Stem cell differentiation into insulin-producing pancreatic beta cells represents a transformative avenue for type 1 diabetes therapy. Existing computational approaches predominantly rely on convolutional neural networks (CNNs) applied to microscopic cell images; while achieving high classification accuracy, such methods suffer from limited interpretability and dependence on expensive imaging infrastructure. This paper presents BetaXplain, an explainable machine learning framework that predicts differentiation success from quantitative gene expression measurements rather than visual features. The system models differentiation outcome as a binary classification problem over a feature vector of five biologically validated transcription factor biomarkers — PDX1, NKX6.1, NGN3, INS, and MAFA — derived from the public GSE83139 gene expression dataset. We evaluate three classifiers — Random Forest, Support Vector Machine, and XGBoost — with XGBoost achieving the highest accuracy of 92.4%, precision of 91.8%, and recall of 93.1%. Prediction outputs are augmented with gene importance rankings and radar-chart visualizations that provide biologically interpretable explanations of each decision. Comparative analysis against CNN-based baselines demonstrates that BetaXplain matches predictive performance while substantially improving transparency, reducing infrastructure requirements, and enabling direct biological insight. The framework constitutes a step toward clinically actionable, interpretable AI in regenerative medicine.

Index Terms—stem cell differentiation, gene expression, explainable AI, XGBoost, pancreatic beta cells, bioinformatics, transcription factor biomarkers