PDO Drug Screening Data for Evaluating and Training AI Drug Response Prediction Models
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "PDO Drug Screening Data for Evaluating and Training AI Drug Response Prediction Models".
Marcus: Here's the gist of the paper. Anti-cancer drug response prediction (DRP) is critically important for precision oncology and therapeutic development These studies generated data for improving and evaluating DRP models, including multi-omics profiles of PDO lines, drug feature data, and treatment response measurements By systematically evaluating multiple model architectures and learning strategies across independently generated PDO datasets, the work characterizes the extent to which knowledge learned from large-scale CCL datasets transfers to PDOs and whether limited PDO data can improve predictive performance in the target domain Data Generation Strategy The research addressed a gap between existing DRP models developed using immortalized cancer cell lines (CCLs) and patient tumors by generating new drug screening data using patient-derived organoids (PDOs).
Ines: Four drug screening studies spanning colorectal cancer, head and neck cancer, and sarcoma PDOs were conducted to produce matched treatment response measurements and molecular profiles across multiple experimental platforms and assay protocols Three of these studies were designed with overlapping PDOs and compounds, enabling direct comparison of drug response measurements generated independently using different screening technologies and experimental procedures.
Marcus: Model Training Scenarios The study evaluated six publicly available DRP models under three learning strategies to assess how CCL and PDO data can be used for predicting PDO drug response.
Ines: These scenarios included: 1.
Paper summary: Ines: So, we're talking about this paper called "PDO Drug Screening Data for Evaluating and Training AI Drug Response Prediction Models". It's focused on making drug response prediction more accurate for actual patient tumors.
Marcus: Exactly. The core idea here is that most of the drug response prediction models out there are trained on cancer cell lines, which can be quite different from what we actually see in a patient's tumor at the genomic and molecular level <ref:2611.15769#pg2>.
Ines: Right. The paper claims they generated new data using patient-derived organoids, or PDOs, to bridge that gap between cell lines and tumors eight <ref:2611.15769#pg3,gap between cell lines and>. They set out to see how knowledge from the large cell line datasets transfers to these PDOs and if adding just a little bit of PDO data can actually improve the prediction performance in the real target domain.
Yuki: From a population genetics standpoint, that makes sense. We know that immortalized cell lines are just one snapshot, and patient tumors have all those different biological histories <ref:2611.15769#pg3>. I'm curious how much of that transfer learning actually works in practice for these complex systems.
Marcus: Well, they tested six different existing drug response prediction models using three different ways to train them. They looked at training models only on the large cell line data, then with a mix of cell line and some PDO data, and finally pretraining on the cell lines and then fine-tuning with the PDO data <ref:2611.15769#pg3>.
Ines: And they used different ways to check their results too, like random partitioning for cross validation, which is pretty standard, and cancer-blind partitioning for a stricter test of how well the model generalizes across different patient cases <ref:2611.15769#pg3>.
Yuki: That brings up a question about the sample size. When they talk about limited PDO data improving performance, how much is 'limited' actually limited in terms of what you can support genetically?
Marcus: The paper found that incorporating even a portion of the PDO data into the training improved the prediction performance, showing an increase in model-average R squared for PDO prediction across various datasets <ref:2611.15769#pg3>.
Ines: And they also looked at transfer learning—pretraining on cell lines and then fine-tuning on PDO data—and that generally improved the prediction performance compared to training them jointly when the PDO domain data was only a small fraction of the total training set <ref:2611.15769#pg3>.
Yuki: That sounds like a promising direction for building models that are more robust to those experimental differences between culture and in vivo settings. But what about the downside of that transfer learning? Does it introduce new problems?
Marcus: Yes, the authors mentioned that this transfer learning approach came at the cost of forgetting previously learned patterns from the cell lines <ref:2611.15769#pg3>. So you get better performance on PDOs, but you might lose some of what you learned about those cell lines.
Ines: That's a trade-off we see all the time in machine learning, right? We want to adapt to new data without completely erasing the valuable knowledge from the initial training set. But how robust are these findings when you look at specific datasets?
Yuki: I'm interested in those specifics. They mentioned that prediction performance on one particular dataset, the IMPROVE dataset two was consistently the lowest among all four drug screening studies <ref:2611.15769#pg3>.
Marcus: And they even found negative r squared values in some cases there <ref:2611.15769#pg3>. The authors attributed that to an issue with insufficient mixing of the cell culture medium, which is a very concrete experimental detail <ref:2611.15769#pg3>.
Ines: That’s a fair caveat. It shows that even with good methodology, biological variability and technical noise in the data collection process can still cause significant issues for the models <ref:2611.15769#pg3>.
