State-Aware Interpretability in Graph Neural Networks for Stem Cell Fate Prediction

The question of how a cell determines its fate—whether to remain in a pluripotent stem-like state or differentiate into a specialized form—lies at the core of modern biology. This process is not controlled by a single gene but emerges from complex interactions among multiple regulatory factors. These interactions are often non-linear, dynamic, and context-dependent. From a computational perspective, this makes stem cell fate prediction a challenging task due to the high dimensionality, noise, and interconnected nature of gene expression data. Although machine learning models have shown promise in predicting cellular states, many operate as black boxes, limiting their usefulness for biological interpretation. Bridging this gap between predictive accuracy and interpretability is the primary motivation of this work. In this study, we introduce a state-aware and interpretable graph-based framework for modeling stem cell fate by integrating concepts from systems biology and deep learning. Biologically, we focus on key pluripotency-associated genes—Oct4, Sox2, Klf4, cMyc, Nanog, Lin28, and Esrrb—which form a tightly interconnected regulatory network. To capture this structure, we generate a synthetic dataset that mimics realistic gene expression behavior, incorporating variability and noise observed in biological systems. 
From a modeling perspective, we evaluate three approaches: Random Forest, Multi-Layer Perceptron (MLP), and Graph Neural Networks (GNNs). Random Forest provides a strong baseline with interpretable feature importance, while MLP captures non-linear relationships but treats genes independently. In contrast, the GNN models genes as nodes in a graph, allowing information to propagate through edges representing biological interactions. This enables the model to learn from collective gene behavior rather than isolated features. 
Our results show that while all models achieve reasonable performance, the GNN consistently demonstrates better generalization and robustness, particularly in ROC-AUC and F1-score. Importantly, its performance depends strongly on the underlying graph structure: replacing biologically inspired connections with random ones significantly reduces accuracy. This highlights the value of incorporating domain knowledge into machine learning models.