TL;DR
This paper introduces PRECISE, a novel domain adaptation method that bridges the gap between drug response predictions from pre-clinical cancer models (cell lines, PDXs) and actual human tumors. By learning a shared representation, PRECISE enables more reliable translation of findings, successfully recovering known biomarker-drug associations in human tumors.
What Problem Does This Paper Address?
Developing effective cancer treatments relies on understanding how tumors respond to drugs. Pre-clinical models like cell lines and patient-derived xenografts (PDXs) are invaluable for generating large-scale drug response data. However, these models, while sharing core biological processes with human tumors, also possess distinct characteristics (e.g., lack of immune system, different microenvironments). This "domain shift" makes it challenging to directly apply drug response predictors trained on pre-clinical data to predict patient outcomes. The paper addresses this critical need for methods that can efficiently and reliably transfer knowledge from abundant pre-clinical data to the more limited, but clinically relevant, human tumor data.
Methods
The core of PRECISE is a domain adaptation strategy. It begins by independently extracting "domain-specific factors" (principal components) from both pre-clinical and human tumor transcriptomic data. The key innovation lies in how these factors are aligned and combined:
- Subspace Alignment: A linear transformation is used to align the pre-clinical factors with the tumor factors, quantifying their similarity.
- Principal Vectors (PVs): Singular Value Decomposition (SVD) is applied to this transformation to identify "principal vectors" – pairs of directions that capture shared information between the two domains. These PVs are ranked by their similarity, with the top pairs representing the most conserved biological processes.
- Consensus Representation: Instead of directly using source or target PVs, PRECISE interpolates between them. This interpolation is guided by minimizing the Kolmogorov-Smirnov (KS) statistic between the source and target data projected onto the interpolated features, aiming to find a "sweet spot" that balances information from both domains. This creates a "consensus representation."
- Regression Model: A drug response predictor (using Ridge regression) is then trained on the pre-clinical data projected onto this consensus representation. This predictor is subsequently applied to human tumor data.
This approach differs from simpler methods by explicitly accounting for the domain shift through a principled alignment and interpolation process, rather than just correcting for batch effects.
Datasets
The study utilizes several large-scale datasets:
- Cell Line Dataset: GDSC1000, containing gene expression profiles and drug sensitivity (IC50 values) for over 1,000 cell lines.
- PDX Dataset: Novartis PDXE, providing gene expression profiles for nearly 400 patient-derived xenografts.
- Human Tumor Dataset: Gene expression profiles, mutation data, and copy number aberrations from The Cancer Genome Atlas (TCGA) for breast and skin melanoma cancers.
These datasets are standard in cancer genomics research, and the paper leverages their scale to train and validate its domain adaptation approach.
Key Insights and Ideas
- Quantifiable Domain Shift: The paper rigorously demonstrates that pre-clinical models and human tumors, while sharing some biological signals, are not directly interchangeable. The variance explained by pre-clinical principal components in human tumor data is significantly lower than that explained by tumor-specific components, underscoring the need for adaptation.
- Principal Vectors as Shared Mechanisms: The identified Principal Vectors (PVs) effectively capture common biological processes. Top PVs show high similarity and enrichments in pathways like the cell cycle, expected to be conserved. Less similar PVs capture processes like immune response, which are less conserved due to model system differences.
- Interpolation for Robustness: The consensus representation, achieved through interpolation between source and target PVs, is crucial. It balances domain-specific information, leading to predictors that are less performant on the source domain but more reliable on the target (human tumor) domain.
- Validation via Biomarker Recovery: The most compelling insight is PRECISE's ability to recover known, independent biomarker-drug associations (e.g., ERBB2 amplification and Lapatinib response, BRAFV600E mutation and Trametinib response) in human tumors, a feat that simpler methods struggle with.
Key Results
- Reduced but Competitive Performance: PRECISE achieves a slight reduction in predictive performance on pre-clinical data (median relative reduction of 3.9% in Pearson correlation) compared to a direct Ridge regression model. However, this is a trade-off for improved transferability.
- Successful Biomarker Recovery: PRECISE reliably recovers known associations between biomarkers and drug response in human tumors. For instance, it shows a significant correlation between ERBB2 copy number and predicted Lapatinib sensitivity (Spearman correlation $\rho = -0.32$) and differentiates Trametinib response based on BRAFV600E mutations.
- Superior to Baseline Methods: Compared to standard Ridge regression (with or without batch correction like ComBat), PRECISE demonstrates a more consistent ability to recover these critical biomarker-drug associations, particularly for Trametinib response.
Limitations and Future Work
- Transcriptomics Focus: The current study is limited to transcriptomics data. The authors acknowledge that other molecular data types (e.g., genomics, proteomics) also contribute to drug response and suggest that a multi-omic domain adaptation approach would be beneficial.
- Choice of Dimensionality Reduction: The method's performance can depend on the initial dimensionality reduction technique (PCA was used here). Exploring other methods that might capture biological signals more effectively is a potential avenue.
- Hyperparameter Tuning: While the paper details a robust cross-validation strategy for hyperparameter tuning (e.g., shrinkage coefficient in regression), the selection of the number of principal components and principal vectors still involves some empirical choices.
Future work could involve extending PRECISE to multi-omic data, exploring alternative dimensionality reduction and interpolation strategies, and applying it to a wider range of cancer types and drugs.
Summary updated