Benchmarking Cell Deconvolution Algorithms

We performed two in silico experiments to assess the detection limits of different deconvolution algorithms. In the first experiment (Supplementary Fig. 6), we used the same cell line GEPs described above to compare CIBERSORT and RLR with five other GEP deconvolution methods^{4 (link)–8 (link)}. We evaluated detection limit using Jurkat cells (spike-in concentrations of 0.5%, 1%, 2.5%, 5%, 7.5%, and 10%), whose reference GEP (median of three replicates in GSE11103) was added into randomly created background mixtures of the other three blood cell lines. Five mixtures were created for each spike-in concentration. Predicted Jurkat fractions were assessed in the presence of differential tumor content, which we simulated by adding HCT116 (described above) in ten even increments, from 0% to 90%. Of note, we also used the same marker or signature genes described for simulated tumors (above). In a second experiment (Supplementary Fig. 7a), we compared CIBERSORT with QP^{5 (link)}, LLSR^{4 (link)}, PERT^{6 (link)}, and RLR. We spiked naïve B cell GEPs from the leukocyte signature matrix into four random background mixtures of the remaining 21 leukocyte subsets in the signature matrix. The same background mixtures were used for each spike-in. We also tested the addition of unknown content by adding defined proportions (0 to 90%) of randomly permuted expression values from a naïve B cell reference transcriptome (median expression profile from samples used to build LM22, Supplementary Table 1). We then repeated this analysis for each of the remaining leukocyte subsets in LM22 (Supplementary Fig. 7b).