Abstract
Metagenomics promises to provide new valuable insights into the role of microbiomes in eukaryotic hosts such as humans. Due to the decreasing costs for sequencing, public and private repositories for human metagenomic datasets are growing fast. Metagenomic datasets can contain terabytes of raw data, which is a challenge for data processing but also an opportunity for advanced machine learning methods like deep learning that require large datasets. However, in contrast to classical machine learning algorithms, the use of deep learning in metagenomics is still an exception. Regardless of the algorithms used, they are usually not applied to raw data but require several preprocessing steps. Performing this preprocessing and the actual analysis in an automated, reproducible, and scalable way is another challenge. This and other challenges can be addressed by adjusting known big data methods and architectures to the needs of microbiome analysis and DNA sequence processing. A conceptual architecture for the use of machine learning and big data on metagenomic data sets was recently presented and initially validated to analyze the rumen microbiome. The same architecture can be used for clinical purposes as is discussed in this paper.
Highlights
Current studies are showing the importance and contribution of communities of microorganisms, known as the microbiota, for human development [1], diet–microbiota interactions [2], interactions with the immune system [3,4], and diseases [5,6]
As this paper focuses on demonstrating classification improvements for Machine Learning (ML)
Clustering algorithms on the other hand try to find and group similar data points without using predetermined classes. Another unsupervised type of ML often used in metagenomics that can be distinguished from classification and clustering is dimensionality reduction [13]
Summary
Current studies are showing the importance and contribution of communities of microorganisms, known as the microbiota, for human development [1], diet–microbiota interactions [2], interactions with the immune system [3,4], and diseases [5,6]. The traditional way to attempt to answer these and other research questions would be to take samples of the microorganisms from their environment and to culture these in a lab Afterward, they could be studied and compared to other samples to detect similarities or differences in the composition of microorganisms between samples. A single microbiome study can contain hundreds of gigabytes or more of raw sequencing data During processing, this can get multiplied many times as intermediate results in different formats need to be produced. A concern arises in the application of deep learning models to metagenomics classification of phenotypes (linking metagenomic data to observable characteristics of the microorganisms or hosts), where there are more features than samples, which is often the case in predictive modeling of metagenomes.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
