Integrating Artificial Intelligence-Driven Digital Pathology and Genomics to Establish Patient-Derived Organoids as a Novel Alternative Model for Drug Response in Head and Neck Cancer

Abstract

Patient-derived organoids (PDOs) are emerging as advanced 3D ex vivo novel alternative method (NAM) preclinical models, offering significant advantages over traditional cell lines and monolayer cultures for therapeutic development. In this study, we established PDOs from surgically resected fresh tissues of human papillomavirus (HPV)-negative head and neck squamous cell carcinoma (HNSCC) across anatomical sites, tumor T-categories, and sample types. These PDOs faithfully recapitulate the tumor's pathology, mutational profile, and drug response. To enable rapid classification of PDO identity, we developed a new convolutional neural network (CNN) model, TransferNet-PDO, which accurately distinguished tumor versus normal PDOs in culture using digital histopathology images (AUC≥0.88). PDOs maintained stable cultures and were cryopreserved between passages 5 and 12. Immunohistochemistry (IHC) staining (PanCK, p63, Cytokeratin 13, Ki67) confirmed squamous phenotype and histologic aggression of the original tumor. For tumors harboring TP53 mutations by whole-exome sequencing (WES), PDOs retained the corresponding p53 functional status as confirmed by IHC (enhanced or loss of expression). Somatic mutational landscape revealed that PDOs preserved driver somatic mutations, copy number variations (CNVs), and clonal architecture including low-prevalence subclones. Drug sensitivity assessment of PDOs showed that cisplatin reduced cell viability, whereas cetuximab and lenvatinib had minimal effects. Chemoradiation led to greater tumor organoid killing compared to radiation or chemotherapy alone. This study presents an integrated HNSCC PDO platform combining tissue biobanking, organoid establishment, multi-omics characterization, functional drug screening, and AI-driven histopathologic classification, providing a comprehensive and scalable system for translational cancer research.

Keywords: Artificial Intelligence; Digital pathology; Drug response; Genomics; Head and neck cancer; Human specimens; Novel alternative model; Organoids.

Figure 1.. TransferNet-PDO accurately distinguishes malignant from normal PDOs in HNSCC.

(A) Study design for PDO development and analysis. Five main modules are shown: (1) Fresh tumor specimens from patients with HNSCC are collected at the time of surgical resection; (2) PDOs are developed and expanded through serial passaging; tissue aliquots are banked using multiple preservation methods (FFPE, snap frozen, viably frozen) when available; (3) A new CNN classifier (TransferNet-PDO) is developed to distinguish tumor versus normal PDOs from H&E images; (4) PDOs undergo whole-exome sequencing to characterize somatic alterations, tumor mutational burden (TMB), and clonal architecture in comparison to the original tumor; (5) Drug sensitivity is assessed by cell viability assays and transcriptomic analysis. (B) ransferNet-PDO deep-learning workflow. 256×256 tiles were extracted from H&E-stained PDO while slide images (WSIs) and input into a pre-trained Hover-Net ResNet50 backbone for nuclear segmentation. In the initial transfer learning phase (Phase 0), classification weights were re-trained while backbone features were frozen. In the fine-tuning phase (Phase 1), residual layers of the ResNet50 backbone were unfrozen, and the entire network was optimized using a reduced learning rate. Model performance on segmentation was evaluated using dice coefficients across epochs, with the best checkpoint selected at epoch 32. Final classifications on tumor/normal were made at the single-cell level based on nuclear features, and PDO-level predictions were aggregated using a winner-take-all (WTA; also known as majority voting) strategy. (C) Representative examples of classified tumor PDOs, normal PDOs, and mixed tumor/normal PDO cultures. Individual nucleated cells are classified as tumor (red) or normal (green). (D) Performance of TransferNet-PDO in classifying individual cells as tumor or normal. (E) Performance of PDO-level classification via WTA among classified cells. For D and E: (left) sensitivity, specificity, and ROC-AUC across training, validation, and three independent WSI validation cohorts; (right) corresponding ROC curves for each cohort. Detailed methodology is described in Supplementary Methods.

Figure 2.. HNSCC PDOs show high-fidelity genomic landscape.

