Dimensionality Reduction based Transfer Learning applied to Pharmacogenomics Databases.

Area of Science:

• Pharmacogenomics
• Computational Biology
• Bioinformatics

Background:

• Numerous pharmacogenomics databases exist for personalized therapy predictive modeling.
• Limited consistency between databases hinders cross-database model application despite common cell lines and drugs.

Purpose of the Study:

• To investigate if models trained on one pharmacogenomics database can improve predictions on another.
• To explore methods for enhancing cross-database predictive modeling in pharmacogenomics.

Main Methods:

• Utilized two pharmacogenomics databases for comparative analysis.
• Employed common basis vectors (including PCA-based) to represent database information.
• Assessed prediction performance against naive cross-database model application.

Main Results:

• Representing pharmacogenomics databases with common basis vectors significantly improved prediction performance.
• PCA-based basis vectors demonstrated robustness, even with low correlated input features.
• This method outperformed the naive application of models trained on one database to another.

Conclusions:

• Common basis vector representation is a viable strategy to improve cross-database prediction in pharmacogenomics.
• This approach enhances the utility of existing pharmacogenomics data for personalized medicine.
• PCA-based methods offer a robust solution for integrating diverse pharmacogenomics datasets.