Machine Learning Prioritizes Pancreatic Noncoding Variants Affecting HNF4A Binding Motifs

An interpretable machine-learning framework combining single-nucleus chromatin accessibility measurements with tree-based modeling can prioritize noncoding genetic variants and uncover nonlinear regulatory relationships in human pancreatic progenitor cells, according to a study by Kelly A. Frazer, Benjamin Henson, Jennifer P. Nguyen, Wilfredo G. Gonzalez Rivera, and Matteo D’Antonio. The investigation addressed a central challenge in functional genomics: although noncoding genetic variants underlie many complex human traits, predicting their precise regulatory impacts within cell type-specific contexts remains difficult because chromatin accessibility, variant allele frequencies, and transcription factor binding interact in complex, nonlinear ways that constrain standard linear statistical approaches.

To evaluate these regulatory interactions during development, the researchers generated single-nucleus ATAC sequencing data from seven induced pluripotent stem cell-derived pancreatic progenitor cell samples. From these accessible chromatin regions, the authors identified 93,803 heterozygous single nucleotide polymorphisms and quantified allele-specific effects for each variant using binomial tests. They then trained an XGBoost model alongside standard logistic regression and regularized logistic regression baselines to predict allele-specific effects using four categories of features: chromatin read depth, peak score, variant minor allele frequency, and binding predictions for five pancreatic transcription factors, namely HNF4A, HNF1A, NEUROD1, NKX6-1, and PDX1.

Model performance was evaluated by the area under the receiver operating characteristic curve using fivefold cross-validation as well as designated test and validation single nucleotide polymorphism sets. The XGBoost framework significantly outperformed both linear models, showing its most pronounced performance advantage at stringent allele-specific effect thresholds, consistent with its capacity to learn nonlinear feature interactions. Chromatin read depth, peak score, and minor allele frequency proved to be the strongest individual predictors; however, stratifying variants across read-depth and minor allele frequency deciles demonstrated that XGBoost retained superior predictive accuracy across deciles, confirming that its performance advantage was not explained by these primary variables alone.

To interpret the model’s predictions and group variants by their functional characteristics, the team conducted Shapley value analysis and projected variants into lower-dimensional space. This dimensionality reduction revealed seven distinct variant clusters defined by characteristic chromatin states and allele frequencies, supported by highly significant pairwise differences in allele frequency distributions among the groups. One cluster was notably enriched for lower minor allele frequency variants exhibiting strong predicted interactions with HNF4A binding motifs; while its lower frequency profile was expected because minor allele frequency and its Shapley contribution were inputs to the embedding, its elevated allele-specific effect enrichment and distinct HNF4A interaction structure provided the biologically informative insight.

Further analyses using Combined Annotation Dependent Depletion supported a stronger relationship between HNF4A-related predictions and predicted deleteriousness within this lower-frequency cluster. The authors concluded that this pattern aligns with selective constraint operating on HNF4A-bound regulatory elements in pancreatic progenitors. These results establish an interpretable methodology for prioritizing noncoding variation in specific developmental contexts and provide a biological hypothesis for how lower-frequency noncoding variants may contribute to intrafamilial phenotypic variability observed in Maturity-Onset Diabetes of the Young type 1 (MODY1).

Subject of Research: Machine learning prioritization of noncoding regulatory variants in pancreatic progenitor cells

Article Title: XGBoost-SHAP analysis prioritizes lower-frequency variants predicted to affect HNF4A binding motifs in pancreatic progenitors

Article References: Frazer, K. A., Henson, B., Nguyen, J. P., Gonzalez Rivera, W. G., & D’Antonio, M. (2026). XGBoost-SHAP analysis prioritizes lower-frequency variants predicted to affect HNF4A binding motifs in pancreatic progenitors. PLOS Genetics, 22(11), e1012361. https://doi.org/10.1371/journal.pgen.1012361

Image Credits: AI Generated

DOI: 10.1371/journal.pgen.1012361

Keywords: genomics, machine learning, XGBoost, HNF4A, pancreatic progenitors, transcription factors, allele-specific effects, MODY1, XGBoost-SHAP, analysis, prioritizes, lower-frequency