A Bayesian model for single-cell transcript expression analysis with MERFISH data

Johannes Köster

2016

MERFISH

FISH

Fluorescence in-situ hybridization:

• label RNA with fluorescent probes
• see RNA molecules in single cells

Problem:

needs 1 color per transcript

MERFISH

...

MERFISH

Problem with raw counts:

20% loss and misidentification

MERFISH

Known error probabilities:

1→0 error: 10%

0→1 error: 4%

Goal:

• Bayesian model on top of error probabilities
• estimate gene/transcript expression
• "Bayesian"-style differential expression analysis

Approach

Urn model for

expression likelihoods

Bayesian model for differential expression

Example results

Results

Simulation

Pr(0→1)    Pr(1→0)

simulated

hybridization

Bayesian model recovers biased counts

Bayesian model recovers biased counts

Application

Characterize batch effects in real data

Published dataset:

• ~200 fibroblast cells
• 7 batches
• same biological condition

Question:

Are expression profiles influenced by batch effects?

t-SNE analysis

cell size

cell position

batch

Approach

• calculate posterior estimate of coefficient of variation (CV) between means of batches
• null model: CV < 0.5
• control expected FDR at 5%

Differentially expressed genes

Gene ontology enrichment

controlled FDR at 5% with Benjamini-Yekuteli

Conclusion

Bayesian model for gene expression analysis on MERFISH data:

• provides estimates of expression, fold change and coefficent of variation
• credible intervals, expected FDR

Simulation:

can correct for biases in MERFISH data

Application:

applied to characterize batch effects

Acknowledgements

Shirley Liu

Myles Brown

Bo Li

Peng Jiang

Eric Severson