Yuki: So, what does this mean for us looking at the broader landscape of drug response prediction? Does this suggest that focusing only on cell lines is fundamentally limiting our ability to predict patient outcomes?
Marcus: It suggests that while cell line data is an essential starting point because it's reproducible and easy to get large amounts of, it clearly doesn't capture everything about patient tumors <ref:2611.15769#pg2>. The utility of these new PDO datasets is that they help us develop and evaluate DRP models that are actually closer to predicting patient response eight <ref:2611.15769#pg3>.
Ines: So, the paper’s main contribution is demonstrating the utility of these newly generated datasets for developing and evaluating DRP models for PDO drug response prediction <ref:2611.15769#pg3>. It moves us from just using cell lines to testing how much we can learn from patient-like samples.
Yuki: I think it means that expanding and harmonizing these PDO drug screening datasets is an important step toward building DRP models that actually work for patients <ref:2611.15769#pg3>. We need more data like this to understand the true complexity of cancer response across different biological backgrounds.
Marcus: Right. And alongside that, we still need computational methods that explicitly account for those systematic differences between experimental platforms and drug response distributions <ref:2611.15769#pg3>. That's where the next challenge lies.
Conclusion: Ines: So, we've been looking at this paper about using patient derived organoids or PDOs to help train and test drug response prediction models.
Marcus: Yeah, the title is "PDO Drug Screening Data for Evaluating and Training AI Drug Response Prediction Models." It’s about taking those organoid data points and seeing if they actually make a difference in predicting how a real tumor will react to medicine.
Ines: Basically, the authors are showing how incorporating this patient-like data into existing cell line prediction models actually improves the performance numbers on PDO responses.
Marcus: They tested six different AI models and three ways of training them—just cell line data, a mix with some PDO data, or pretraining on cell lines and then fine tuning on the organoid stuff.
Yuki: From a population genetics view, this is interesting because it shows whether that limited set of patient samples can actually support a better prediction for the whole group of patients.
Ines: And they found that adding even a small amount of PDO data improves those model average R squared values across different datasets.
Marcus: Transfer learning, where you train on cell lines first and then fine tune with PDO data, generally beats just training everything together when the organoid data is only a small piece of the total training set.
Yuki: But they did flag that this transfer learning comes with a cost—the model starts forgetting some of those initial drug response patterns it learned from just the cell lines.
Ines: That trade-off between better accuracy and losing old knowledge is something we always have to think about in these systems.
Marcus: And they also found some specific limitations, like one dataset showed negative r squared values because the cell culture medium wasn't mixed enough, which shows how much technical stuff can mess with the data.
Yuki: So it sounds like the big implication is that we need to expand and harmonize these PDO screening datasets if we want DRP models that actually work for patients, but we also still need better ways to handle those experimental differences.
Ines: Exactly. So what does this mean for the world in terms of precision oncology?
Yitan Zhu, Kaitlyn Marlor, Vincent Lavelle, Thomas S Brettin, Matteo Sacco, Costanza Siniscalchi, Henry W Raeder, Michael Ryan Weil
Division of Data Science and Learning, Argonne National Laboratory · Computing, Environment and Life Sciences, Argonne National Laboratory Department of Medicine, Section of Hematology/Oncology, University of Chicago Department of Human Genetics, University of Chicago Cancer Research Technology Program Cancer Data Science Initiatives at Frederick National Laboratory for Cancer Research Department of Computer Science The University of Chicago
q-bio.QM, cs.AI
Submitted: 2026-11-11
Updated: 2026-11-11
Code: https://github.com/zhuyitan/PDO_DRP_Analysis
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 77/100
The gist: Anti-cancer drug response prediction (DRP) is critically important for precision oncology and therapeutic development These studies generated data for improving and evaluating DRP models…
Key concepts
- Anti-cancer Drug Response Prediction (DRP)
- This is the goal of the research: building AI models that can accurately predict how a specific drug will affect a patient's tumor, using molecular data from organoids (PDOs). This is crucial for precision oncology.
- Cancer Cell Lines (CCLs) vs. Patient-Derived Organoids (PDOs)
- CCLs are immortalized cell lines used in existing models, while PDOs are grown from actual patient tumors. The study aimed to see if knowledge learned from the more abundant CCL data could be effectively transferred to predict responses in the more biologically relevant PDOs.
- Transfer Learning
- This is a technique where a model first learns general patterns from a large, well-studied dataset (like CCL data) and then adapts those learned patterns to a smaller, specific dataset (like PDO data). The study found this method often boosted performance but risked 'forgetting' the initial general knowledge.
