The inference of molecular information from hematoxylin–eosin (HE) specimens may reduce the ancillary testing burden in digital pathology. To assess whether deep convolutional networks predict, at the per-nucleuslevel, continuous DAB optical density (OD) and binary positivity for a biologically coherent panel of DNA replication–related proteins (CDC6, CDT1, MCM7, ORC1, CDC7, and Geminin) plus Ki-67 directly from HE nuclear images. We constructed a same-specimen paired HE/IHC dataset from 21 endometrioid carcinoma cases (7 per grade). After color unmixing via NMF-derived stain vectors and HoVer-Net–based nuclear segmentation, 100 × 100 nucleus-centered HE crops were paired one-to-one with IHC-derived per-nucleus OD. ImageNet-initialized backbones (ResNet-50 baseline vs. EfficientNet-B0 and MobileNetV3-Small) were trained with regression (OD) and classification (positivity) heads, and multi-task learning across markers was also evaluated. Case-wise splits ensured no patient overlap across training/validation/test. Across markers, the per-nucleus prediction of protein expression from HE nuclear morphology was feasible with moderate discriminative performance, with the strongest signal being observed for MCM7 (AUC-ROC ≈ 0.71; F1 ≈ 0.72). Performance was dependent on the markers used: MCM7 and Ki-67 consistently showed stronger discrimination (AUC-ROC ≈ 0.70–0.72), Geminin and CDC6 were moderately predictable in some settings, and CDT1 remained near the chance level (AUC-ROC ≈ 0.50). Among the architectures evaluated, ResNet-50 demonstrated the most stable generalization, and multi-task training yielded modest average gains, but was not consistently beneficial. Regarding the Ki-67 labeling index, the nuclear-crop approach showed moderate agreement with WSI-based digital IHC (ROI level r ≈ 0.54; MAE ≈ 15.2 pp). The major limitations of this study were the small cohort size (21 cases) and lack of external validation. Per-nucleus protein expression—both binary positivity and continuous OD—was shown to be inferable from HE nuclear morphology alone. These results suggest clinical utility and will motivate future studies using larger external cohorts and self-supervised pretraining.
Urata et al. (Sat,) studied this question.