Predictive Modeling of Anticancer Drug Sensitivity Using REFINED CNN.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Convolutional neural networks (CNNs) excel at spatial data but struggle with tabular datasets lacking inherent structure.
• Predicting anticancer drug sensitivity often uses tabular data, limiting the application of CNNs.
• Existing methods may not fully exploit feature relationships in non-sequential, non-image data.

Purpose of the Study:

• To present a novel computational procedure, REpresentation of Features as Images with NEighborhood Dependencies (REFINED).
• To enable the effective application of CNNs to tabular datasets for tasks like anticancer drug sensitivity prediction.
• To improve predictive performance by creating image-like representations of high-dimensional feature vectors.

Main Methods:

• Developed REFINED to map high-dimensional feature vectors into compact 2D images.
• Integrated REFINED with CNN models for deep learning applications.
• Compared CNNs with REFINED against fully connected networks on high-dimensional feature vectors.

Main Results:

• REFINED successfully transforms tabular data into image formats suitable for CNNs.
• The REFINED-CNN approach demonstrated enhanced predictive performance.
• Reduced model parameterization and improved embedded feature extraction were observed compared to traditional methods.

Conclusions:

• REFINED provides a viable strategy to adapt CNNs for tabular data analysis.
• This method offers improved accuracy in anticancer drug sensitivity prediction.
• REFINED facilitates better utilization of complex biological datasets in deep learning models.