The only option you have is to treat every single cell as a replicate. That has the obvious downside that any effect can be a donor-unique effect, since you have no biological replication to rule that out. There is literature to guide choice of tools for such DE analysis (https://pubmed.ncbi.nlm.nih.gov/29481549/). I always use limma-trend. My personal best practice here, somewhat similar to what is in the linked paper is to properly prefilter the data for genes with sufficient expression level in a certain fraction of cells, and to test against a fold change with limma, avoiding significances with tiny effect sizes. For example, testing against a FC of 1.1 with limma-treat will already bring number of nonsense DEGs down a lot.