Although one would expect to use genomic data to discriminate diseases/phenotypes, this is a complex and hard task as key information to achieve it is spread over high-dimensional spaces. To address this issue, we envisaged that clustering similar variables, e.g. genes, would facilitate the identification of such information. Thus, here we have developed DeepInsight, a method comprising three steps: rearrange variables, extract features, and construct a classification model. Non-image data are transformed into image data allowing us to apply and take advantage of the learning capabilities of Convolutional Neural Networks (CNNs). Our experiments showed that with real data, such as cancer transcriptome, DeepInsight achieved much higher accuracy than typical machine learning methods. Our methodology has the potential to contribute to various biomedical applications and improve feature and structure extraction in non-image data.