I wonder if we could combine a celldex reference with the tumor cells from SCPCL000822 to create a custom reference? This way we would have more than the two classes. This might avoid the overfitting problem and allow for SingleR to hopefully better distinguish between potential normal cell types and then tumor cells.
This is kind of what I was thinking, using the multiple reference functionality of SingleR. I kind of expect that you could keep the "Normal" label in the Tumor/normal reference from 822, as it should not fit better than any of the more specific references, and could serve as a bit of a negative control. Or maybe you could use the SingleR calls for 822 "normal" cells, but replace all of tumor calls with the "tumor" label. Maybe try both.
This issue is to track looking at using different types of references to classify tumor cells in SCPCL000824. We should then compare to the annotations from marker gene based methods output in #532.
This is kind of what I was thinking, using the multiple reference functionality of SingleR. I kind of expect that you could keep the "Normal" label in the Tumor/normal reference from 822, as it should not fit better than any of the more specific references, and could serve as a bit of a negative control. Or maybe you could use the SingleR calls for 822 "normal" cells, but replace all of tumor calls with the "tumor" label. Maybe try both.
Originally posted by @jashapiro in https://github.com/AlexsLemonade/OpenScPCA-analysis/issues/532#issuecomment-2177237451
This issue is to track looking at using different types of references to classify tumor cells in SCPCL000824. We should then compare to the annotations from marker gene based methods output in #532.