Predicted meta-omics: A potential solution to multi-omics data scarcity in microbiome studies.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Gut microbiome imbalances are linked to diseases like IBD, diabetes, and cancer.
• Metagenomics and amplicon sequencing offer limited insights into microbial functions.
• Other meta-omics data provide deeper insights but are expensive and labor-intensive.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting meta-omics data from metagenomics data.
• To infer connections between different layers of microbial functions using computational approaches.
• To assess the utility of predicted meta-omics features in downstream biological applications.

Main Methods:

• Evaluation of various machine learning models, including elastic net regression, random forests, and neural networks.
• Utilizing paired meta-omics datasets for model training and validation.
• Assessing prediction accuracy for transcript, metabolite, and protein abundances.

Main Results:

• Elastic net regression and random forests accurately predicted transcript (correlation up to 0.77) and metabolite (correlation up to 0.74) abundances.
• Predicting protein profiles proved more challenging than transcript or metabolite prediction.
• Multi-output regression neural networks showed comparable performance when trained on fewer output features.
• Predicted features enabled inflammatory bowel disease classification with performance similar to experimental data.

Conclusions:

• Machine learning can effectively predict microbial transcript and metabolite abundances from metagenomics data.
• Predicted meta-omics features are valuable for downstream applications like disease classification.
• This computational approach offers a cost-effective alternative to experimental meta-omics data generation.