Abstract: It is widely supposed that the tissue specificity of gene expression indicates gene function. Now, an extensive analysis of gene expression in the mouse reveals that quantitative measurement of expression levels in different tissues can contribute powerfully to the prediction of gene function.
Introduction
For decades, biologists have relied on a simple heuristic: "where a gene is expressed tells us what it does." If a gene is active exclusively in the heart, it likely serves a cardiac function. However, this binary "on/off" approach is a blunt tool. Many genes are expressed across multiple tissues, and gene "knockouts" often fail to produce defects in every tissue where the gene is active. With the human and mouse genomes sequenced, the next frontier is functional genomics—mapping these thousands of genes to their specific physiological roles.
A landmark study by Zhang et al. proposes a more sophisticated paradigm. By treating quantitative expression levels across dozens of tissues as a high-dimensional "functional signature," researchers can predict the biological roles of thousands of previously uncharacterized genes with high confidence.
Methodology: From Binary Data to Multivariate Signatures
The researchers moved beyond simple presence/absence data by employing a multivariate approach centered on quantitative profiling:
- Quantitative Profiling: Using cDNA microarrays, they measured the precise mRNA levels of approximately 40,000 genes across 55 diverse mouse tissues, ranging from specific brain regions to various embryonic stages.
- Data Refinement: To eliminate noise, the dataset was filtered to 21,622 genes showing significant differential expression. These results were validated against known tissue markers and RT-PCR.
- Machine Learning Integration: The team utilized Support Vector Machines (SVMs) to analyze expression levels. By treating the 55 tissue measurements as a 55-dimensional vector, the SVMs could identify patterns—or "signatures"—associated with specific Gene Ontology (GO) biological process categories.
- Functional Inference: The model was trained on genes with established functions and then applied to predict the roles of unannotated genes.
Key Insights: The Importance of Quantity
The most significant finding is that the amount of expression is as informative as the location. Quantitative measurements capture nuances lost in binary data. A gene's function is often defined not by a single tissue, but by its relative activity across a broad spectrum of tissues.
As seen in the expression heatmap, genes belonging to the same GO category exhibit strikingly similar multivariate signatures. For instance, genes involved in "synaptic transmission" show a distinct profile across brain tissues compared to those involved in "muscle contraction." Interestingly, the data also suggests that some broad GO categories, such as "cell-cell adhesion," actually consist of multiple distinct expression sub-groups, indicating that our current functional classifications may need refinement.
Results and Predictive Success
The predictive power of this quantitative approach is substantial:
- Functional Inference: The model successfully recovered the known functions of 7,387 genes already present in the Gene Ontology.
- Discovery of New Functions: More importantly, it predicted physiological roles for 12,123 previously unannotated genes.
- High-Confidence Candidates: A subset of 1,092 genes was identified with precision scores exceeding 50%. These represent high-priority targets for experimental validation, many of which have already been supported by subsequent analysis of protein domains and existing literature.
Limitations and Future Directions
While the use of SVMs proved effective, future research might explore simpler, more interpretable models like Linear Discriminant Analysis to pinpoint exactly which tissues provide the most functional information. Additionally, identifying a "minimal set" of tissues could streamline future large-scale profiling efforts.
The ultimate goal is to extend this quantitative signature approach across species. By comparing functional signatures between humans, mice, and other models, researchers can begin to understand how gene functions evolve and how subtle changes in expression patterns contribute to complex physiological differences.
Summary updated