Thank you for presenting this nice approach. I have a question about the minimizing MMD step. Since the BERMUDA is trying to minimize the MMD of same cell types within different batches in the embedding space. I wonder whether it is possible that this step will make some cell types difficult to identify after batch effect removal? For instance, taking the XOR example, assuming we are in the embedding space and we have cell type 1 with two batches where their distribution centered at (0,0) and (1,1); cell type 2 with two batches with distribution centered at (0,1) and (1,0), respectively. If we try to minimize the MMD between two batches in these cell types, it seems that we may mix them at (0.5, 0.5). Will using HVG partially alleviate this problem?
Hi Tongxin,
Thank you for presenting this nice approach. I have a question about the minimizing MMD step. Since the BERMUDA is trying to minimize the MMD of same cell types within different batches in the embedding space. I wonder whether it is possible that this step will make some cell types difficult to identify after batch effect removal? For instance, taking the XOR example, assuming we are in the embedding space and we have cell type 1 with two batches where their distribution centered at (0,0) and (1,1); cell type 2 with two batches with distribution centered at (0,1) and (1,0), respectively. If we try to minimize the MMD between two batches in these cell types, it seems that we may mix them at (0.5, 0.5). Will using HVG partially alleviate this problem?
Looking forward to your response.
Best, Larry