Phase 6: Results

In this phase, you will gather and report findings for your top predictive model.

Report the best model and its test set performance (e.g., AUC). List the top predictors identified via Variable Importance. Describe insights from confounder analysis (Phase 3) and Model Interpretation (if applicable). Discuss the biological relevance of the top predictors.

1. Combine Findings

1. Identify the top model from phase 4 by considering

   • Model performance metrics

   • ROC Curves

   • Biological relevance of top predictors in Variable Importance

   • Confounder Check (phase 3)

2. Pull together all your findings, including

   • Clustered t-SNE plots for responder classification, if applicable (Phase 2)

   

   • t-SNE plots and analysis from Confounder check (Phase 3)

   

   • Model performance metrics (Phase 4)

   

   • Training and Testing ROC Curves

   

   • Model Interpretation plots, if applicable

   

   • Variable Importance bar plot

   

   • Features across dataset dot plots for top predictive features

   

2. Analyze GO Terms & Biological Themes

The GO terms present in your dataset are a result of pathway enrichment analysis, which is a powerful tool external to PANDORA that helps identify biological themes from gene expression. You can use GO term databases to identify GO terms to uncover overall biological themes in responder groups and model prediction.

• Pathway Enrichment Analysis Tools:

  • clusterProfiler in R

  • DAVID

  • Metascape

  • Enrichr

• GO term databases

  • GO

  • KEGG

  • Reactome

GO Terms alongside predictive variables can be used to identify biological themes using the following workflow:

1. Identify GO terms from your top predictors

   1. Open the Gene Ontology Resource webpage

   2. Search for all your top GO predictive terms in the form GO:#

      1. i.e. GO:0070206, GO:1903214

   3. Click term history to see ancestor chart, child terms, and co-occurring terms

   4. Create a list of all biological processes and themes related to your GO Terms

2. Check the expression levels of baseline terms in responder groups

   1. Select your predictive processed dataset from the Workspace (This dataset should only contain baseline features and your responder columns)

   2. Navigate to Discovery -> Start -> Hierarchical Clustering

   3. Configure Clustering Column Selection

      1. Select your Responder column for the Columns

      2. Set First (n) rows such that it is larger than the total number of baseline features

      
   4. Configure Clustering Display Options

      1. Enable Grouped display

      2. Select the responder column for Grouped column

      
   5. Click Plot image

3. Analyze the resultant heatmap

   1. Take note on how the expression of top predictive variables varies among the responder classes.

   2. With biological themes in mind from predictive variables and top GO terms, consider the biological themes among responder classes.

   
4. Make plots reflecting biological themes (optional)

   1. Outside of PANDORA, you may create additional plots, such as radar plots, reflecting the different immune profiles of responder classes based on the baseline or fold change expression levels of features in each class.

You've now identified and analyzed your strongest model through consideration of model performance, biological interpretation, and confounder analysis. By pulling all your analysis together, you have now created a comprehensive picture of your model to draw biologically relevant insights from.