I had a question about over adjustment with SVA. I used the svaseq() on my dataset and ended up with 11 SV's. I had problems with DE analysis because my matrix was "computationally singular". It seems that the first couple of SV's in our dataset are highly correlated > 0.7 with cell type proportions. After reducing to 3 SV's, based on the "be" method (the leek method recommended 65), my dataset was no longer computationally singular. Were the 11 SV's just over adjusting and capturing all the heterogeneity in our sample (does this mean the data is of poor quality - given that all heterogeneity was able to be captured in 11 SVs)? Also, what gives rise to such a large discrepancy in the number of SV's proposed by the "be" and "leek" methods? Do you suggest adding cell type proportions to our svaseq modelmatrix formula and then also adjusting in downstream analysis?