Terminology
Summary
Anti-cancer drug response prediction (DRP) is critically important for precision oncology and therapeutic development These studies generated data for improving and evaluating DRP models, including multi-omics profiles of PDO lines, drug feature data, and treatment response measurements By systematically evaluating multiple model architectures and learning strategies across independently generated PDO datasets, the work characterizes the extent to which knowledge learned from large-scale CCL datasets transfers to PDOs and whether limited PDO data can improve predictive performance in the target domain
Data Generation Strategy
The research addressed a gap between existing DRP models developed using immortalized cancer cell lines (CCLs) and patient tumors by generating new drug screening data using patient-derived organoids (PDOs). Four drug screening studies spanning colorectal cancer, head and neck cancer, and sarcoma PDOs were conducted to produce matched treatment response measurements and molecular profiles across multiple experimental platforms and assay protocols Three of these studies were designed with overlapping PDOs and compounds, enabling direct comparison of drug response measurements generated independently using different screening technologies and experimental procedures. p. 1 We conducted four drug screening studies
spanning colorectal cancer, head and neck cancer, and sarcoma PDOs, producing matched drug response
measurements and molecular profiles across multiple experimental platforms and assay protocols. Three of
these studies were designed with overlapping PDOs and compounds, enabling direct comparison of drug
response measurements generated independently using different screening technologies and experimental
procedures.Where the paper says it
Model Training Scenarios
The study evaluated six publicly available DRP models under three learning strategies to assess how CCL and PDO data can be used for predicting PDO drug response. These scenarios included:
-
training models exclusively on large-scale cancer cell line (CCL) data. p. 1
Where the paper says it
Six publicly available DRP models representing diverse machine learning and deep learning architectures were evaluated under three learning strategies. The first strategy trains models exclusively on large-scale cancer cell line (CCL) data and directly applies them to predict drug responses in PDOs.
-
training models on a combination of CCL data and a subset of PDO data. p. 1
Where the paper says it
The first strategy trains models exclusively on large-scale cancer cell line (CCL) data and directly applies them to predict drug responses in PDOs. The second strategy trains models on a combination of CCL data and a subset of PDO data, with predictive performance evaluated on the remaining PDO data.
-
pretraining models on CCL data followed by fine-tuning on PDO data. p. 6
Where the paper says it
Additional model-specific preprocessing, architecture, and training details are provided in Supplementary Information Section 1. 3 Analysis Schemes We built DRP models and evaluated their prediction performance on PDO response under three different scenarios designed to assess how CCL and PDO data can be used for predicting PDO drug response: (1) training exclusively on CCL data, (2) joint training on CCL and PDO data, and (3) transfer learning with pretraining on CCL data followed by fine-tuning on PDO data (Figure 2).
Performance Evaluation Metrics
Model performance was assessed using two cross-validation schemes: random partitioning, in which data are randomly split into training, validation, and test sets, and cancer-blind partitioning, in which cancer cases are mutually exclusive across these sets. Predictive performance was quantified by comparing predicted and true normalized AUC values using the coefficient of determination (R2) and the Pearson correlation coefficient (PCC). p. 6 An exception was the Pilot 1 PDO dataset
with 47 PDO lines, where 10-fold data partitioning with an 8:1:1 train, validation, and test ratio was also
applied in the cancer-blind cross-validation. Predictive performance was quantified by comparing predicted
and true normalized AUC values using the coefficient of determination (R2
) and the Pearson correlation
coefficient (PCC).Where the paper says it
Key Findings on Model Utility
The analysis demonstrated that incorporating a portion of the PDO data into model training improved PDO response prediction performance, showing an increase in model-average R2 for PDO prediction across various datasets. Transfer learning, which involves pretraining models on CCL data and subsequently fine-tuning them using PDO data, generally further improved PDO response prediction performance compared to joint training when the target domain data constituted only a small fraction of the overall training set. p. 6 Additional model-specific preprocessing, architecture, and training details are
provided in Supplementary Information Section 1. 3 Analysis Schemes
We built DRP models and evaluated their prediction performance on PDO response under three different
scenarios designed to assess how CCL and PDO data can be used for predicting PDO drug response: (1)
training exclusively on CCL data, (2) joint training on CCL and PDO data, and (3) transfer learning with
pretraining on CCL data followed by fine-tuning on PDO data (Figure 2). which knowledge learned from large-scale CCL datasets transfers to PDOs and whether limited PDO data
can improve predictive performance in the target domain. Together, these analyses demonstrate the utility of
the newly generated datasets for developing and evaluating DRP models for PDO drug response prediction.Where the paper says it
Where the paper says it
Limitations and Future Directions
