Fig. 7: Zero-shot pan-cancer generalization on TCGA. | npj Digital Medicine

Fig. 7: Zero-shot pan-cancer generalization on TCGA.

From: Decoding the ERS–CAF immunoregulatory axis via multimodal AI and its pan-cancer prognostic and therapeutic predictive value

Fig. 7: Zero-shot pan-cancer generalization on TCGA.

The chordoma-trained model is applied as-is (frozen weights, no TCGA tuning) to TCGA-PAAD/STAD/COADREAD WSIs. Patient-level scatter plots visualize concordance against transcriptomic reference scores for the three targets. Concordance is summarized with 95% CIs, and permutation controls show correlations are not explained by chance alignment. a TCGA-PAAD: zero-shot concordance (patient-level). b TCGA-STAD: zero-shot concordance (patient-level). c TCGACOADREAD: zero-shot concordance (patient-level). d Pearson r with 95% CIs (Fisher transform; by cancer and target). e Permutation control: shuffled predictions yield r ≈ 0.

Back to article page