Network-based drug sensitivity prediction.

Area of Science:

• Genomics
• Computational Biology
• Pharmacogenomics

Background:

• Drug sensitivity prediction is crucial for drug discovery, often using high-throughput genomic data.
• Existing computational methods, including deep neural networks, largely ignore modular relations among genomic features.
• Gene co-expression networks offer a way to capture these relationships for improved prediction.

Purpose of the Study:

• To investigate the role of gene co-expression networks in drug sensitivity prediction.
• To develop novel network-based methods for identifying predictive genomic features.
• To compare the performance of these new methods against existing algorithms and deep learning models.

Main Methods:

• Introduced a network-based feature selection method utilizing gene co-expression networks.
• Proposed two graph-based neural network models integrating gene network information.
• Conducted a large-scale comparative study with canonical algorithms (Elastic Net, Random Forest, PLS, SVR) and deep neural networks.
• Utilized a non-small cell lung cancer (NSCLC) cell line RNA-seq dataset with 50 drug treatments.

Main Results:

• Network-based feature selection improved prediction performance over Pearson correlation coefficients.
• Random Forest demonstrated superior performance among canonical algorithms and deep neural networks.
• The proposed graph-based neural network models outperformed standard deep neural network models.
• Prediction performance was found to be drug-dependent, potentially linked to drug mechanisms of action.

Conclusions:

• Network-based feature selection and prediction models enhance drug response prediction accuracy.
• Genomic feature relationships within networks are more robust than individual feature correlations in high-dimensional, low-sample datasets.