For example, if I have 1 million cells in suspension and from this suspension I separate single cells, how many cells should be sequenced individually to have a statistically significant result?
What is significant? I mean, what apsect of the data are you looking at and how do you define an effect size? And what effect size will you then judge significant? Or - just in case you were talking about statistical significance - at what level do you want to reject the null hypothesis?
If you seek for effects between cells, I don't understand why you separate the cells instead of straight-away measuring the average profiles of many cells...
It definitely depends on your level of sub-population heterogeneity, but I've read numbers ranging anywhere from 10 to 10,000 (10K being a representation of a reconstructed bulk population). Most papers that I have read generally do 10-100 per treatment/condition and carefully word their scientific questions.
I think a good justification for a sample size of 50 single cells is provided in Box 2 of this paper:
"In a standard bulk RNA sequencing (RNA-seq) experiment, precision is limited only by sequencing depth. Typically, ten million reads are generated, and a threshold of 50 reads per kb per million reads (RPKM) is considered adequate to call a gene as expressed. For a gene that is 1 kb long, this corresponds to 500 reads, thus leading to a minimum coefficient of variation (CV; which is equal to the standard deviation divided by the mean) of 4%, as given by the Poisson distribution. In a fairly typical single mammalian cell containing 200,000 mRNA molecules, 50 RPKM corresponds to about ten mRNA molecules. Again, assuming a Poisson distribution across cells, the expected CV is 32%, but this can be reduced by pooling data from many cells. How many cells are needed to reduce this error to that of the bulk experiment? The answer is 50, because the pooled data from 50 cells will contain 500 mRNA molecules. These are ideal numbers, and in practice more cells will be required. For example, if the efficiency of converting mRNA to cDNA is only 10% (which is not an unrealistic assumption), then tenfold more cells will be required. Similarly, when additional noise is introduced (for example, by PCR amplification) the number of cells required will increase correspondingly. Furthermore, if the sample is heterogeneous, then enough cells must be analysed so that all representative cell types are observed. Finally, all these estimates assume that the single-cell measurements are accurate, as systematic inaccuracies (for example, due to amplification bias) will not be cured by collecting more cells."