Table 4. Perceptual consistency of gene expression in reconstructed images.

Correlation between predictions of gene expression from real and reconstructed tiles, averaged per patient, demonstrating a high perceptual similarity of the gene expression of the real and generated images. For TCGA, a deep learning model was trained to predict gene expression from real tiles from TCGA-BRCA using threefold cross-validation. The correlation between predictions for real/generated images is aggregated for the three held-out validation sets. For the CPTAC-BRCA validation, a deep learning model trained across the entire TCGA-BRCA dataset was used to generate predictions. Also listed are the correlations between model predictions and true gene expression, as well as the same correlations made from the reconstructed (Gen) tiles. The similar correlation coefficients from real tiles and reconstructed tiles illustrates the reconstructed tiles retain informative data with regard to histologic subtype.

Source	Gene	n	Pearson r, real versus gen.	P value	Pearson r, real versus expression	P value	Pearson r, gen. versus expression	P value
TCGA	CD3G	941	0.9 (0.88–0.91)	<1 × 10−100	0.45 (0.39–0.49)	<1 × 10−100	0.38 (0.32–0.43)	<1 × 10−100
TCGA	COL1A1	941	0.84 (0.82–0.86)	6.88 × 10−269	0.45 (0.4–0.5)	<1 × 10−100	0.31 (0.26–0.37)	<1 × 10−100
TCGA	MKI67	941	0.91 (0.9–0.92)	0.00 × 100	0.43 (0.38–0.48)	<1 × 10−100	0.35 (0.3–0.41)	<1 × 10−100
TCGA	EPCAM	941	0.9 (0.88–0.91)	0.00 × 100	0.32 (0.26–0.37)	<1 × 10−100	0.25 (0.19–0.31)	<1 × 10−100
CPTAC	CD3G	97	0.73 (0.62–0.81)	3.08 × 10−17	0.3 (0.1–0.47)	<1 × 10−100	0.21 (0.02–0.4)	0.04
CPTAC	COL1A1	97	0.74 (0.64–0.82)	2.89 × 10−18	0.56 (0.4–0.68)	<1 × 10−100	0.55 (0.39–0.67)	<1 × 10−100
CPTAC	MKI67	97	0.82 (0.75–0.88)	5.34 × 10−25	0.04 (−0.16–0.24)	0.67	0.07 (−0.13–0.26)	0.51
CPTAC	EPCAM	97	0.44 (0.27–0.59)	5.21 × 10−6	0.1 (−0.1–0.29)	0.33	−0.04 (−0.24–0.16)	0.69