilance research in special populations. By clarifying key differences between children and adults and establishing a feasible computational workflow, this work contributes to advancing safer medication use in vulnerable pediatric populations and enhancing regulatory risk assessment.

Materials and methods

Data Extraction and Preprocessing

We extracted ADR reports related to pediatric population (defined as individuals aged 14 years or younger, consistent with common clinical practice definitions^73–76) from the FAERS database covering 2004 to 2024. Initial data preprocessing involved deduplication based on unique report identifiers, mapping of ADR events to MedDRA Preferred Terms (version 26.1)⁷⁷, and standardization of drug names via DrugBank⁷⁸. To improve data quality, reports with clear inconsistencies, such as pediatric cases with physiologically implausible recorded body weights, were carefully removed, as these likely represented data entry errors or misclassified adult cases. Additionally, non-drug-related adverse events, including surgical complications and medical device malfunctions, were excluded to focus the analysis on potential drug-induced reactions.

Pharmacovigilance Signal detection and labeling

To systematically identify genuine drug–ADR associations while minimizing both false positives and false negatives, we implemented a comprehensive signal detection framework that balances statistical sensitivity with clinical relevance. Our approach strategically categorized ADRs into FDA Black Box Warning ADRs and regular ADRs, a critical distinction that allows for tailored detection thresholds based on clinical severity. In the United States, Black Box Warnings represent the FDA’s most serious safety alerts, visually distinguished by a black border surrounding the text on prescription drug labels, and indicate potentially severe or life-threatening risks⁷⁹. This categorization enabled us to address alert fatigue by implementing differential detection strategies: increasing sensitivity for critical Black Box ADRs to ensure capture of rare but serious signals, while enhancing specificity for regular ADRs to minimize false positives.

Four established disproportionality analysis methods were selected based on their complementary statistical foundations: the frequentist approaches of Reporting Odds Ratio (ROR) and Proportional Reporting Ratio (PRR), which offer computational simplicity, higher sensitivity, and straightforward clinical interpretability but may be more susceptible to false positives with sparse data; Empirical Bayesian Geometric Mean (EBGM) for its shrinkage estimation that stabilizes signals in sparse data matrices and reduces false positives⁸⁰; and Bayesian Confidence Propagation Neural Network (BCPNN) for its information-theoretic foundation that more effectively manages confounding in complex reporting databases⁸¹. While the frequentist methods excel at signal detection sensitivity, the Bayesian methods provide enhanced specificity through their incorporation of prior information, creating a balanced framework that maximizes the identification of genuine safety signals while minimizing spurious associations^37,82. The ROR was calculated as

$$\begin{array}{c}\hfill \mathrm{ROR}=\frac{\mathrm{a}/\mathrm{c}}{\mathrm{b}/\mathrm{d}}=\frac{\mathrm{ad}}{\mathrm{bc}}\hfill \end{array}$$ 1

where $\mathrm{a}$ is the count of reports containing both the drug and ADR, $\mathrm{b}$ is the count of reports with the drug but not the ADR, $\mathrm{c}$ is the count of reports with the ADR but not the drug, and $\mathrm{d}$ is the count of reports with neither the drug nor the ADR. The 95% confidence interval (CI) was computed based on the standard error of $\ln \left(\mathrm{ROR}\right)$:

$$\begin{array}{c}\hfill \mathrm{SE}\left(\ln \mathrm{ROR}\right)=\sqrt{\frac{1}{\mathrm{a}}+\frac{1}{\mathrm{b}}+\frac{1}{\mathrm{c}}+\frac{1}{\mathrm{d}}}\hfill \end{array}$$ 2

$$\begin{array}{c}\hfill 95\%\mathrm{CI}=\exp \left(\ln \left(\mathrm{ROR}\right)\pm 1.96\times \mathrm{SE}\left(\ln \mathrm{ROR}\right)\right)\hfill \end{array}$$ 3

