MUSIC Topic Modeling of LUHMES CROP-seq Data

Single cell expression data

Source:
High-throughput single-cell functional elucidation of neurodevelopmental disease-associated genes reveals convergent mechanisms altering neuronal differentiation, GEO accession: GSE142078.

Perturbations:
CRISPR knock-down of 14 autism spectrum disorder (ASD)–associated genes (3 gRNAs per gene) + 5 non-targeting gRNAs.

Cells:
Lund human mesencephalic (LUHMES) neural progenitor cell line.
Cells from 3 batches were merged together into 1 analysis. All cells have only a single type of gRNA readout.

MUSIC pipeline

Preprocessing

Cell quality control
SAVER data imputation
Filtering cells with invalid edits
(if a cell's differentially expressed gene profile is more similar to the control cells than to the other cells with the same perturbation.)
Selecting genes with large dispersion difference between case and control
Normalizing and rounding the expression values to non-negative integers

Final gene and cell size: [1] 3197 4146

Result interpretation

Annotating the functions of each topic:

MUSIC obtains the occurrence probabilities of genes available in each topic. It then selects the top 10% genes of each topic based on their occurrence probabilities, and perform functional enrichment analysis using all genes in topic modeling as background.

Characterizing topc-perturbation relationships:

For a specific topic, MUSIC prioritizes the perturbation effect by calculating the specific topic probability difference (TPD) between the case and control groups.

For the \(i\)-th perturbation on the \(j\)-th topic, the TPD against the control group is computed as the Student's \(t\)-statistics between \(\{z_{mj}\}_{m\in \text{perturbtion }i}\) and \(\{z_{nj}\}_{n\in \text{control}}\),
where \(z_{mj}\) is the probability of topic \(j\) in cell \(m\) normalized w.r.t. the control group: \(z_{mj} = \frac{\theta_{mj}-\mu_{\text{control}}}{\sigma_{\text{control}}}\).