Deep learning algorithms have obtained numerous achievements in image classification, speed recognition, video processing. Visualizing metagenomic data is a challenge because of its complexity and high-dimensional. In this paper, we introduce several approaches based on dimensionality reduction algorithms and data density to visualize features which reflect the species abundance. The sophisticated methods used in this study, that are unsupervised approaches, carry out dimensionality reduction and map the data into a 2-dimensional space. From the visualizations obtained, deep learning techniques are leveraged to enhance the prediction performance for colorectal cancer. We show by experiments on five Metagenome-based colorectal cancer datasets from different regions such as Chinese, Austrian, American, German and French cohorts that the proposed visualizations allow to visualize bio-medical signatures and improve the prediction performance compared to classical machine learning.