Abstract Accurate classification of cancer subgroups is essential for precision medicine, tailoring treatments to individual patients based on their cancer subtypes. In recent years, advances in high-throughput sequencing technologies have enabled the generation of large-scale transcriptomic data from cancer samples. These data have provided opportunities for developing computational methods that can improve cancer subtyping and enable better personalized treatment strategies. Here in this study, we evaluated different feature selection schemes in the context of meningioma classification. While the scheme relying solely on bulk transcriptomic data showed good classification accuracy, it exhibited confusion between malignant and benign molecular classes in approximately ~8% of meningioma samples. In contrast, models trained on features learned from meningioma single-cell data accurately resolved the sub-groups confused by bulk-transcriptomic data but showed limited overall accuracy. To integrate interpretable features from the bulk (n=78 samples) and single-cell profiling (~10K cells), we developed an algorithm named CLIPPR which combines the top-performing single-cell models with RNA-inferred copy number variation (CNV) signals and the initial bulk model to create a meta-model, which exhibited the strongest performance in meningioma classification. CLIPPR showed superior overall accuracy and resolved benign-malignant confusion as validated on n=792 bulk meningioma samples gathered from multiple institutions. Finally, we showed the generalizability of our algorithm using our in-house single-cell (~200K cells) and bulk TCGA glioma data (n=711 samples). Overall, our algorithm CLIPPR synergizes the resolution of single-cell data with the depth of bulk sequencing and enables improved cancer sub-group diagnoses and insights into their biology.

Leveraging Single-Cell Sequencing to Classify and Characterize Tumor Subgroups in Bulk RNA-Sequencing Data