Converting tabular data into images for deep learning with convolutional neural networks.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Convolutional neural networks (CNNs) excel with sequential or spatial data like speech and images.
• Tabular data often lacks inherent feature order, limiting CNN applicability.
• Predicting anti-cancer drug response from molecular data is crucial but challenging.

Purpose of the Study:

• To develop a novel algorithm, image generator for tabular data (IGTD), for transforming tabular data into image representations suitable for CNNs.
• To evaluate IGTD's effectiveness in preserving feature relationships and improving predictive performance.
• To enhance the prediction of anti-cancer drug response using CNNs trained on IGTD-generated images.

Main Methods:

• Developed the image generator for tabular data (IGTD) algorithm to map tabular features to image pixels, optimizing for feature proximity.
• Applied IGTD to transform cancer cell line (CCL) gene expression profiles and drug molecular descriptors into image formats.
• Trained CNNs on these IGTD-generated image representations for anti-cancer drug response prediction.

Main Results:

• IGTD creates compact image representations that better preserve the neighborhood structure of original features compared to existing methods.
• CNNs trained on IGTD images of CCLs and drugs outperformed those trained on alternative image representations.
• IGTD-based CNNs showed superior performance in predicting anti-cancer drug response over models using original tabular data.

Conclusions:

• IGTD offers an effective method for converting tabular data into image representations for CNNs.
• The IGTD approach significantly improves the prediction accuracy of anti-cancer drug response.
• This method holds promise for applying deep learning models to diverse tabular datasets in computational biology and drug discovery.