The prediction performance on IMPROVE Dataset 2 was consistently the lowest among the four drug screening studies, with negative R2 values observed in some cases, which researchers attributed to an issue with insufficient mixing of the cell culture medium. The moderate levels of correlations observed between response measurements from different studies highlight a challenge for both drug screening experiments and PDO response modeling. p. 9 371, n = 280). All three correlations are at a moderate level, indicating a substantial variability remained between AUC measurements generated from studies using different experimental
protocols and assay technologies. In this work, we address this need by generating new PDO drug screening data specifically designed to
support the development and evaluation of AI-based DRP models. We conducted four drug screening studies
spanning colorectal cancer, head and neck cancer, and sarcoma PDOs, producing matched drug response
measurements and molecular profiles across multiple experimental platforms and assay protocols. Together, these analyses demonstrate the utility of
the newly generated datasets for developing and evaluating DRP models for PDO drug response prediction. 2 Materials and Methods
2.Where the paper says it
Where the paper says it
Where the paper says it
Data and Code Availability
The full dataset is available at https://doi.org/10.5281/zenodo.22920452. The code used for training and evaluating DRP models in this paper is available here: //github. p. 0 These studies generated data for improving and evaluating DRP models, including multi-omics profiles of
PDO lines, drug feature data, and treatment response measurements. Using these data, we retrained and
evaluated multiple publicly available DRP models under three model-building scenarios: training models
using cell line data only, training models using combined cell line and PDO data, and pretraining models
on cell line data followed by fine-tuning on PDO data. Using these data, we retrained and
evaluated multiple publicly available DRP models under three model-building scenarios: training models
using cell line data only, training models using combined cell line and PDO data, and pretraining models
on cell line data followed by fine-tuning on PDO data. We then evaluated and compared model prediction
performance on PDOs across these analysis scenarios.Where the paper says it
Where the paper says it
Acknowledgement
This research was part of the Innovative Methodologies and New Data for Predictive Oncology Model Evaluation (IMPROVE) project under the National Cancer Institute - Department of Energy (NCI-DOE) collaboration program that has been funded in whole or in part with federal funds from the National Cancer Institute, National Institutes of Health, Task Order No. 75N91019F00134 and from the Department of Energy under Award Number ACO22002-001-00006. p. 2 Table 1: Summary of datasets and experimental sources. The final row reports the total number of unique
experiments, drugs, cell lines, and organoids across all sources.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Transfer learning from cell lines to patient-derived organoids can improve prediction performance in target domains by fine-tuning models on PDO data, which is shown to be
more effective in improving PDO prediction performance than training models on combined PDO and CCL data.
-
Incorporating a portion of the PDO data into model training substantially improves response prediction performance, demonstrating that even though the target-domain data constituted only a small fraction of the overall training set,
they provided valuable information that was not captured by CCL screening data but was important for accurate PDO response prediction.
-
Models trained using joint CCL and PDO data jointly in random split cross-validations achieve a significantly higher model-average R2 for PDO prediction compared to those trained exclusively on CCL data, indicating that
incorporating PDO data directly into model training substantially improves PDO prediction performance.
-
The study identifies a challenge in reproducibility, noting that
Moderate levels of correlations in the range of [0.371, 0.645] are observed on response measurements from different studies,
suggesting future computational methods shouldexplicitly consider measurement variations
to improve generalizability across heterogeneous experimental settings. -
AI systems can be designed to explicitly account for measurement variations, which may help bridge the gap between experimental systems and patient tumors by addressing
variations of response measurements caused by differences in screening platforms may therefore limit model generalizability even when the underlying biological response is similar.
Abstract
Anti-cancer drug response prediction (DRP) is critically important for precision oncology and therapeutic development. Currently, most DRP models are developed using drug screening data from immortalized cancer cell lines, which can differ substantially from patient tumors at the genomic and molecular levels, even though patient tumor response is the ultimate prediction target. To help bridge this gap and bring DRP models closer to their intended clinical application, we conducted four drug screening studies using patient-derived organoids (PDOs), which more faithfully recapitulate patient tumors than immortalized cell lines. These studies generated data for improving and evaluating DRP models, including multi-omics profiles of PDO lines, drug feature data, and treatment response measurements. Using these data, we retrained and evaluated multiple publicly available DRP models under three model-building scenarios: training models using cell line data only, training models using combined cell line and PDO data, and pretraining models on cell line data followed by fine-tuning on PDO data. We then evaluated and compared model prediction performance on PDOs across these analysis scenarios. In addition, we also assessed the consistency of drug screening viability data generated by different laboratories using different assay technologies and/or experimental protocols.
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation