Drug sensitivity modeling

This repo contains scripts which I wrote to build predictive models on cancer cell line drug sensitivities. This was part of a homework I had. The tasks with the required data were provided a priori.

Aim

The tasks are described in the input/comp-bio-homework-description.pdf file.

1. Given the data perform a dimensionality reduction with a preferred method and viusalize it.
2. Using gene expression data identify biomarkers for Lapatinib sensitivity.
3. Build a predictive model for Lapatinib sensitivity and evaluate it's performance.

Note

The created predictive model is by no means exhaustive and was not optimized.

Summary

Task 1

• I decided to use principal component analysis (PCA).
• I retrieved only the first 10 principal components as from PC5 the explained variability does not change significantly (see: output/pca.png)
• The samples are colored according to the cancer type, some degree of separation of the samples is noticeable on the PC1.

Task 2

• As the gene expression provided was TPM transformed and only cancer cell line samples were provided, there was no possibility of differential gene expression analysis. Instead, I used a feature selection approach.
• I used two regression approaches: random forest and lasso regression. The output of these approaches was overlapped and only commonly selected features were kept.
• A total of 60 genes were selected as influential for Lapatinib sensitivity of cancer cells (see: output/selected_genes.tsv and output/selected_gene_ids.tsv).
• A systematic approach was also performed to get an idea of what functions are affected by the drug treatment. The analysis resulted in only a single enriched function, i.e. regulation of blood pressure (see: output/singificant_terms.xlsx). This could make sense as the RTK is involved in metabolism, growth and differentiation in cancers and blood supply of tumors could be essential.

Task 3

• To evaluate the model I used a simple train/test split approach. I retained 20% of the samples as testing data.
• Firstly I used a random forest regression approach. I also implemented a short optimization step for the model fitting to try to improve the evaluation metrics. The model performance was lackluster with a R2 = 0.36 and MSE = 1.64 (see: output/rf_pred-vs-actual.png).
• Secondly, I used the lasso regression in order to get better predictions. I used similar approaches as above, without the optimization step. The overall model predictions showed improvement (see: output/lasso_pred-vs-actual.png) with an R2 = 0.56 and a MSE = 1.13.
• Lastly I tried to use a classifier. As a classifier I chose the random forest approach. Similar steps were used as above to optimize the model fit and evaluate its performance. The classified performed somewhat poorly (see: output/roc_curve.png) with a ROC AUC = 0.77. Accuracy of model is 0.68 and precision is 0.68 with a recall of 0.59.