XGBoost-SHAP analysis prioritizes lower-frequency variants predicted to affect HNF4A binding motifs in pancreatic progenitors
Kelly A. Frazer, Benjamin Henson, Jennifer P. Nguyen, Wilfredo G. Gonzalez Rivera, Matteo D’AntonioBackground
Noncoding genetic variants underlie many complex traits, yet predicting their regulatory impact in cell type-specific contexts remains a major challenge. Transcription factor binding is a primary mechanism through which noncoding variants influence gene regulation, but the relationships among chromatin accessibility, variant frequency, and transcription factor binding are complex and nonlinear, limiting the performance of traditional linear models.
Methods
We generated single-nucleus ATAC sequencing data from seven induced pluripotent stem cell-derived pancreatic progenitor cell samples and identified 93,803 heterozygous single nucleotide polymorphisms overlapping accessible chromatin regions. Allele-specific effects (ASE) were quantified for each variant via binomial tests. We trained XGBoost, logistic regression, and regularized logistic regression models to predict allele-specific effects via chromatin read depth, peak score, variant minor allele frequency (MAF), and transcription factor binding prediction for five pancreatic transcription factors: HNF4A, HNF1A, NEUROD1, NKX6-1, and PDX1. Model performance was assessed by the area under the receiver operating characteristic curve (AUC) via fivefold cross-validation as well as using test and validation SNP sets. Shapley (SHAP) value analysis was used to interpret model predictions and to cluster variants on the basis of their regulatory feature contributions.
Results
XGBoost significantly outperformed linear models, particularly at stringent allele-specific effect thresholds, consistent with its ability to capture nonlinear interactions among features. Chromatin read depth, peak score, and minor allele frequency were the strongest predictors. Stratification by read-depth and MAF deciles showed that XGBoost retained better overall predictive performance across deciles than the linear models, indicating that its advantage was not explained by these variables alone. Dimensionality reduction of variants on the basis of Shapley value profiles revealed seven clusters with distinct chromatin states and variant allele frequencies. These findings were supported by highly significant pairwise differences in allele frequency distributions across the seven clusters. One cluster was enriched for lower-MAF SNPs showing strong predicted interactions with HNF4A binding motifs. Because MAF and its SHAP contribution were inputs to the embedding, the lower-MAF identity of this cluster was expected; its high ASE enrichment and distinctive HNF4A interaction structure were the biologically informative findings. Combined Annotation Dependent Depletion (CADD) analyses further supported a stronger relationship between HNF4A-related predictions and predicted deleteriousness in this cluster.
Conclusions
Integration of single-nucleus chromatin accessibility data with an interpretable XGBoost-SHAP framework provides a method for prioritizing noncoding variants and identifying nonlinear regulatory relationships in a specific developmental context. Within this framework, the stronger relationship between HNF4A-related predictions and CADD-predicted deleteriousness in the lower-MAF cluster is consistent with selective constraint on HNF4A-bound regulatory elements and suggests a hypothesis for how lower-frequency noncoding variation may contribute to intrafamilial phenotypic variability in HNF4A-related Maturity-Onset Diabetes of the Young (MODY1).