Technical Dropout vs. Biological Silencing
Thank you for this great work.
You showed that silencing a gene by removing it from the input list affects the embeddings of other genes within the same cell. This raises a question I didn't see addressed in the paper, how does Geneformer distinguish between a gene that is truly not expressed versus one that was expressed but simply not captured due to technical dropout?
Since dropout is a fundamental source of noise in single-cell data, I was wondering whether the model has any mechanism to handle this.
Thank you in advance
Thank you for your question. If a gene is commonly dropping out without the remainder of the transcriptome changing, the model will observe this noisiness in the many observations of cell states in the training data so this will likely be encoded information. For the in silico perturbation, the comparison is between the unperturbed cell that contains the gene and a perturbed cell that does not contain it, so the two are directly contrasted to determine the direction of the predicted shift in response to the deletion. It is advisable to use data from platforms that detect a good number of genes as opposed to some that have very few genes detected due to shallow profiling.