Similarly, the PRR is defined as $\mathrm{PRR}=\frac{\mathrm{a}/\left(\mathrm{a}+\mathrm{b}\right)}{\mathrm{c}/\left(\mathrm{c}+\mathrm{d}\right)}$ with its 95% CI derived from the standard error of $\mathrm{SE}\left(\ln \mathrm{PRR}\right)$:$\mathrm{PRR}=\frac{\frac{\mathrm{a}}{\mathrm{a}+\mathrm{b}}}{\frac{\mathrm{c}}{\mathrm{c}+\mathrm{d}}}$

$$\begin{array}{c}\hfill \mathrm{SE}\left(\ln \mathrm{PRR}\right)=\sqrt{\frac{1}{\mathrm{a}}-\frac{1}{\mathrm{a}+\mathrm{b}}+\frac{1}{\mathrm{c}}-\frac{1}{\mathrm{c}+\mathrm{d}}}\hfill \end{array}$$ 4

$$\begin{array}{c}\hfill 95\%\mathrm{CI}=\exp \left(\ln \left(\mathrm{PRR}\right)\pm 1.96\times \mathrm{SE}\left(\ln \mathrm{PRR}\right)\right)\hfill \end{array}$$ 5

The BCPNN method estimates the Information Component (IC) as

$$\mathrm{IC}={\log }_2\frac{P\left(x,y\right)}{P\left(x\right)P\left(y\right)}={\log }_2\frac{a\left(a+b+c+d\right)}{\left(a+b\right)\left(a+c\right)}$$ 6

where $P\left(x,y\right)$ is the joint probability of the drug–ADR pair, and $P\left(x\right)$, $P\left(\mathrm{y}\right)$ are the marginal probabilities. The variance $\mathrm{V}\left(\mathrm{IC}\right)$ is calculated as

$$\begin{array}{c}\hfill \text{V}\left(\text{IC}\right)={\left(\frac{1}{\ln 2}\right)}^2\left[\frac{\mathrm{N}-\mathrm{a}+\mathrm{\gamma }-{\mathrm{\gamma }}_{11}}{\left(\mathrm{a}+{\mathrm{\gamma }}_{11}\right)\left(1+\mathrm{N}+\mathrm{\gamma }\right)}+\frac{\mathrm{N}-\mathrm{a}-\mathrm{b}+\mathrm{\alpha }-{\mathrm{\alpha }}_1}{\left(\mathrm{a}+\mathrm{b}+{\mathrm{\alpha }}_1\right)\left(1+\mathrm{N}+\mathrm{\alpha }\right)}+\frac{\mathrm{N}-\mathrm{a}-\mathrm{c}+\mathrm{\beta }-{\mathrm{\beta }}_1}{\left(\mathrm{a}+\mathrm{c}+{\mathrm{\beta }}_1\right)\left(1+\mathrm{N}+\mathrm{\beta }\right)}\right]\hfill \end{array}$$ 7

with $\mathrm{N}=\mathrm{a}+\mathrm{b}+\mathrm{c}+\mathrm{d}$. The parameter $\mathrm{\gamma }$ is calculated as

$$\begin{array}{c}\hfill \mathrm{\gamma }={\mathrm{\gamma }}_{11}\frac{\left(\mathrm{N}+\mathrm{\alpha }\right)\left(\mathrm{N}+\mathrm{\beta }\right)}{\left(\mathrm{a}+\mathrm{b}+{\mathrm{\alpha }}_1\right)\left(\mathrm{a}+\mathrm{c}+{\mathrm{\beta }}_1\right)}\hfill \end{array}$$ 8

where ${\mathrm{\alpha }}_1={\mathrm{\beta }}_1=1$, $\mathrm{\alpha }=\mathrm{\beta }=2$, ${\mathrm{\gamma }}_{11}=1$, and ${\mathrm{\gamma }}_2=1$ are prior parameters⁸¹. To establish a reliable criterion for signal detection that accounts for statistical uncertainty, we calculated the lower confidence bound of IC (IC-2SD) as

$$\begin{array}{c}\hfill \text{IC}-2\text{SD}=\mathrm{E}\left(\mathrm{IC}\right)-2\times \sqrt{\mathrm{V}\left(\mathrm{IC}\right)}\hfill \end{array}$$ 9

where $\mathrm{E}\left(\mathrm{IC}\right)$ represents the expected value of IC and $\sqrt{\mathrm{V}\left(\mathrm{IC}\right)}$ is its standard deviation. This value (IC-2SD) serves as a conservative approximation of the lower 95% confidence interval for the information component, widely adopted in pharmacovigilance practice.

