Geodesy

Gene

Marketing

Principle of compositionality

Benchmark (surveying)

Ecology

Biology

Identification (biology)

Benchmarking

Data mining

Computer science

Filter (signal processing)

Computational biology

Business

Genetics

Microbiome

Bioinformatics

Geography

Sampling (signal processing)

Inference

Data science

Artificial intelligence

Metagenomics

Machine learning

Computer vision

Nature Communications

While metagenomic sequencing has become the tool of preference to study host-associated microbial communities, downstream analyses and clinical interpretation of microbiome data remains challenging due to the sparsity and compositionality of sequence matrices. Here, we evaluate both computational and experimental approaches proposed to mitigate the impact of these outstanding issues. Generating fecal metagenomes drawn from simulated microbial communities, we benchmark the performance of thirteen commonly used analytical approaches in terms of diversity estimation, identification of taxon-taxon associations, and assessment of taxon-metadata correlations under the challenge of varying microbial ecosystem loads. We find quantitative approaches including experimental procedures to incorporate microbial load variation in downstream analyses to perform significantly better than computational strategies designed to mitigate data compositionality and sparsity, not only improving the identification of true positive associations, but also reducing false positive detection. When analyzing simulated scenarios of low microbial load dysbiosis as observed in inflammatory pathologies, quantitative methods correcting for sampling depth show higher precision compared to uncorrected scaling. Overall, our findings advocate for a wider adoption of experimental quantitative approaches in microbiome research, yet also suggest preferred transformations for specific cases where determination of microbial load of samples is not feasible.

Benchmarking microbiome transformations favors experimental quantitative approaches to address compositionality and sampling depth biases