Predicting drug response through learning from network-based integration of multi-omics data in cancer cell lines

Predicting the treatment response continues to be a significant hurdle due to the inherent heterogeneity among tumours, leading to divergent responses to identical therapeutic approaches. This thesis introduces a strategy aiming to overcome this challenge by integrating a network-based hybrid machine learning model to detect drug response patterns, potentially elucidating the intricate biological mechanisms at play. Toward this goal, our approach deploys baseline omics data from more than 900 cancer cell lines. We assemble multi-omics data, including mutation, proteomic, and transcriptomic profiles from a broad spectrum of cell lines, along with their drug sensitivity data and protein targets of the drugs. The collected information is integrated onto a human protein-protein interactome, and using a personalized PageRank network propagation technique, we generate around 28,000 unique context-specific subnetworks, each representing a distinct drug-cell line pair. From this pool, we filtered out the grey zone contexts and focused on 1,194 pairs involving 426 cell lines and 45 drugs, labelled as either sensitive or resistant, facilitating more accurate predictions. Subsequently, these complex networks are converted into a format compatible with machine learning models via Graph2Vec, an unsupervised graph embedding algorithm, resulting in vectorized network representations encapsulating their topological and attribute-related features into a vector. These transformed networks serve as the input for a machine learning classifier, enabling it to predict drug responses based on the distinct properties inherent to each drug-cell line combination. Validity of this prediction model is thoroughly examined through a stringent testing, including a 10-fold cross-validation, resulting in an accuracy of 0.83 and an Area Under the Precision-Recall Curve (AUC) of 0.90. Moreover, additional 10-fold blind testing on independent cell line, drug, and tissue sets yielded an AUC of 0.72, 0.89, and 0.81, respectively. These findings underscore the model's potential in enhancing the precision of drug response predictions and illuminating the causal network modules underpinning drug resistance.