A study of 216 children and adolescents with acute lymphoblastic leukemia (ALL) in Lahore, Pakistan, has identified a genetic pattern that may help predict which young patients develop heart damage from the anthracycline chemotherapy drugs that cure their cancer. The research, published in the Egyptian Journal of Medical Human Genetics, combined careful statistical modeling of five gene variants with polygenic risk scoring and machine learning, and its most striking result is methodological: a properly weighted genetic score discriminated cases with an area under the curve (AUC) of 0.802, while a naive unweighted score performed no better than a coin flip at 0.498.
Anthracyclines such as doxorubicin are a mainstay of ALL induction therapy, but they carry a well-known cardiac cost. The drugs intercalate into DNA and inhibit topoisomerase II in rapidly dividing cells, and in cardiomyocytes they also disrupt the mitochondrial electron transport chain, generating reactive oxygen species that damage membranes, mitochondrial DNA and contractile proteins. The resulting injury can drive cardiomyocyte death, fibrosis and a progressive decline in left ventricular function. In this cohort, cardiotoxicity was defined by American Society of Echocardiography criteria as a drop in left ventricular ejection fraction of more than 10 percentage points to below 53 percent, or clinical cardiac abnormalities after induction. Seventy-nine of the 216 patients (36.57 percent) met these criteria, while 137 did not.
The researchers genotyped five variants in genes involved in drug metabolism and transport: SOD2 rs4880, CBR1 rs9024, PNPLA3 rs738409, ABCC1 rs4148350 and ABCG2 rs2231142. Because five variants tested across four inheritance models generate many comparisons, the team applied Bonferroni correction, with thresholds of p < 0.0025 at the model level and p < 0.005 at the genotype level. After correction, ABCG2 rs2231142 stood alone as statistically robust. Its heterozygous AC genotype was strongly enriched among patients who developed cardiotoxicity, appearing in 26.6 percent of the affected group versus 3.6 percent of the unaffected group, while the homozygous AA genotype was far more common in those spared (52.6 percent versus 12.7 percent).
That non-additive pattern is biologically puzzling. Under a codominant model, which fit best by the Akaike Information Criterion (AIC = 240.7), the AC genotype carried an odds ratio of 5.25 (95 percent CI: 1.84 to 14.95; p < 0.001). Yet in multivariable logistic regression the minor A allele of ABCG2 appeared protective, with an odds ratio of 0.45 in the fully adjusted model. The authors note this contradicts the conventional view that rs2231142, the Q141K variant, reduces transporter activity and should therefore raise toxicity by impairing clearance of anthracycline metabolites. They suggest the C allele may tag a distinct functional haplotype in this population, or that linkage patterns differ from those in predominantly European cohorts, and they call for population-specific functional studies.
CBR1 rs9024 told a more internally consistent story. The AA genotype was enriched in the cardiotoxicity group (60.8 percent versus 40.1 percent), the recessive model fit best (AIC = 279.1; p = 0.003), and the variant retained significance in multivariable regression, with an odds ratio of 1.70 (95 percent CI: 1.19 to 2.44; p = 0.004) in the fully adjusted model. This aligns with prior evidence that the A allele increases CBR1 expression and activity, enhancing conversion of doxorubicin into the cardiotoxic metabolite doxorubicinol, and with the same group’s earlier work on the UKALL 2003 protocol. However, the genotype-level association did not survive Bonferroni correction (corrected p = 0.09), so the authors frame it as a well-supported candidate requiring prospective validation. SOD2, PNPLA3 and ABCC1 showed no significant associations in any analysis.
