TL;DR
This paper tackles the challenge of missing data in microRNA (miRNA) expression profiles, which hinders accurate analysis. They introduce a novel imputation method that leverages Gene Ontology (GO) functional similarity derived from miRNA target genes, outperforming traditional expression-based methods on breast cancer miRNA datasets.
What Problem Does This Paper Address?
Microarray data, including miRNA expression profiles, often suffers from missing values. These gaps can arise from experimental limitations, noise, or preprocessing errors. Simply discarding genes or samples with missing data can lead to a loss of valuable information and biased analyses. Therefore, accurately imputing these missing values is crucial for reliable downstream analysis, such as identifying disease biomarkers. While imputation methods exist, applying domain knowledge to miRNA data is challenging because miRNAs themselves lack direct functional annotations. This paper addresses this gap by proposing a method to infer miRNA functional similarity from the known functional annotations of their target genes.
Methods
The core of this paper is a new imputation method that combines traditional expression-based similarity with biologically-derived functional similarity. Here's a breakdown:
- Expression-based Similarity: The paper starts with the widely used K-Nearest Neighbors (KNN) algorithm. KNN measures the similarity between genes based on the Euclidean distance of their expression levels across samples.
- GO-based Functional Similarity: This is the novel contribution.
- Target Gene Identification: They first identify target genes for each miRNA using computational prediction tools (like miRanda) and public databases.
- GO Annotation Retrieval: For each target gene, they retrieve its associated Gene Ontology (GO) terms, which describe gene functions, biological processes, and cellular components.
- GO Term Similarity: They compute the semantic similarity between GO terms using the G-SESAME method, which considers the structure of the GO graph.
- Gene Functional Similarity: The similarity between two genes is then calculated based on the GO term similarities of their respective target genes. They employ the Best Match Average (BMA) strategy for this integration.
- miRNA Functional Similarity: Finally, the functional similarity between two miRNAs is derived from the functional similarities of their target genes.
- Combined Similarity Measure: A new distance metric is defined by combining the GO-based functional similarity and the Euclidean distance of expression profiles. This new measure is then used within the KNN imputation framework.
Datasets
The study evaluated their method on two publicly available miRNA microarray datasets from breast cancer research:
- GSE26659: Measured using Agilent-019118 Human miRNA Microarray 2.0.
- GSE40525: Also measured using Agilent-019118 Human miRNA Microarray 2.0.
These datasets contain probes for human and viral microRNAs. The authors simulated missing values by randomly selecting miRNAs and assigning missing elements within their profiles, creating datasets with varying percentages of missing data.
Key Insights and Ideas
- Bridging the Annotation Gap: The most significant insight is that functional similarity of miRNAs can be effectively inferred indirectly through their target genes and the well-established GO annotation system. This circumvents the lack of direct miRNA functional annotations.
- Complementary Information: Biological knowledge, even when indirectly derived, acts as a valuable complement to expression data. It helps to refine similarity measures and improve imputation accuracy.
- miRNA Data Characteristics: The study notes that miRNA expression data might inherently have lower noise and smaller scale compared to mRNA data, potentially leading to better performance of conventional imputation methods on miRNA data than previously observed for mRNA data.
- Impact of K: The optimal number of neighbors (K) for KNN imputation in miRNA data appears to be smaller (around 4 or 5) compared to mRNA data.
Key Results
- The proposed GO-based imputation method consistently achieved lower Normalized Root Mean Square Error (NRMSE) compared to conventional KNN and Local Least Squares (LLS) methods on both GSE26659 and GSE40525 datasets.
- The improvement was more pronounced for KNN than for LLS, suggesting that the GO-based similarity measure has a stronger impact on neighbor selection in KNN.
- The GO-based method demonstrated improved performance across various maximum missing rates, indicating its robustness.
Limitations and Future Work
- Simulation Strategy: The authors acknowledge that the simulation strategy for generating artificial missing values, while more realistic than completely random models, could be further refined to better mimic real-world missing data distributions.
- Integration Strategy: The method for integrating pairwise GO similarity into miRNA similarity (BMA) is one approach, and other strategies could be explored.
- Annotation Incompleteness: The inherent incompleteness of GO annotations and potential inaccuracies in computational target gene prediction are acknowledged limitations.
- Future Directions: The paper suggests exploring other biological knowledge sources (e.g., miRNA-associated diseases) and refining the integration strategies to further enhance imputation quality. The authors also highlight the potential for their method to be applied to miRNAs from other species.
Summary updated