TL;DR
This paper introduces a novel approach to breast cancer prognosis by analyzing the activity of signaling pathways, rather than just individual gene expression. They found that these pathway-level "mechanistic biomarkers" significantly outperform traditional gene-based signatures, offering a more interpretable and accurate way to predict patient outcomes.
What Problem Does This Paper Address?
Predicting breast cancer prognosis and response to treatment is crucial for personalized medicine. While gene expression signatures have shown promise, they often lack clear biological interpretation, making it difficult to understand why a particular signature predicts outcome. This "black box" nature hinders the translation of these signatures into actionable clinical insights. The authors aim to bridge this gap by leveraging the known roles of signaling pathways in cancer progression to create more interpretable and effective prognostic biomarkers.
Methods
The core of this study lies in the application of the Canonical Circuit Activity Analysis (CCAA) method, implemented in the hipathia R package. This method transforms gene expression data into "signaling pathway activity" at different levels: individual circuits, effector circuits, and cellular functions.
Here's a breakdown of the approach:
- Data Preprocessing: RNA-seq counts from the TCGA breast cancer dataset were normalized and batch effects were corrected.
- Pathway Modeling: KEGG pathways were used to define signaling circuits. A circuit is essentially a series of connected nodes (genes/proteins) representing how a signal propagates from an input to an output.
- Activity Calculation: The CCAA method models signal flow through these circuits. The "activity" of a circuit is determined by the signal intensity reaching its final effector protein. This activity is then aggregated for effector circuits and further for cellular functions based on GO annotations.
- Prognostic Feature Generation: The calculated pathway activities were used to create various "profiles" (e.g., circuit activity, effector circuit activity, function activity). These profiles, along with gene expression profiles (both pathway-related and non-pathway-related genes), were used as predictive features.
- Classification: A suite of 12 state-of-the-art classification algorithms were tested for their ability to predict patient survival (deceased vs. alive) using these different profiles. A 5-fold cross-validation repeated 10 times was employed to assess performance, with the Area Under the ROC Curve (AUROC) as the primary metric.
Datasets
The primary dataset used in this study is the The Cancer Genome Atlas (TCGA) breast cancer dataset. This dataset provided RNA-seq counts for 18,708 genes from 879 tumor samples, along with vital status information for the corresponding donors. No new benchmarks were introduced; the study utilized existing TCGA data.
Key Insights and Ideas
- Pathway Activity as a Superior Biomarker: The most significant finding is that pathway-level features (signaling circuit activities) provide better prognostic performance than individual gene expression levels. This suggests that understanding how genes interact within functional pathways offers a more robust signal for predicting cancer outcomes.
- Mechanistic Interpretability: Unlike traditional gene signatures, pathway activity biomarkers are inherently more interpretable. They directly relate to cellular functions and biological processes known to be involved in cancer progression, offering insights into the "why" behind the prognosis.
- Beyond Signaling Proteins: The study also demonstrated that combining signaling circuit activity with the expression of genes not involved in signaling can further enhance the predictive model. This suggests that non-signaling genes can play complementary roles in cancer progression that are not captured by pathway analysis alone.
- Efficiency of Pathway Features: A notable observation is that a small subset of genes involved in signaling pathways (represented by circuit activity) can achieve similar predictive performance to using all genes. This highlights the power of recoding gene expression into pathway activities, suggesting a more efficient way to capture relevant biological information.
Key Results
- Pathway Activity Outperforms Gene Expression: The
path.valsprofile (representing signaling circuit activities) achieved the highest AUROC scores, outperforming gene-based profiles likepath.genes.vals,other.genes.vals, andgenes.vals. - Specific Pathway Circuits Identified: The study identified specific signaling circuits (e.g., from the HIF-1 and NF-kappa B pathways) that are highly relevant for breast cancer prognosis, linking them to known cancer hallmarks like growth/survival and inflammatory responses.
- Non-Signaling Genes Contribute: While not significantly improving performance over pathway-only features, the combination of pathway activity and non-signaling gene expression (
path.and.other.genes.vals) identified genes like ABCB5 and LMO4, which are implicated in drug resistance, cell cycle progression, and tumor suppression, offering a broader mechanistic view. - Support Vector Machines Dominate: Support Vector Machines (SVMs) with various kernels were consistently identified as the best-performing classifiers across different profile types, indicating their robustness for both gene-level and pathway-level features.
Limitations and Future Work
- Coverage of Pathways: The authors acknowledge that the prognostic power of pathway-level profiles might be limited by the coverage of genes within known biological pathways. Not all genes or biological processes are fully represented in current pathway databases.
- Simplistic Function Averaging: The performance of function-based profiles (
fun.vals,go.vals) was slightly worse than other pathway-level profiles. This is attributed to the simplistic averaging of signaling activity values based on GO annotations, which can be incomplete or ambiguous. - Future Directions: The study suggests that further investigation of the top-ranked "other-genes" (genes not in signaling pathways) could reveal new cancer drivers and potential therapeutic targets. The authors also emphasize the potential of their approach for personalized medicine by creating patient-specific mechanistic biomarkers. The
hipathiatool and associated R package are made available for further research.
Summary updated