(A) Somatic mutations in known HNSCC cancer genes across six HPV-negative HNSCC patient tumors and corresponding PDOs across serial passages (indicated as p#). (top) bar plots show total tumor mutational burden (TMB) per sample, defined as the number of protein-altering somatic variants (SNVs and small insertions/deletions [indels]) by whole-exome sequencing (WES); (bottom) the heatmap displays mutations in known genes involved in HNSCC cancer development and progression, categorized by functional class: tumor suppressors, oncogenes, epigenetic regulators, and others. Each row indicates a unique variant shown to the right of the heatmap; each column represents a tumor or PDO sample. Variant classifications are color-coded. (B) Representative H&E and p53 IHC staining of tumor tissues and matched PDOs from two patients (HN23–10730 with positive p53, and HN24–10810 with negative p53). PDOs show histologic and immunohistochemical concordance with the original tumor. (C) Representative example of tumor and serial PDO passages from patient HN24–10810 (top to bottom: p1-p10) showing genome-wide somatic copy number variations (CNV). Denoised copy ratio is shown on y-axis. Chromosome numbers are shown on x-axis. CNV patterns remain largely stable across passages, reflecting genomic fidelity. (D) Heatmap of log₂ copy number changes across the genome for all six patient-PDO pairs from A. Each row represents a tumor (denoted by asterisk, first row of each panel) or PDO sample; chromosome number and cytobands are shown on the x-axis. Gains (red) and losses (blue) are shown relative to diploid baseline, demonstrating preservation of major CNV events in PDOs.

Figure 3.. HNSCC PDOs maintain clonal architecture over time.

(A) Representative example of H&E and p53 IHC staining of PDOs from patient HN23–10730 across passages 4, 6, and 9. PDOs retain malignant histology and consistent nuclear p53 accumulation over time, reflecting the underlying TP53 mutational status. Scale bar = 100 μm. (B) Clonal architecture reconstruction from WES somatic mutation and CNV analysis for six HPV-negative HNSCC cases from Fig. 2. Fishplots depict subclonal dynamics from the original tumor across PDO passages, with each vertical line representing a sample in the same order from Fig. 2A. TP53 mutant clones were the founding clones in 5 of 6 cases. Case HN24–10900 showed a clonal shift after antifungal treatment (blue arrow). Minor subclones (white arrows) were preserved. Known pathogenic mutations in HNSCC driver genes are annotated next to each plot, with colored dots on the phylogenetic tree corresponding to specific subclones from the fishplots.

Figure 4.. Tumor PDOs demonstrate clinically relevant sensitivity to standard-of-care therapies in HNSCC.

(A-C) Dose-response curves for cisplatin in A, cetuximab in B, and lenvatinib in C on PDOs from multiple patients. PDOs showed consistent sensitivity to cisplatin (IC50 range: 1–20 μM), while cetuximab and lenvatinib had minimal effects at most tested doses. Viability was measured using the CellTiter-Glo^® assay and normalized to untreated control. (D) Gene set enrichment analysis (GSEA) of RNAseq gene expression from Lenvatinib (10² nM, 24 hours)-treated PDOs (FDR-adjusted P<0.05) from four patients (eight samples total). Top MSigDB Hallmark 50 gene sets (H50) are shown on the row. Dots are colored by direction (NES>0: upregulated; NES<0: downregulated), and their size are scaled by -log10(FDR-adjusted p-value). A full list of H50 GSEA results is provided in Table S9. (E) GSEA plots showing lenvatinib induced downregulation of MYC targets and OXPHOS, and upregulation of type I/II IFN signaling in HNSCC PDOs. Enrichment scores are shown on y-axis, and ranked gene positions are shown on x-axis. (F) Chemoradiation response in two tumor PDO lines. For each treatment group, four replicates were included. P-values in parathesis next to each treatment group represent whether the slope of fitted models (indicating radiation dose-dependent cell viability reduction) was significantly different from zero within that group (full statistics are provided in Table S10). Denotation: * P<0.01, *** P<0.001, **** P<0.0001. Two-sided hypergeometric test was used in D. Linear mixed-effects models were used in F, with radiation dose as the fixed effect and replicate id as the random effect. P-values shown in D and F were after multiple comparison adjustments by BH-FDR method.