Specifically, the EBGM is calculated as

$$\begin{array}{c}\hfill \mathrm{EBGM}=\frac{\mathrm{a}\left(\mathrm{a}+\mathrm{b}+\mathrm{c}+\mathrm{d}\right)}{\left(\mathrm{a}+\mathrm{c}\right)\left(\mathrm{a}+\mathrm{b}\right)}\hfill \end{array}$$ 10

$$\begin{array}{c}\hfill 95\%\mathrm{CI}=\exp \left(\ln \left(\mathrm{EBGM}\right)\pm 1.96\times \sqrt{\frac{1}{\mathrm{a}}+\frac{1}{\mathrm{b}}+\frac{1}{\mathrm{c}}+\frac{1}{\mathrm{d}}}\right)\hfill \end{array}$$ 11

For FDA Black Box Warning ADRs, a high-sensitivity detection strategy was employed by removing the minimum case count thresholds for ROR and PRR, aiming to capture rare but critical signals. The BCPNN method applied a relaxed cutoff (IC − 2 SD > 0), while EBGM signals were considered positive when the lower bound of the 95% CI exceeded 2. In contrast, regular ADRs were subject to stricter criteria: conventional ROR and PRR thresholds requiring a minimum of three reports, a more conservative BCPNN cutoff (IC − 2 SD ≥ 3), and consistent EBGM thresholds⁸³. Final ADR labels were assigned based on a majority voting scheme, whereby a drug–ADR pair was labeled positive only if at least two of the four methods identified it as a signal.

Data preprocessing and endpoint selection

Due to the inherent sparsity and imbalance in adverse drug reaction (ADR) data across drugs, rigorous filtering procedures were implemented to ensure the quality and usability of the dataset for downstream modeling. Initially, the processed drug–ADR association data, derived from the signal detection phase, was transformed into a wide-format matrix, where rows represented unique drugs and columns represented distinct ADR endpoints. Matrix entries indicated the presence (1) or absence (0) of a confirmed signal for each drug–ADR pair. To mitigate the impact of sparse ADR occurrences, endpoints were excluded based on missing data thresholds set differentially for FDA Black Box Warning ADRs and regular ADRs: Black Box ADRs with >85% missing values and regular ADRs with >80% missing values were removed. This step helped to eliminate ADRs with insufficient reporting that could introduce noise and instability into the predictive models. Further, to address class imbalance within retained endpoints, a Balance Ratio metric was introduced for each ADR:

$$\begin{array}{c}\hfill \text{Balance Ratio}=\frac{∣{\mathrm{N}}_1-{\mathrm{N}}_0∣}{{\mathrm{N}}_1+{\mathrm{N}}_0}\hfill \end{array}$$ 12

where ${\mathrm{N}}_1$ and ${\mathrm{N}}_0$ denote the number of positive (ADR occurrence) and negative (no ADR) samples respectively for each ADR endpoint. ADRs excluded during the initial filtering but exhibiting relatively balanced distributions (Balance Ratio <0.3) and missingness less than 90% were reconsidered and retained. This two-tiered filtering strategy preserved potentially meaningful yet less frequent ADR signals while maintaining data quality. Canonical SMILES strings for each drug were retrieved from ChEMBL⁸⁴ and standardized through salt removal and tautomer normalization using the RDKit. Biologics, mixtures, and other complex molecules lacking valid canonical SMILES were excluded to ensure reliable structural descriptor generation.

Biological fingerprint feature engineering

We utilized the 25-dimensional biological fingerprint descriptors (BioFeat) developed by Bertoni et al.⁴⁰, which provide a comprehensive and multi-scale bioactivity representation of small molecules across five hierarchical levels: molecular, target, biological network, cellular, and clinical domains. This set of descriptors is constructed to capture richer pharmacological characteristics by integrating a wide range of bioactivity data curated in the Chemical Checker resource than traditional chemical descriptors alone.

Each of the five biological levels comprises five specific sub-features, including but not limited to:

• Molecular Level: 2048-bit Morgan fingerprints (radius=2) for 2D structure, E3FP conformational fingerprints capturing 3D conformations, Murcko scaffolds highlighting core molecular frameworks, MACCS keys representing common substructures, and physicochemical properties such as molecular weight, logP, and hydrogen bond donor/acceptor counts.

• Target Level: Pharmacological mechanisms, metabolism-related genes, protein structural annotations, binding affinity data from ChEMBL and BindingDB, alongside high-throughput screening assays from PubChem.

• Biological Network Level: Functional ontologies, endogenous metabolic pathways, molecular-target interaction pathways, gene ontology biological processes, and protein-protein interaction networks.

• Cellular Level: Transcriptional responses across cell lines, cytotoxicity data (GI50) from cancer panels, chemical genetic screens in yeast mutants, drug-induced morphological changes, and cell-based growth assays.

• Clinical Level: ATC classifications, drug indication information, adverse effect profiles, disease phenotype associations, and drug-drug interaction datasets.

