Abstract
In the past, microarray studies have been criticized due to noise and the limited overlap between gene signatures. Prior biological knowledge should therefore be incorporated as side information in models based on gene expression data to improve the accuracy of diagnosis and prognosis in cancer. As prior knowledge, we investigated interaction and pathway information from the human interactome on different aspects of biological systems. By exploiting the properties of kernel methods, relations between genes with similar functions but active in alternative pathways could be incorporated in a support vector machine classifier based on spectral graph theory. Using 10 microarray data sets, we first reduced the number of data sources relevant for multiple cancer types and outcomes. Three sources on metabolic pathway information (KEGG), protein-protein interactions (OPHID) and miRNA-gene targeting (microRNA.org) outperformed the other sources with regard to the considered class of models. Both fixed and adaptive approaches were subsequently considered to combine the three corresponding classifiers. Averaging the predictions of these classifiers performed best and was significantly better than the model based on microarray data only. These results were confirmed on 6 validation microarray sets, with a significantly improved performance in 4 of them. Integrating interactome data thus improves classification of cancer outcome for the investigated microarray technologies and cancer types. Moreover, this strategy can be incorporated in any kernel method or non-linear version of a non-kernel method.
Introduction
Microarray technology provides a high-resolution snapshot of gene expression within tumors, yet it faces a significant reliability hurdle. Independent studies focusing on the same cancer type often yield disparate "gene signatures." This inconsistency stems from the inherent redundancy of biological systems: while different patients may harbor mutations in different genes, those genes frequently participate in the same functional pathways.
Standard machine learning classifiers typically treat genes as independent features. Consequently, if Patient A has a mutation in Gene 1 and Patient B has a mutation in Gene 2, a traditional model perceives them as distinct cases. However, if these genes are neighbors within a metabolic pathway, the biological and clinical outcomes may be identical. To address this, researchers have developed a method to "denoise" cancer prognosis models by encoding the human interactome—the complex network of biological interactions—into machine learning kernels.
Methodology
The study utilizes Weighted Least Squares Support Vector Machines (LS-SVM) with a novel approach to calculating similarity based on spectral graph theory.
Graph Encoding and Kernels
The researchers represented 15 distinct biological databases (including KEGG for pathways and OPHID for protein interactions) as graphs, where nodes represent genes and edges represent known interactions. For each graph, a Laplacian matrix was calculated, and its pseudoinverse (the "G-matrix") was used to define an "Interactome Kernel."
Instead of a standard linear kernel, the model employs a quadratic form: $x^T G x$. This mathematical structure forces the model to account for "second-order interactions." If two genes are proximal within the biological network, the G-matrix assigns them a high similarity weight, effectively "smoothing" the expression data across the network and capturing functional similarities that independent gene analysis would miss.
Data Sources and Ensembles
The study evaluated 16 public microarray datasets covering breast cancer, ovarian cancer, prostate cancer, and lymphoma. Ten datasets were used for training and source selection, while six were reserved for independent validation.
After testing 15 different interactome data sources, three were identified as the most predictive:
- KEGG: Metabolic pathways
- OPHID: Protein-protein interactions
- microRNA.org: miRNA-gene targeting
The final decision support model combines these three sources using a simple averaging ensemble (the "mean rule").
Key Results and Insights
The integrated model demonstrated a significant performance boost over the standard "microarray-only" baseline. In the validation phase, the interactome-enhanced model showed significantly improved performance in 4 out of 6 independent datasets. Across all 16 datasets, the improvement was highly statistically significant (p < 0.001).
The "Big Three" of Biology
A critical finding was that not all biological data is equally beneficial. While metabolic pathways, protein-protein interactions, and miRNA networks provided a clear signal, other sources, such as protein domain data, introduced more noise than value. Furthermore, the study found that simple averaging of predictions outperformed complex weighting schemes, likely because the small sample sizes typical of clinical data make complex ensembles prone to overfitting.
Robustness and Technology Independence
The method proved robust across various cancer types and microarray platforms (e.g., U133a, U95av2). This suggests that biological networks serve as a universal "prior" that enhances predictive power regardless of the specific technology used to measure gene expression.
Limitations and Future Directions
Despite its success, the approach has limitations. Calculating the pseudoinverse of the Laplacian is computationally intensive, requiring the researchers to limit their analysis to the 5,000 most variable genes. Additionally, the model's accuracy is tied to the completeness and quality of current interactome databases. Future work may explore "dynamic" networks that adapt to specific patient mutations or tissue types, moving beyond the "static" networks used in this study.
Summary updated