Leveraging TCGA gene expression data to build predictive models for cancer drug response

Evan A. Clayton, Toyya A. Pujol, John F. McDonald, Peng QiuBMC Bioinformatics 2020 · DOI

TL;DR

This paper leverages gene expression data from The Cancer Genome Atlas (TCGA) to build machine learning models that predict patient response to chemotherapy drugs like 5-Fluorouracil (5-FU) and Gemcitabine (GCB). The models achieve up to 86% accuracy, demonstrating that primary tumor gene expression is a valuable predictor of drug response and can aid oncologists in treatment decisions.

What Problem Does This Paper Address?

Personalized medicine aims to tailor cancer treatments to individual patients based on their unique genetic makeup. However, predicting a patient's response to a specific drug is challenging due to the inherent molecular heterogeneity of cancer. While previous efforts have relied on cell line data, these often struggle to accurately translate to in vivo patient responses due to differences in genetic and environmental factors. The availability of comprehensive pan-cancer datasets from TCGA, including both gene expression profiles and clinical drug response data from primary tumor samples, presents a novel opportunity to build more accurate and clinically relevant predictive models for drug response.

Methods

The researchers employed a two-stage machine learning approach:

  1. Clustering for Dimensionality Reduction: They used the optCluster R package to identify optimal clustering algorithms and the ideal number of clusters. This process groups genes with similar expression patterns into "gene modules," significantly reducing the number of features while aiming to minimize information loss. This is a key step to handle the high dimensionality of gene expression data.
  2. Classification for Prediction: The random forest algorithm was used for classification. Random forest is an ensemble method that builds multiple decision trees, making it robust against overfitting, a common issue in machine learning. It also provides feature importance rankings, allowing the identification of genes most predictive of drug response.

They compared various clustering and classification methods, ultimately finding Clara clustering combined with random forest to be the most effective. Demographic data was also tested but did not improve accuracy.

Datasets

The study utilized data from The Cancer Genome Atlas (TCGA). This dataset includes:

  • Gene expression profiles: Specifically, Upper-Quartile Normalized Fragments per Kilobase per Million mapped reads (UQ-FPKMs) from patients' primary tumor samples.
  • Clinical trial data: This includes information on drugs administered, patient demographics, temporal data, and patient response to treatment.

The study focused on two drugs, 5-Fluorouracil (5-FU) and Gemcitabine (GCB), due to the availability of the largest number of patient records for these treatments across multiple cancer types.

Key Insights and Ideas

  • In Vivo Data is Crucial: The study strongly supports the idea that using primary tumor gene expression data (in vivo) is more effective for predicting drug response than relying solely on cell line data (in vitro).
  • Gene Modules as Predictors: Grouping genes into biologically relevant modules through clustering is an effective strategy for reducing dimensionality and improving predictive model performance.
  • Interpretability Matters: The models not only achieve high accuracy but also identify biologically relevant genes and pathways, offering insights into the mechanisms of drug response and potential new therapeutic targets.
  • Drug-Specific Mechanisms: The observed differences in model performance between 5-FU and GCB suggest that some drugs target shared cancer mechanisms, while others are specific to certain cancer types or families.

Key Results

  • High Prediction Accuracy: The developed machine learning models achieved prediction accuracies of up to 86%.
  • GCB Model Outperforms 5-FU: The Gemcitabine (GCB) model generally performed better, with a cross-validation accuracy of 82.3% and a model validation accuracy of 85.7%, compared to the 5-FU model's cross-validation accuracy of 84.1% and a significantly lower model validation accuracy of 52.9%.
  • Model Validation Challenges: The 5-FU model's performance dropped considerably during independent validation, particularly on stomach adenocarcinoma (STAD) data, suggesting that pan-cancer models may not perform equally well across all cancer types.
  • Enriched Pathways: Gene set enrichment analysis revealed that predictive gene modules were enriched in pathways related to drug metabolism, inflammation (chemokine and cytokine signaling), T cell activation, and WNT signaling, highlighting their biological relevance to drug response.

Limitations and Future Work

  • Sample Size: The authors acknowledge that the study's limitation is the sample size, particularly for the 5-FU model, which impacted its validation performance. They suggest that larger datasets are needed to further refine these models.
  • Cancer Type Specificity: The significant drop in accuracy for the 5-FU model on STAD data indicates that pan-cancer models might struggle with cancer types that have dissimilar drug response mechanisms. Future work could explore building models for molecularly similar cancers of the same histology or anatomy.
  • Extrapolation from Cell Lines: The paper highlights that extrapolating human drug responses from cell line data is a significant challenge, reinforcing the value of in-vivo data.
  • Future Directions: The authors envision that these predictive models will become increasingly powerful and aid oncologists in making critical treatment decisions, potentially guiding the selection of first or second-line therapies. They also suggest that feature ranking can identify biologically relevant genes that may lead to new therapeutic targets.

Summary updated

BibTeX
@article{clayton2020leveraging,
  author = {Evan A. Clayton and Toyya A. Pujol and John F. McDonald and Peng Qiu},
  title = {{Leveraging TCGA gene expression data to build predictive models for cancer drug response}},
  journal = {BMC Bioinformatics},
  year = {2020},
  doi = {10.1186/s12859-020-03690-4},
}
Figure