These descriptors were predicted and completed using Siamese neural networks trained to infer missing bioactivity signatures, producing consistent 128-dimensional embeddings per feature. The Chemical Checker framework has been shown to effectively extend bioactivity annotations to poorly characterized compounds, enhancing descriptive power beyond traditional chemical fingerprints. We incorporated these biological fingerprints into our modeling framework because of their ability to capture rich biological context and improve predictive performance compared to conventional structural descriptors. For benchmarking purposes, standard cheminformatics descriptors, including RDKit 2D descriptors and Morgan fingerprints, were also computed using RDKit.

Machine learning model development and optimization

To address the high sparsity and class imbalance in ADR datasets for special populations, we developed and evaluated a comprehensive machine learning framework encompassing multiple algorithms: XGBoost, Support Vector Machines (SVM), Multilayer Perceptron (MLP), Logistic Regression, and the Transformer-based Tabular Prior-data Fitted Network (TabPFN) model. TabPFN leverages pre-training on synthetic data to capture complex feature interactions with minimal tuning. These models were selected for their ability to handle small-to-medium sized, high-dimensional tabular datasets with varying balances of interpretability and predictive power⁸⁵.

Hyperparameter optimization was conducted using Optuna, with stratified 10-fold cross-validation to ensure robust model performance. BioFeat served as the primary input features, with comparative benchmarking performed using classical cheminformatics descriptors like Morgan fingerprints and RDKit 2D descriptors to validate the predictive value of our biological fingerprints.

Model performance assessment and interpretability

Model performance was quantitatively evaluated on the independent test set using a suite of standard metrics to comprehensively characterize predictive accuracy and reliability, particularly under class imbalance:

• Area under the receiver operating characteristic curve (ROC AUC) measures the ability of the model to distinguish positive and negative classes across varying thresholds; mathematically, it represents the probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative one.

• Area Under the Precision-Recall Curve (PRC AUC) summarizes the precision–recall trade-off and is particularly sensitive to performance on minority positive classes. Precision $\mathrm{P}$ and recall $\mathrm{R}$ are given by

$$\begin{array}{c}\hfill \mathrm{P}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}},\mathrm{R}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}\hfill \end{array}$$ 13

where $\mathrm{TP}$, $\mathrm{FP}$, $\mathrm{FN}$ are true positives, false positives, and false negatives respectively.

• Accuracy evaluates the overall correctness of classification:

$$\begin{array}{c}\hfill \text{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}\hfill \end{array}$$ 14

• F1 Score is the harmonic mean of precision and recall, balancing their trade-off:

$$\begin{array}{c}\hfill \text{F}1=2\times \frac{\mathrm{P}\times \mathrm{R}}{\mathrm{P}+\mathrm{R}}\hfill \end{array}$$ 15

• Matthews Correlation Coefficient (MCC) yields a balanced measure of classification quality by accounting for all confusion matrix components:

$$\begin{array}{c}\hfill \text{MCC}=\frac{\mathrm{TP}\times \mathrm{TN}-\mathrm{FP}\times \mathrm{FN}}{\sqrt{\left(\mathrm{TP}+\mathrm{FP}\right)\left(\mathrm{TP}+\mathrm{FN}\right)\left(\mathrm{TN}+\mathrm{FP}\right)\left(\mathrm{TN}+\mathrm{FN}\right)}}\hfill \end{array}$$ 16

Together, these metrics provide a rigorous and multifaceted evaluation across sensitivity, specificity, precision, recall, and overall predictive power, with ROC AUC and PRC AUC serving as primary indicators for classification discrimination and robustness in imbalanced datasets.

Independent test set construction and representativeness

For each ADR endpoint, we identified drugs with non-missing labels and performed a stratified holdout at the drug level (90% training set, 10% independent test set; fixed random seed = 42), preserving the endpoint’s positive/negative ratio to ensure test set representativeness. All hyperparameter tuning and model selection were conducted via stratified 10-fold cross-validation within the 90% training partition only; the independent test set was held out and used exactly once for final evaluation. For descriptor pipelines requiring normalization, scaling parameters were fit on training folds and applied to validation/test data to prevent leakage.

To gain insights into the model’s decision-making process and identify key biological factors driving ADR risk predictions, SHapley Additive exPlanations (SHAP) analysis was conducted on tree-based models such as XGBoost⁴¹. SHAP values quantify each feature’s contribution to individual predictions by averaging marginal contributions over all possible feature subsets, enabling insight into model decision-making processes.

