NCI Division of Cancer Biology Junior Investigators Meeting (Abstract): Submission #7

Abstract Differentiating meaningful biological variation from noise is a fundamental problem in single-cell transcriptomics. As an alternative for cell-state identification, previous studies employed stochastic 10-cell sequencing, a mini-bulk method that compares 10-cell samples to larger pooled samples to identify heterogeneous expression states while maintaining reproducibility. However, 10-cell sequencing alone cannot combine samples across batch variables, such as sex or individual patients, which reduces analytical sample sizes and limits biological insight. To improve this method controlling for batch effects and technical noise, we applied a generalized additive model of location and scale. This model treats observed variance as a linear combination of factor variances, such as those from technical error and batch effect, and isolates the residual, biologically meaningful variance. We applied this model to stochastic 10-cell sequencing data from a mouse model of gliomagenesis and to human estrogen receptor-positive breast cancer samples. In each setting, the generalized additive model of location and scale predicted heterogeneously expressed genes that previous analyses missed. We validated these model predictions using RNA fluorescence in situ hybridization in independent samples. The generalized additive model of location and scale increases biological information extracted from limited sample sizes. This model may be applicable to other mini-bulk methods seeking to isolate meaningful data from noise.