The polygenic scoring comparison produced the study’s clearest lesson. An unweighted additive score, simply summing minor allele counts across all five variants, yielded an AUC of 0.498, essentially chance. A beta-weighted score that coded each variant according to its best-fitting inheritance model and weighted it by its logistic regression coefficient, reversing the sign for the protective ABCG2 allele, achieved an AUC of 0.802. Stratifying patients into tertiles of the weighted score revealed a significant dose-response gradient: cardiotoxicity rates rose from 11.0 percent in the lowest tertile to 36.6 percent in the middle and 62.5 percent in the highest (trend p < 0.001). The authors caution, however, that the score was derived and evaluated in the same dataset, a circular procedure that likely inflates both the AUC and the tertile gradient, and cumulative anthracycline dose data were unavailable to rule out confounding by differential dosing.
Machine learning added an exploratory layer. Five classifiers, logistic regression, support vector machine, decision tree, XGBoost and LightGBM, were trained on 80 percent of the data and tested on a held-out 20 percent (n = 43), with and without SMOTE oversampling applied only to the training set. Without oversampling, XGBoost and LightGBM each achieved a test AUC of 0.73, with cross-validation AUCs of 0.79 and 0.78. SHAP interpretability analysis ranked ABCG2 rs2231142 as the most influential feature across all models, with CBR1 rs9024 consistently second, mirroring the regression findings. The authors stress that the test set was too small for stable estimates, that the near-identical AUCs of LightGBM and of a Decision Tree with SMOTE reflect sampling variance rather than algorithmic superiority, and that the results are proof-of-concept for a genetic feature set rather than a deployable model.
The limitations are substantial and the authors enumerate them candidly. Cumulative anthracycline dose, the strongest established predictor of cardiotoxicity, was not recorded, nor were cardioprotective drugs such as dexrazoxane. Echocardiograms were obtained at a single time point fifteen days after induction, capturing only acute cardiotoxicity, and the study is cross-sectional and single-center. The sample of 216 was powered to detect large effects, leaving modest variant effects undetectable, and all predictive models were evaluated on the data that generated them, so reported AUCs are likely optimistic. Family history of cancer, the only clinical variable significantly associated with cardiotoxicity (30.4 percent versus 17.5 percent; p = 0.043), may itself reflect residual confounding or multiple testing, and the authors explicitly warn against over-interpreting it.
What the study does establish is a methodological point with broad relevance to pharmacogenomics: inheritance model selection and verified effect direction are not statistical refinements but prerequisites for usable polygenic scores. Ignoring them reduced discrimination to chance even when individually significant variants were present in the same five-variant panel. Within this Pakistani pediatric cohort, ABCG2 rs2231142 emerges as the most robust single-variant determinant of anthracycline-induced cardiotoxicity, CBR1 rs9024 as an independently significant risk variant under the recessive model, and the weighted polygenic score as a promising but unvalidated stratification tool. The authors conclude that prospective validation in a dose-annotated, multi-center cohort, integrating conventional cardiac risk factors and cumulative drug exposure, is an essential prerequisite before any clinical implementation.
Subject of Research: Genetic predictors of anthracycline-induced cardiotoxicity in pediatric acute lymphoblastic leukemia
Article Title: Integrating inheritance model selection, weighted polygenic scoring, and interpretable machine learning for cardiotoxicity risk prediction in pediatric ALL
Article References: Aslam, S., Sarwar, S., Ahmed, M., & Shabana (2026). Integrating inheritance model selection, weighted polygenic scoring, and interpretable machine learning for cardiotoxicity risk prediction in pediatric ALL. Egyptian Journal of Medical Human Genetics, 27(1), Article 79. https://doi.org/10.1186/s43042-026-00908-7
Image Credits: AI Generated
DOI: 10.1186/s43042-026-00908-7
Keywords: anthracycline cardiotoxicity, pediatric ALL, ABCG2 rs2231142, CBR1 rs9024, polygenic risk score, machine learning, pharmacogenetics, LightGBM, UKALL 2003, left ventricular ejection fraction, Pakistan, Bonferroni correction
News Source: Teresa Odom. (October 29, 2026). Research report: Integrating inheritance model selection, weighted polygenic scoring, and interpretable machine learning for cardiotoxicity risk prediction in pediatric ALL. Scienmag.
