Abstract 4669: Label free identification of cancer cell death pathways via holotomography and deep learning as an early pharmacodynamic biomarker

Abstract Accurate measurement of regulated cell death (RCD)—including apoptosis, necroptosis, and necrosis—is critical for oncology drug development and mechanism-of-action studies. Conventional fluorescence assays introduce phototoxicity, labeling bias, and incompatibility with long-term or high-frequency pharmacodynamic monitoring. We developed a fully label-free platform that integrates 3D holotomography (HT) and deep learning to classify RCD phenotypes directly from intrinsic refractive-index (RI) signatures, enabling non-perturbative, mechanism-aware drug-response biomarkers.HeLa cells were induced into apoptosis, necroptosis, or necrosis using canonical biochemical triggers, with fluorescence markers (Annexin V, PI, Hoechst) used solely for ground truth. 3D RI tomograms acquired by HT-X1 Plus were converted to 2D maximum-intensity-projection (MIP) patches to train an ImageNet-pretrained CNN to classify five states (live-control, live-treated, apoptosis, necroptosis, necrosis) using a sliding-window/majority-vote strategy. Temporal concordance was evaluated through synchronized HT–fluorescence time-lapse imaging and flow cytometry. For subtle drug-induced phenotypes (doxorubicin, cisplatin), performance of full 3D volumetric models was compared with 2-D projections to assess the necessity of depth information. Cross-cell-line robustness was tested on A549 cells with minimal fine-tuning.The five-state classifier achieved 99.3% accuracy on held-out HeLa datasets, with misclassifications limited to the apoptosis–necroptosis boundary. HT-based predictions identified early necroptotic transitions 2–4 hours before Annexin V/PI fluorescence, and population-level dynamics closely matched flow cytometry, establishing an earlier, dye-free pharmacodynamic window. In drug-response experiments, 3D volumetric models outperformed all 2D approaches, capturing spatially heterogeneous, mechanism-rich morphological signatures (76–88% accuracy in 3D vs. 50–55% in 2D MIP and 0% in SUM projections). The HeLa-trained model generalized poorly to A549 cells initially (50.4% accuracy), but small-data fine-tuning restored near-perfect performance, demonstrating practical assay portability across cancer cell types.Holotomography-based AI provides a fully label-free, segmentation-free, real-time biomarker for distinguishing RCD pathways and quantifying early drug responses with high accuracy. The platform detects necroptosis hours earlier than biochemical assays, resolves subtle drug-induced morphologies, and adapts rapidly to new cell types. These capabilities position HT-AI as a scalable pharmacodynamic tool for mechanism-of-action profiling, cytotoxicity testing, and high-content oncology drug discovery, enabling longitudinal, non-destructive phenotyping beyond fluorescence-based methods. Citation Format: Minwook Kim, Park Weisun, Geon Kim, Sanggeun Oh, Juyeon Park, Jihwan Yu, Hyun-Suk Min, Sumin Lee, Won Do Heo, YongKeun Park. Label free identification of cancer cell death pathways via holotomography and deep learning as an early pharmacodynamic biomarker abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4669.