![Skip factor à 0, 1, 2, 3
• The skipFactor value to
skip the outliers. Higher
values imply that less
gene are considered
outliers. skipFactor
equal to 0 does not
skip;
1
Zeta à 0.35, 0.4, 0.45,
0.5
• The zeta parameter that
sets the threshold value
which controls the
activation of a linguistic
label;
2
piVal à 0.4 to 0.8 step by
0.05
• The piVal parameter is
equal to the percentage of
values of a class to
determine the fuzzy
patterns. It can take values
in the interval [0,1];
3
Overlapping à 1, 2
• Determines the
number of discrete
labels;
4
Discriminant Fuzzy Pattern – Grid search](https://image.slidesharecdn.com/lung-presentation-ridotta-170926153240/85/A-pathway-and-SVM-based-tool-for-tumor-classification-9-320.jpg)
A pathway and SVM based tool for tumor classification
- 3. The Data • Lung Squamous Cell Carcinoma (LSCC) • 106 patients • 11837 genes • 23074 methylation values • 352 miRNAs • Survival information
- 4. Pipeline
- 5. Similarity Network Fusion - SNF • SNF is a intermediate multi-view clustering methodology for patients sub-typing. Patients similarity network Fusion iterations Fused patients similarity network miRNAs methy
- 6. SNF - Grid search • The algorithm is run different time with the following parameters: • Number of iterations: 200 • K: 10 to 30 step by 1 • Number of nearest neighbors • 𝛼 : 0.3 to 0.8 step by 0.1 • Variance for local model • For each combination of K and 𝛼, the number of clusters was evaluated through two heuristics: eigen-gaps K12 and eigen-gaps K2. • Each clustering was evaluated through the survival analysis by using the log-rank test.
- 7. SNF - The results P-Value = 0.0015 K = 23 𝛼 = 0.6 Number of Iteration = 200
- 9. Skip factor à 0, 1, 2, 3 • The skipFactor value to skip the outliers. Higher values imply that less gene are considered outliers. skipFactor equal to 0 does not skip; 1 Zeta à 0.35, 0.4, 0.45, 0.5 • The zeta parameter that sets the threshold value which controls the activation of a linguistic label; 2 piVal à 0.4 to 0.8 step by 0.05 • The piVal parameter is equal to the percentage of values of a class to determine the fuzzy patterns. It can take values in the interval [0,1]; 3 Overlapping à 1, 2 • Determines the number of discrete labels; 4 Discriminant Fuzzy Pattern – Grid search
- 10. Enrichment Analysis • For each group of genes selected by DFP parameters the enrichment analysis was performed • The p-value is calculated based on the hypergeometric model • We only used KEGG and Reactome Database
- 12. The pathways • The selected pathways are 67: 28 KEGG and 39 Reactome • The selected genes are 1384. The DFP parameters are: • skip factor 2 • zeta 0.3 • piVal 0.6 • overlapping 1
- 13. Classification with SVM • For each pathway a Linear SVM was executed on each pair of classes • Two level cross-validation • 3 outer folds • 2 inner folds • C: 1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6
- 16. name pathway id pathway p-value Accuracy Size classes Cytokine-cytokine receptor interaction K _hsa04060 0.04 0.93 21 5vs2 Cell cycle K _hsa04110 0.03 0.89 12 1vs2 Cell cycle K _hsa04110 0.02 0.97 13 2vs3 Cell cycle K _hsa04110 0.05 0.90 11 5vs3 Osteoclast differentiation K _hsa04380 0.03 1.00 10 1vs4 Antigen processing and presentation K _hsa04612 0.03 1.00 8 1vs4 Antigen processing and presentation K _hsa04612 0.05 0.92 7 2vs3 Antigen processing and presentation K _hsa04612 0.03 1.00 7 5vs4 T cell receptor signaling pathway K _hsa04640 0.04 1.00 11 1vs4 T cell receptor signaling pathway K _hsa04640 0.01 0.81 12 5vs1 Th1 and Th2 cell differentiation K _hsa04658 0.04 1.00 9 1vs4 Th1 and Th2 cell differentiation K _hsa04658 0.04 0.93 9 2vs3 Th17 cell differentiation K _hsa04659 0.04 1.00 11 1vs4 Th17 cell differentiation K _hsa04659 0.04 0.96 12 2vs3 T cell receptor signaling pathway K _hsa04660 0.03 1.00 8 1vs4 T cell receptor signaling pathway K _hsa04660 0.04 1.00 8 5vs4 B cell receptor signaling pathway K _hsa04662 0.04 0.84 7 1vs2 B cell receptor signaling pathway K _hsa04662 0.03 0.89 8 5vs2 Leukocyte transendothelial migration K _hsa04670 0.01 0.91 12 1vs2 Leukocyte transendothelial migration K _hsa04670 0.01 0.90 13 1vs3 Leukocyte transendothelial migration K _hsa04670 0.04 0.92 14 5vs3
- 18. Second step of classification: pathway probabilities combinations • For each pairs of class, we combine the pathways using the class probabilities of SVM as new features. • We try all the combination of pathways using linear SVM • C: 1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6
- 19. Graph Interaction • We create a graph interaction for each combination of classes • The vertices of the graph are the genes in pathways • The size is equal at SVM weight, if a genes is in common between the pathways we pick the max weight. • For the edges: 1. We calculated the correlation between the genes 2. Split the correlation in positive and negative and calculate the MST 3. Only the edges belonged to MST are in the final graph • We highlighted the pathways with shapes of different colours