Fig. 4

scBFA outperforms quantification models when the gene detection noise is less than gene quantification noise. Rows represent different settings of (gene) detection noise (\( {\sigma}_{\boldsymbol{\pi}}^2 \)), and columns represent different settings of (gene) quantification noise (\( {\sigma}_{\mu}^2 \)). The diagonal represents simulations where the detection noise is equal to the quantification noise (\( {\sigma}_{\mu}^2={\sigma}_{\pi}^2 \)), and the plots above the diagonal represent simulations where the detection noise is less than the quantification noise. Each y-axis indicates the cross-validation performance (MCC) of cell type predictors trained on embeddings learned from the simulated data, while each x-axis represents the gene detection rate that is manipulated by the parameter δ. Here, the ground truth embedding matrix is obtained by fitting ZINB-WaVE to the LPS benchmark under HVG selection. The dispersion parameter r is set to be 1 in these simulations.