TL;DR
This paper introduces the Pan-PreClinical (PPC) project, a massive atlas of ex vivo drug screening data from nearly 2,000 samples and 3,000 drugs across 134 cancer types. They developed a novel Bayesian model to harmonize this diverse data, revealing tissue-specific drug sensitivities and uncovering systematic biases in traditional cell line models. This work lays the groundwork for a foundation model to accelerate functional precision oncology.
What Problem Does This Paper Address?
Developing effective cancer therapies often relies on understanding how drugs affect tumors. While traditional cell lines are easy to work with, they don't always accurately reflect how real patient tumors will respond. Patient-derived ex vivo models (like organoids and xenografts) are more representative but are expensive and complex to generate and screen at scale, hindering comprehensive pan-cancer analyses. This lack of large, harmonized ex vivo datasets limits our ability to build predictive models for drug response and discover new therapeutic strategies. The PPC project aims to overcome these limitations by creating a unified, large-scale atlas of ex vivo drug screening data.
Methods
The core of the PPC project is a novel contrastive Bayesian tensor factorization model. This model is designed to handle the heterogeneity of data from different studies, experimental conditions, and drug concentrations. It uses contrastive learning to regularize embeddings for both samples and drugs, leveraging annotations like primary site and drug mechanism. This allows for harmonizing drug names, predicting missing dose-response data, and enabling pan-cancer analyses. They also developed a foundation model based on a BERT-like architecture for efficient "few-shot" prediction of drug responses, meaning it can make accurate predictions with very limited new data.
Datasets
The paper introduces the Pan-PreClinical (PPC) dataset, a comprehensive collection of ex vivo drug screening data. It includes:
- 2.1 million experiments
- 1,982 ex vivo samples (patient-derived cells, organoids, xenografts)
- 3,100 drugs
- Data from 134 cancer indications across 26 studies. This dataset is integrated with genomic, transcriptomic, and clinical data where available. They also integrated 12 large pan-cancer cell line datasets for comparative analysis.
Key Insights and Ideas
- Ex vivo models are crucial but hard to scale: The paper strongly advocates for ex vivo models over cell lines for better patient response prediction, but highlights the significant challenges in creating large, harmonized datasets.
- Cell lines have systematic biases: Comparing ex vivo data to cell line data revealed that cell lines tend to favor drugs targeting proliferation and undervalue drugs targeting cell-cell communication, a critical insight for interpreting existing cell line screening data.
- Harmonization is key: The novel Bayesian tensor factorization model is essential for integrating diverse ex vivo datasets, enabling robust pan-cancer analyses that were previously impossible.
- Foundation models for ex vivo data: The development of a few-shot foundation model demonstrates a path towards more efficient and scalable drug discovery using ex vivo models, even with limited sample availability.
- Drug sensitivities predict molecular profiles: The study shows that drug screen results can predict underlying genomic and transcriptomic features of tumors, linking functional response to molecular mechanisms.
Key Results
- The PPC dataset harmonized data from 26 studies, resulting in 327,335 unique dose-response curves for 3,151 drugs.
- The model identified 303 tissue-specific drug sensitivities.
- The PPC foundation model achieved 100-400% higher correlation than baseline models in few-shot learning scenarios, especially when data is limited.
- The model's embeddings could predict gene mutation status with AUROC up to 0.83 and pathway activity scores with high accuracy.
- Cell lines were found to be significantly more sensitive to cell-cycling drugs (234 out of 700 significant differences) while ex vivo samples were more sensitive to drugs targeting intercellular signaling pathways.
Limitations and Future Work
- Data Sparsity: The degree of matched omics data (genomics, transcriptomics) varied widely across studies, leading to a sparse dataset that limited the depth of mechanistic analysis.
- Treatment History: Complete treatment history for samples was not always available, making it difficult to disentangle the effects of disease progression, metastasis, and prior treatments on drug response.
- Cell Viability Assays: Bulk cell viability assays can't always distinguish between proliferative arrest and cell death, and gathering time-point data or tracking single cells could improve interpretation.
- Emerging Ex Vivo Platforms: The authors anticipate incorporating data from newer ex vivo assays that better model the tumor microenvironment (TME), such as co-cultures and microfluidic systems, to further enhance the PPC dataset's translational value.
- Scalability: While the foundation model improves efficiency, generating large-scale ex vivo datasets remains a challenge. Future work could focus on optimizing data generation and integration strategies.
Summary updated