Tumor-lymphoid aggregate interface structures predict immunotherapy response in hepatocellular carcinoma - Test H&E images and notebook

This entry contains the external validation data and relevant codes to reproduce the results in the manuscript "Tumor-lymphoid aggregate interface structures predict immunotherapy response in hepatocellular carcinoma", where a tumor-microenvironment (TME) signature that predicts immune checkpoint blockade (ICB) response in hepatocellular carcinoma (HCC) from routine H&E images is tested. The signature was originally derived from joint analysis of spatial multi-omics data (please see the manuscript), and we used an in-house method (preprint: https://www.biorxiv.org/content/10.64898/2026.06.03.729847v1) to transfer the signature to H&E. Main Contents: HOPE_for_HCC-ICB.ipynb: Main notebook for reproducing the validation results; training_data/: Patch annotations + pre-computed training features derived from the discovery cohort; validation_data/MDACC_HCC-ICB-HE/: external test cohort 1, containing 23 images from 13 patients; validation_data/UHB_HCC-ICB-HE/: external test cohort 2, containing 37 images from 33 patients; inference/: Pre-computed embeddings for the external test cohorts (via UNI2-h) Requirements Python 3.11, PyTorch, timm, torchvision, zarr, tifffile, opencv-python, pandas, numpy, scikit-learn, scipy, seaborn, matplotlib, tqdm. The UNI2-h checkpoint is not redistributed (at: https://huggingface.co/MahmoodLab/UNI2-h). Download it separately and set UNI2h_CKPT_PATH. Only needed to re-extract features — not to reproduce the reported results from the pre-computed embeddings. Citation Please cite the main manuscript and the methods preprint listed above.