Interpretable gene networks from single-cell foundation models reveal conserved neurogenic dysfunction in Parkinson’s disease
Interpreting large-scale single-cell transcriptomic data remains a major challenge for understanding disease mechanisms. Recent single-cell foundation models learn rich representations of gene relationships across millions of cells, yet methods for translating these embeddings into biologically interpretable gene networks remain limited. Here we present scGENet, a computational framework that constructs context-specific gene interaction networks from foundation model–derived gene embeddings. By fine-tuning pretrained models on transcriptomic data from human midbrain organoids, scGENet generates transcriptome-scale gene modules that capture biologically meaningful cellular programs. Benchmarking across multiple foundation models demonstrates that networks derived from a fine-tuned scGPT brain model show the highest concordance with curated neuronal pathways, Parkinson’s disease (PD) genetic risk loci, and independent patient-derived transcriptional signatures. Applying this framework to human iPSC-derived PD midbrain organoids reveals transcriptional modules associated with neuronal differentiation, synaptic signaling, and cell-cycle regulation. Single-nucleus RNA sequencing further links these programs to altered cellular composition, including reduced dopaminergic neurons, expansion of radial glia–like progenitors, and a dopaminergic neuron subtype expressing SNCA and VGLUT2. Integration with independent human substantia nigra datasets identifies a conserved neurogenic program disrupted across genetic and idiopathic PD. Together, these results establish a generalizable strategy for extracting interpretable gene networks from single-cell foundation models, enabling systematic discovery of disease-relevant molecular programs across diverse tissues and datasets.