This project applies unsupervised machine learning techniques to classify molecular subgroups in medulloblastoma (MB) using lncRNA expression profiles from RNA-seq data of 167 MB samples. We utilized the ConsensusClusterPlus R package, a robust tool for consensus clustering, to explore the optimal number of subgroups. ConsensusClusterPlus supports various clustering algorithms, including partitioning around medoids (PAM), hierarchical clustering, and k-means, using distance metrics like Euclidean distance and Pearson correlation.
We also evaluated the classification agreement with the original molecular subgroups from the ICGC cohort using Cohen’s Kappa coefficient. Our findings identified four robust molecular subgroups corresponding to the well-known MB subtypes: Group 3, Group 4, SHH, and WNT.
Objectives
Identify the optimal number of molecular subgroups in medulloblastoma using high-variance lncRNAs.
Validate the agreement between the new lncRNA-based classification and the original molecular subgroups using Cohen’s Kappa.
Visualize the clustering stability across different numbers of subgroups using consensus matrices, CDF, and delta area plots.
Provide insights into the molecular distinction between the identified subgroups based on lncRNA profiles.
Methodology
1. Data Preprocessing
The dataset consists of 167 MB samples, each with RNA-seq normalized expression data. From this dataset, we selected the 5000 lncRNAs with the highest variance to ensure we captured significant expression differences across MB subgroups.
2. Consensus Clustering
Consensus clustering was performed using ConsensusClusterPlus. We evaluated the number of clusters (k) ranging from 2 to 10, with 1000 resamplings to ensure the stability of the clustering solution.
The clustering algorithm used partitioning around medoids (PAM) and Euclidean distance for hierarchical clustering.
Cumulative distribution function (CDF) plots and delta area plots were used to assess clustering stability and to determine the optimal number of subgroups.
3. Validation Using Cohen’s Kappa
To assess the agreement between the lncRNA-based clustering (k=4) and the original MB classification, we calculated Cohen’s Kappa coefficient using the psych package in R.
We computed both unweighted and weighted Kappa values, along with 95% confidence intervals, to evaluate the strength of the agreement.
Results
1. Optimal Number of Subgroups (k = 4)
The consensus clustering analysis identified four subgroups as the optimal number of molecular subgroups for MB.
The CDF plot (Figure 1) shows the stabilization of clustering as the number of clusters increased, with k=4 showing the highest clustering stability. The delta area plot (Figure 2) further confirms that the most significant improvement in clustering stability occurred at k=4.
Figure 1: Cumulative Distribution Function (CDF) for k=2 to k=10 clusters.
Figure 2: Delta Area Plot showing relative changes in area under the CDF curve.
2. Consensus Matrix for k = 4
The consensus matrix for k=4 (Figure 3) reveals clear and well-separated clusters, corresponding to the four known molecular subgroups of MB: Group 3, Group 4, SHH, and WNT.
Each subgroup exhibited high intra-cluster consensus, indicating the robustness of the clustering solution.
Figure 3: Consensus matrix for k=4, showing clear separation between MB subgroups.
3. Cohen’s Kappa: Agreement with Original MB Classification
To validate the clustering solution, we calculated Cohen’s Kappa to measure the agreement between the lncRNA-based classification and the original MB molecular diagnosis.
Weighted Kappa: 0.87 (95% CI: 0.78 - 0.95) suggests excellent agreement, reinforcing the accuracy of the lncRNA-based classification.
Figure 4: Cohen’s Kappa results, comparing lncRNA-based classification (k=4) with original MB subgroups.
Discussion and Insights
The four subgroups identified correspond to the well-established molecular subgroups of medulloblastoma: Group 3, Group 4, SHH, and WNT, supporting the validity of our lncRNA-based approach.
The high Kappa values indicate that the lncRNA expression profiles provide a reliable molecular signature for MB classification, closely mirroring the ICGC classification based on traditional molecular markers.
The robust intra-cluster consensus highlights the ability of lncRNAs to distinguish between MB subgroups, offering potential biomarkers for molecular classification and therapeutic targeting.
Future Directions
Functional validation of key lncRNAs: Investigate the role of top lncRNAs in driving the molecular subtypes of MB.
Integration of multi-omics data: Incorporate genomic, epigenomic, and proteomic data to enhance the understanding of MB biology.
Therapeutic insights: Explore the potential for targeting lncRNA-mediated pathways in MB treatment.
Dependencies
R (v4.0 or higher)
ConsensusClusterPlus R package (v1.48.0)
psych R package for Kappa analysis
RNA-seq normalized expression data from 167 medulloblastoma patients (ICGC cohort)
References
ConsensusClusterPlus: Wilkerson, M.D. & Hayes, D.N. (2010). ConsensusClusterPlus: A class discovery tool with confidence assessments and item tracking. Bioinformatics, 26(12), 1572-1573.
Cohen’s Kappa: Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20, 37-46.
Unsupervised machine learning method establishes medulloblastoma molecular subgroup classification using lncRNA expression
Overview
This project applies unsupervised machine learning techniques to classify molecular subgroups in medulloblastoma (MB) using lncRNA expression profiles from RNA-seq data of 167 MB samples. We utilized the ConsensusClusterPlus R package, a robust tool for consensus clustering, to explore the optimal number of subgroups. ConsensusClusterPlus supports various clustering algorithms, including partitioning around medoids (PAM), hierarchical clustering, and k-means, using distance metrics like Euclidean distance and Pearson correlation.
We also evaluated the classification agreement with the original molecular subgroups from the ICGC cohort using Cohen’s Kappa coefficient. Our findings identified four robust molecular subgroups corresponding to the well-known MB subtypes: Group 3, Group 4, SHH, and WNT.
Objectives
Methodology
1. Data Preprocessing
2. Consensus Clustering
3. Validation Using Cohen’s Kappa
Results
1. Optimal Number of Subgroups (k = 4)
The consensus clustering analysis identified four subgroups as the optimal number of molecular subgroups for MB.
The CDF plot (Figure 1) shows the stabilization of clustering as the number of clusters increased, with k=4 showing the highest clustering stability. The delta area plot (Figure 2) further confirms that the most significant improvement in clustering stability occurred at k=4.
Figure 1: Cumulative Distribution Function (CDF) for k=2 to k=10 clusters.
Figure 2: Delta Area Plot showing relative changes in area under the CDF curve.
2. Consensus Matrix for k = 4
The consensus matrix for k=4 (Figure 3) reveals clear and well-separated clusters, corresponding to the four known molecular subgroups of MB: Group 3, Group 4, SHH, and WNT.
Each subgroup exhibited high intra-cluster consensus, indicating the robustness of the clustering solution.
Figure 3: Consensus matrix for k=4, showing clear separation between MB subgroups.
3. Cohen’s Kappa: Agreement with Original MB Classification
To validate the clustering solution, we calculated Cohen’s Kappa to measure the agreement between the lncRNA-based classification and the original MB molecular diagnosis.
Unweighted Kappa: 0.80 (95% CI: 0.72 - 0.87) indicates substantial agreement.
Weighted Kappa: 0.87 (95% CI: 0.78 - 0.95) suggests excellent agreement, reinforcing the accuracy of the lncRNA-based classification.
Figure 4: Cohen’s Kappa results, comparing lncRNA-based classification (k=4) with original MB subgroups.
Discussion and Insights
Future Directions
Dependencies
References