Single cell gene set scoring with nearest neighbor graph smoothed data (gssnng)

Gene set scoring (or enrichment) is a common dimension reduction task in bioinformatics that can be focused on differences between groups or at the single sample level. Gene sets can represent biological functions, molecular pathways, cell identities, and more. Gene set scores are context dependent values that are useful for interpreting biological changes following experiments or perturbations. Single sample scoring produces a set of scores, one for each member of a group, which can be analyzed with statistical models that can include additional clinically important factors such as gender or age. However, the sparsity and technical noise of single cell expression measures create difficulties for these methods, which were originally designed for bulk expression profiling (microarrays, RNAseq). This can be greatly remedied by first applying a smoothing transformation that shares gene measure information within transcriptomic neighborhoods. In this work, we use the nearest neighbor graph of cells for matrix smoothing to produce high quality gene set scores on a per-cell, per-group, level which is useful for visualization and statistical analysis. Availability and implementation The gssnng software is available using the python package index (PyPI) and works with Scanpy AnnData objects. It can be installed using ‘pip install gssnng’. More information and demo notebooks: See https://github.com/IlyaLab/gssnng