Handling empty bins in distribution

[kim1117]

Hello,

I have a question regarding the binning process in Genes2Genes.

When the distribution of values is skewed (e.g., most cells concentrated around 0.4–0.6), the bins determined by optbinning can include regions where no cells are present. As a result, when plotting a bar graph, the number of bars does not match n_bins, and the number of bars differs between the query and reference distributions due to these empty bins.

My questions are:

1. Should empty bins be excluded from the analysis since they contain no relevant data? It seems that this issue arises because the bin range is fixed between 0 and 1. What are your thoughts on this?

2. The first and last bins containing cells likely have fewer counts. Is the matching test(assigning IDMWV) robust enough to handle such sparsely populated edge bins?

I would appreciate your insights on these issues. Thank you!

[dinithins]

Hello @kim1117,

Thanks for your questions.

1. G2G uses discrete bins of cells along the pseudotime only when creating a bar plot visualisation. You could try removing those few outlier cells, min-max normalise the rest of the reference cell pseudotimes and query cell pseudotimes separately again, and re-plot. However, I would not worry if the original data created empty bins in the first place, as the bar plot then agrees with the actual cell density plots. Most importantly, it does not matter for trajectory alignment, as G2G does not align pseudotime based on binned cells.

2. For G2G trajectory alignment, when we define the n_bins = N, it equally divides the [0,1] pseudotime range into N discrete time points known as interpolation time points. For each time point, it estimates a weighted gene expression distribution based on all the single cells in the trajectory (weighted by their pseudotime distance to the particular time point). Therefore, the matching test involves comparing these estimated distributions. For your datasets, since there are only very few cells around the early and later time points, the gene expression of those time points would be largely influenced by the cells present in their adjacent time points of high cell density. The much closer but the very few cells will also contribute to this estimation with a weight higher than that of the many cells that are distant. So the gene expression distribution estimated for such low-cell-density time point is kind of an average representation based on all the available data. You could think of this interpolation as something along the lines of missing data imputation based on the available data in the neighbourhood.

Hope this helps.