TL;DR
This paper introduces "relevance networks," a novel method for uncovering functional relationships between gene expression and drug susceptibility in cancer cell lines. By analyzing correlations between thousands of genes and anticancer agents, they identify potential biological links and generate hypotheses for drug response, offering a more nuanced approach than existing clustering techniques.
What Problem Does This Paper Address?
Understanding how gene expression influences a cancer cell's response to chemotherapy is a major challenge in personalized medicine. Traditional methods often focus on single genes or limited pathways, failing to capture the complex, interconnected nature of biological systems. Researchers need ways to identify patterns across vast datasets of gene expression and drug sensitivity to generate testable hypotheses about why some cells respond to certain drugs while others don't. This paper aims to bridge this gap by developing a method that can discover these complex relationships and distinguish true biological signals from random noise.
Methods
The core of this paper is the development and application of relevance networks. This methodology involves:
- Data Integration: Combining two large datasets: baseline RNA expression levels for 7,245 genes across 60 cancer cell lines, and the sensitivity (measured by GI50, the dose required for 50% growth inhibition) of these same cell lines to 5,084 anticancer agents.
- Feature Filtering (Entropy-based): Removing features (genes or agents) with very low entropy. Low entropy indicates a limited range of values, which can artificially inflate correlation coefficients and lead to spurious associations. This step helps to ensure that the remaining features have enough variability to reveal meaningful patterns.
- Pairwise Correlation Calculation: Calculating a modified correlation coefficient, denoted as $\hat{r}^2$, for every possible pair of features (gene-gene, agent-agent, and gene-agent). This $\hat{r}^2$ retains the sign of the original correlation coefficient ($r$), allowing for the identification of both positive and negative relationships.
- Thresholding and Network Construction: Applying a stringent threshold to the absolute value of $\hat{r}^2$ (abs($\hat{r}^2$)). Only pairs with abs($\hat{r}^2$) above this threshold are considered significant and form connections (edges) in the network. Genes and agents are represented as nodes.
- Statistical Validation (Permutation Testing): To rigorously assess the significance of the identified associations and avoid false positives due to the sheer number of comparisons, the authors performed extensive permutation testing. They randomly shuffled the data 100 times for each feature and recalculated correlations. Associations that were highly unlikely to occur by random chance (abs($\hat{r}^2$) $\ge$ 0.80) were deemed statistically significant. This rigorous validation is a key differentiator from some previous approaches.
The authors highlight that relevance networks are advantageous over methods like phylogenetic-type tree clustering and self-organizing maps because they can:
- Discover both positive and negative correlations.
- Cluster disparate biological measures (e.g., gene expression and drug susceptibility).
- Allow features to participate in multiple relationships, reflecting biological reality.
Datasets
- NCI60 Cancer Cell Line Data:
- Gene Expression Data: Baseline RNA expression levels for 7,245 genes (from 6,416 probe sets on Affymetrix HU6800 arrays) measured in 60 human cancer cell lines.
- Anticancer Agent Susceptibility Data: Sensitivity data for 5,084 anticancer agents tested against the NCI60 panel, quantified by the GI50 (dose needed for 50% growth inhibition). Susceptibility was expressed as the negative logarithm of GI50.
These datasets were integrated to form a combined dataset of 12,329 features.
Key Insights and Ideas
- Uncovering Hidden Biological Relationships: The primary insight is that by systematically correlating gene expression with drug response across a large panel of cell lines, one can uncover novel, potentially functional relationships that are not apparent through other methods.
- The Power of Stringent Thresholding and Validation: The paper emphasizes that a high correlation threshold (abs($\hat{r}^2$) $\ge$ 0.80) combined with rigorous permutation testing is crucial for distinguishing true biological signals from the vast amount of noise inherent in high-throughput data. This approach helps generate more reliable hypotheses.
- Beyond Single-Gene Explanations: The relevance network approach moves beyond studying genes in isolation, allowing for the visualization and analysis of interconnected networks of genes and drugs, reflecting the complex biological pathways involved in drug response.
- Bridging Disparate Biological Measures: The methodology successfully links molecular data (gene expression) with phenotypic data (drug susceptibility), offering a powerful tool for systems biology.
Key Results
- After filtering and applying the stringent threshold (abs($\hat{r}^2$) $\ge$ 0.80), 202 relevance networks were constructed, comprising 834 features (genes and agents) and 1,222 associations.
- The vast majority of these associations were between pairs of anticancer agents (suggesting relationships based on chemical structure or mechanism of action) or pairs of genes (suggesting functional or sequence similarity).
- Crucially, one strong association was found between a gene and an anticancer agent: the gene Lymphocyte cytosolic protein-1 (LCP1) and the anticancer agent NSC 624044 (a thiazolidine carboxylic acid derivative). This association had a calculated $\hat{r}^2$ of 0.83.
- The authors categorized associations into three types: identity/synonymy (e.g., different probe sets for the same gene), derivation (e.g., chemically related anticancer agents), and biologic relationship. They found numerous examples of the first two types, serving as positive controls, and identified nine networks suggesting potential biological relationships.
Limitations and Future Work
- Exclusion of Low-Entropy Features: The paper acknowledges that by excluding the 5% of features with the lowest entropy, they may have missed some valid hypotheses, particularly those involving genes or agents that act uniquely in a single cell line or have a very narrow dynamic range.
- Potential for Missed Hypotheses: The stringent threshold, while minimizing false positives, might also lead to missing weaker but still biologically relevant associations. The authors suggest that with more data or cell lines, these weaker associations might become statistically significant.
- No Modeling of Measurement Noise: The analysis does not explicitly model noise in the RNA expression measurements, which can be more significant at lower expression values.
- Future Directions:
- Investigating Exceptions: Studying cases that violate identified model associations could reveal important exceptions.
- Expanding to Clinical Features: Integrating clinical features alongside gene expression and drug susceptibility data.
- Experimental Validation: The most critical future step is to experimentally validate the generated hypotheses, which will require improved genome annotation and biological databases.
- Exploring Alternative Metrics: Investigating metrics like cross-entropy as an alternative to correlation coefficients.
Summary updated