IGTD

Image Generator for Tabular Data

Model Architecture

Image Generator for Tabular Data (IGTD) transforms tabular data into images. Every sample vector is converted into a heatmap image, in which each pixel represents one feature in the original tabular data. The algorithm searches for an optimized assignment of features to pixels, so that similar features are assigned to neighboring pixels and dissimilar features are assigned to pixels that are far apart. For subsequent prediction modeling, Convolutional Neural Networks (CNNs) are trained based on the images generated by IGTD. In the application of drug response prediction, gene expression profiles of cancer cases and molecular descriptors of drugs are converted into their respective images. Two separate subnetworks of convolutional layers are used to generate embeddings of gene expression and drug descriptor images. Then the embeddings of gene expressions and drug descriptors are concatenated and forwarded to multiple dense layers to make predictions on treatment response.

Feature Representation

• Cancer features:

  • Gene Expression: as numeric features

• Drug features:

  • Drug Molecular Descriptors: as numeric features

URLs

• Original GitHub

• IMPROVE GitHub

References

1. Y. Zhu et. al. “Converting tabular data into images for deep learning with convolutional neural networks”, Scientific Reports, 2021