Workshops Archive

PDO Drug Screening Studies for Evaluating and Training AI Drug Response Prediction Models


Workshop: The 12th Computational Approaches for Cancer Workshop (CAFCW26)

Authors: Yitan Zhu, Kaitlyn Marlor, Vincent Lavelle, Thomas Brettin, and Rick Stevens (Argonne National Laboratory)

Abstract: Background: Anti-cancer drug response prediction (DRP) is important for precision oncology and therapeutic development. Most DRP models are trained on drug screening data of immortalized cancer cell lines, which differ substantially from patient tumors in genomic and molecular characteristics and treatment response. Patient-derived organoids (PDOs) better recapitulate patient tumors, but PDO drug screening data remain scarce.

Methods: We conducted four drug screening studies using PDOs from colorectal cancer, head and neck cancer, and sarcoma, generating matched treatment responses and omics data with multiple screening technologies. The four datasets contain: (1) 27 colorectal and 20 head and neck cancer PDO lines, 37 compounds, and 1,316 experiments (pairs of PDOs and compounds); (2) 4 colorectal PDO lines, 100 compounds, and 400 experiments; (3) 5 colorectal PDO lines, 40 compounds, and 200 experiments; and (4) 9 colorectal and 1 sarcoma PDO lines, 100 compounds, and 1,000 experiments. Screening technologies included multiplane high-content imaging using Rhodamine-Phalloidin, Caspase-3/7, To-Pro-3, and Hoechst staining, and ATP-based CellTiter-Glo assays. Overlapping PDO–compound experimental pairs across three of the studies enabled cross-platform correlation analyses to evaluate the reproducibility of PDO drug screening experiments. Using PDO response data, gene expressions, and drug molecular features, we evaluated six DRP algorithms (DeepTTC, UNO, GraphDRP, HiDRA, IGTD, and LightGBM) under two scenarios: (1) training models on cell-line drug screening data including 465,329 experiments and testing on the four PDO datasets; and (2) training models on cell-line data plus part of the PDO data and testing on the held-out PDO data. Both random-split and cancer-blind cross-validations were used; cancer-blind splits shared no cancer models across training, validation, and test sets.

Results: For shared experiments across studies, response measurements showed moderate agreement, with Pearson correlation coefficients of 0.371–0.645. In scenario 1, mean R² across the six algorithms for the four PDO datasets was 0.07, 0.25, -0.26, and 0.24 with random-split cross-validation, and 0.11, 0.25, -0.25, and 0.27 with cancer-blind cross-validation, indicating limited transferability of models trained only on cell-line data to PDOs. In scenario 2, mean R² improved to 0.73, 0.72, 0.55, and 0.71 with random splits, demonstrating a substantial benefit from incorporating PDO data in training. Under cancer-blind validation, scenario 2 achieved R² values of 0.57, 0.48, -1.40, and 0.58, improving prediction performance on three datasets compared with scenario 1.

Conclusion: We provide four AI-ready PDO drug screening datasets containing treatment responses, PDO multi-omics, and drug molecular features for the development and benchmarking of DRP models. Response measurements from different studies and technologies showed moderate consistency, supporting cross-study comparison while revealing platform variability. Across six algorithms and two validation schemes, models trained with both cell-line and PDO data generally outperformed models trained on cell-line data alone, showing the importance of including PDO data in training when predicting PDO responses. Lower and more variable performance under cancer-blind validation highlights the challenge of generalizing across cancer models. Future work will investigate transfer learning approaches that pre-train on large cell-line data and fine-tune on limited PDO data to improve generalization to new PDOs.


Back to The 12th Computational Approaches for Cancer Workshop (CAFCW26) Archive Listing

Back to Full Workshop Archive Listing