Cross-domain model generalization analysis

To rigorously assess differences in ADR prediction between pediatric and adult populations, we conducted a cross-domain generalization analysis. Herein, we constructed two datasets: a pediatric (“Child”) dataset constructed as described above using FAERS data, and an adult (“Adult”) dataset was derived from the publicly available SIDER database. The Adult dataset was processed and standardized following the same pipeline used for the Child dataset, including harmonization of ADR endpoints to MedDRA 26.1 terms, mapping drugs to DrugBank identifiers, and retrieval of molecular SMILES from ChEMBL.

Both datasets were featurized using our established BioFeat, and subsequently modeled with XGBoost employing stratified 10-fold cross-validation and hyperparameter optimization as described.

We defined four training-testing paradigms to evaluate predictive performance within and across age domains: Child-to-Child (training and testing on pediatric data), Child-to-Adult (training on pediatric data, testing on adult data), Adult-to-Adult (training and testing on adult data), and Adult-to-Child (training on adult data, testing on pediatric data). To prevent data leakage, drugs and ADR endpoints common to both datasets were carefully excluded from training sets in cross-domain tasks. Model performance across these scenarios was assessed using ROC AUC, PRC AUC, and accuracy metrics, following the statistical methods described previously, allowing comprehensive characterization of predictive behavior across age domains.

E3 signature ablation study

Because E3 signatures encode adverse effect information derived from adult populations and could inadvertently introduce data leakage into pediatric models, we conducted a targeted ablation study by removing E3 features from the biological fingerprint prior to model training. XGBoost models were then trained and evaluated independently for each ADR endpoint using both the complete BioFeat set and the E3-ablated feature set. The same modeling pipeline (stratified 10-fold CV, hyperparameter tuning via Optuna) was applied to ensure fair comparison.

Model performance was evaluated by ROC AUC, PRC AUC, and accuracy. Paired two-sided t-tests were performed to compare metrics between complete and E3-ablated models for each endpoint. The paired t-statistic was computed as

$$\begin{array}{c}\hfill \mathrm{t}=\overline{d}/\frac{{\mathrm{s}}_{\mathrm{d}}}{\sqrt{\mathrm{n}}}\hfill \end{array}$$ 17

where $\overline{\mathrm{d}}$ is the mean difference of paired metric values, ${\mathrm{s}}_{\mathrm{d}}$ is the standard deviation of the differences, and $\mathrm{n}$ is the number of endpoints.

Effect size was quantified by Cohen’s d:

$$\mathrm{d}=\frac{\overline{\mathrm{d}}}{{\mathrm{s}}_{\mathrm{pooled}}},{\mathrm{s}}_{\mathrm{pooled}}=\sqrt{\frac{{\mathrm{s}}_1^2+{\mathrm{s}}_2^2}{2}}$$ 18

where ${\mathrm{s}}_1^2$ and ${\mathrm{s}}_2^2$ are the variances of the metric in complete and ablated models, respectively. A significance threshold of $\mathrm{p}<0.05$ was applied.

Author contributions

Y.T. conceived the study, designed the methodology, developed the computational models, performed data analysis, wrote the manuscript, and conducted the validation of predicted ADRs. J.Y. contributed to data preprocessing, manuscript revision, and provided constructive feedback. K.L. assisted with data processing, visualization, and algorithm implementation. J.P. helped with data processing, contributed to study conception, and performed literature validation. Y.D. helped with statistical analysis. D.J. provided supervision and critical revision of the manuscript. D.C. provided supervision, conceptual guidance, and comprehensive manuscript revision. Y.T. performed the cross-domain model analysis, ablation studies, and case studies. D.C. and D.J. secured funding and provided research resources. All authors reviewed and approved the final version of the manuscript.

Peer review

Peer review information

Communications Chemistry thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The raw adverse event reports were obtained from the public FDA Adverse Event Reporting System (FAERS) database, available at https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html. The adverse event ontology data were accessed through the Medical Dictionary for Regulatory Activities (MedDRA) at https://www.meddra.org/. Note that MedDRA data requires a subscription for access. The processed datasets and generated features used in this study are available in the GitHub repository at https://github.com/TY-CSU/Predicting-Adverse-Drug-Reactions-in-Children. All other relevant data supporting the findings of this study are available from the corresponding author upon reasonable request.

Code availability

The source code for this study is available from the GitHub repository at https://github.com/TY-CSU/Predicting-Adverse-Drug-Reactions-in-Children.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42004-025-01865-9.