Table 2.

Purity Estimation Benchmarking Results

Dataset	Model	MAE	RMSE	Pearson’s	CCC
GBmap*	*GBMPurityθ	0.046	0.083	0.978	0.974
	*MuSiCψ	0.105	0.163	0.937	0.921
	*CIBERSORTxψ	0.096	0.172	0.929	0.912
	PUREEλ	0.315	0.372	0.667	0.311
	*Scadenθ	0.378	0.403	0.010	0.001
Wang et al11	GBMPurityθ	0.145	0.177	0.917	0.864
	MuSiCψ	0.189	0.271	0.751	0.662
	CIBERSORTxψ	0.347	0.445	0.422	0.264
	PUREEλ	0.283	0.375	0.667	0.237
	Scadenθ	0.309	0.338	−0.082	−0.004
Neftel et al5	CIBERSORTxψ	0.139	0.174	0.815	0.789
	GBMPurityθ	0.162	0.186	0.902	0.687
	MuSiCψ	0.279	0.320	0.854	0.583
	PUREEλ	0.214	0.259	0.864	0.441
	Scadenθ	0.281	0.332	0.388	0.012
EORTC	GBMPurityθ	0.128	0.160	0.757	0.743
	MuSiCψ	0.138	0.181	0.725	0.709
	CIBERSORTxψ	0.170	0.229	0.581	0.570
	PUREEλ	0.167	0.212	0.765	0.529
	Scadenθ	0.195	0.229	−0.073	−0.005
TCGA	*PUREEλ	0.102	0.123	0.803	0.701
	MuSiCψ	0.114	0.160	0.730	0.694
	GBMPurityθ	0.133	0.159	0.690	0.597
	CIBERSORTxψ	0.160	0.200	0.677	0.565
	Scadenθ	0.272	0.299	−0.103	−0.003

Greek superscripts refer to methodology:

θ, deep learning based on pseudobulks;

ψ, machine learning weighting of single-cell reference profiles;

λ, pre-trained pan-cancer model based on genomic estimates. More details on these methods can be found in the Methods section. Asterisks refer to models trained on the respective dataset. Abbreviations: MAE, mean absolute error; RMSE, root mean squared error; CCC, correlation concordance coefficient; R Pearson’s correlation coefficient. Models are ordered within each dataset by descending CCC, with bold text indicating the best-performing metric for that dataset.