[Paper Review] Deep Learning for Metagenomic Data: using 2D Embeddings and Convolutional Neural Networks

Thanh Hai Nguyen, Yann Chevaleyre+3 more|arXiv (Cornell University)|Dec 1, 2017

Gene expression and cancer classification27 references17 citations

TL;DR

This paper proposes MET2IMG, a deep learning framework that transforms metagenomic data into 2D image representations using phylogenetic ordering or t-SNE embedding, enabling effective application of convolutional neural networks (CNNs) for disease prediction. The method achieves state-of-the-art performance on six metagenomic datasets, with 2D CNNs on t-SNE-based images outperforming traditional ML methods and raw data, especially in complex conditions like IBD and T2D.

ABSTRACT

Deep learning (DL) techniques have had unprecedented success when applied to images, waveforms, and texts to cite a few. In general, when the sample size (N) is much greater than the number of features (d), DL outperforms previous machine learning (ML) techniques, often through the use of convolution neural networks (CNNs). However, in many bioinformatics ML tasks, we encounter the opposite situation where d is greater than N. In these situations, applying DL techniques (such as feed-forward networks) would lead to severe overfitting. Thus, sparse ML techniques (such as LASSO e.g.) usually yield the best results on these tasks. In this paper, we show how to apply CNNs on data which do not have originally an image structure (in particular on metagenomic data). Our first contribution is to show how to map metagenomic data in a meaningful way to 1D or 2D images. Based on this representation, we then apply a CNN, with the aim of predicting various diseases. The proposed approach is applied on six different datasets including in total over 1000 samples from various diseases. This approach could be a promising one for prediction tasks in the bioinformatics field.

Motivation & Objective

• To address the challenge of applying deep learning to metagenomic data where feature count (d) far exceeds sample size (N), leading to overfitting in standard deep learning models.
• To develop meaningful 2D image representations of metagenomic data—using phylogenetic ordering and t-SNE embedding—that preserve biological structure and enable effective CNN learning.
• To demonstrate that 2D CNNs applied to these synthetic images can achieve superior classification performance compared to traditional machine learning methods like Random Forest and SVM on metagenomic datasets.
• To evaluate the impact of different data representations (abundance vs. presence) and network architectures (1D vs. 2D convolutions) on prediction accuracy.

Proposed method

• Transforming metagenomic abundance data into 2D images via 'fill-up' visualization: species are ordered by phylogenetic tree and filled row-wise from left to right, top to bottom, with color-coded abundance bins on a logarithmic scale.
• Applying t-SNE dimensionality reduction to project high-dimensional metagenomic samples into 2D space, preserving local similarities, and visualizing them as images where color represents species abundance or presence.
• Training 2D and 1D convolutional neural networks (CNNs) on these image representations, with architectures including 5 convolutional layers (2D) or 2 layers (1D), each with 20 filters and ReLU activation.
• Using wide convolutions to preserve spatial dimensions and max pooling (2x2, stride 2) to downsample feature maps, followed by fully connected layers and a LogSoftMax output layer for binary classification.
• Employing two classification strategies: two-node output with LogSoftMax and cross-entropy loss, and one-node output with sigmoid and binary cross-entropy; the two-node approach yielded better results and was used in final models.
• Optimizing models with stochastic gradient descent (SGD) and Adam optimizer, mini-batch size 16, learning rate 0.0005, momentum 0.1, and weight decay of 1e-5.

Experimental results

Research questions

• RQ1 Can 2D image representations of metagenomic data enable effective training of convolutional neural networks despite the high-dimensionality and low-sample-size nature of such data?
• RQ2 How do different data representations—phylogenetic ordering (PLG) versus t-SNE embedding—impact the performance of CNN-based disease prediction models?
• RQ3 Does using 2D convolutions on image-converted metagenomic data outperform 1D convolutions on raw data or fully connected networks in classification accuracy?
• RQ4 To what extent do image-based representations improve prediction performance compared to state-of-the-art machine learning methods like Random Forest and SVM on metagenomic datasets?
• RQ5 How do feature representations based on abundance (ABD) versus presence (PRE) influence model performance across diverse diseases?

Key findings

• The MET2IMG framework achieved state-of-the-art performance on six metagenomic datasets, with 2D CNNs on t-SNE-based images outperforming traditional ML methods such as Random Forest and SVM.
• For the IBD dataset, the 2D CNN on t-SNE images achieved 81.8% accuracy (Con2D, PRE), surpassing the best MetAML result of 80.9% and demonstrating the highest improvement over baseline.
• On the T2D dataset, the 2D CNN on t-SNE images with PRE representation reached 71.1% accuracy (Con2D, t-SNE), exceeding the best MetAML result of 70.3% and showing strong performance on complex metabolic disorders.
• 2D convolutions consistently outperformed 1D convolutions across all datasets, indicating that spatial structure in 2D image representations enhances feature learning.
• The t-SNE-based image representation required higher image dimensions (e.g., 64x64) to avoid overlapping data points and improve performance, though it demanded more computational resources than the PLG method.
• The two-node LogSoftMax output layer with 2D CNNs yielded better results than the one-node sigmoid approach, confirming its superiority for binary classification in this context.