To identify samples with highly divergent (dysbiotic) metagenomic microbial compositions, as a complement to baseline disease diagnosis, we defined a dysbiosis score based on Bray–Curtis dissimilarities to non-IBD metagenomes. First, a ‘reference set’ of samples was constructed from non-IBD subjects by taking all samples after the 20th week after the subject’s first stool sample. This was chosen because a subset of the non-IBD subjects at the start of their respective time series may not yet have overcome any gastrointestinal symptoms that triggered the initial visit to a doctor, though these were ultimately not caused by IBD. The dysbiosis score of a given sample was then defined as the median Bray–Curtis dissimilarity to this reference sample set, excluding samples that came from the same subject (Fig. 2c ).
To identify samples that were highly divergent from the reference set, we thresholded the dysbiosis score at the 90th percentile of this score for non-IBD samples. This therefore identifies samples with a feature configuration that has a less than 10% probability of occurring in a participant without IBD. By this measure, 272 metagenomes were classified as dysbiotic. Samples from participants with CD or UC were overrepresented in the dysbiotic set, with 24.3% and 11.6% of their samples classified as dysbiotic, respectively. As expected, these samples also tended to locate in the extremes of the taxonomic ordination based on metagenomes (Extended Data Fig.3b, c ). Dysbiosis was unevenly distributed among subjects (Extended Data Fig. 3d ), with some subjects remaining dysbiotic for all or most of their time series, while others remained non-dysbiotic for their entire time series.
To lend additional support to the definition of dysbiosis (that is, as outliers by one type of microbiome profile), we tested the concordance between dysbiosis classifications made using the same statistical definition, but applied to metabolomic rather than taxonomic profiles. That is, we defined a metabolomic dysbiosis score as the median Bray–Curtis dissimilarity of one metabolomic profile to the non-IBD metabolomic profiles (after the 20th week), and defined the dysbiosis threshold as the 90th percentile of this distribution among non-IBD metabolomic profiles. We then compared these dysbiosis classifications with those from the nearest metagenomic sample (up to two weeks, see ‘Cross-measurement type temporal matching’) and found that dysbiotic samples identified by metagenomics were 4.6 times more likely to be dysbiotic by metabolomics (Fisher’s exact P = 5.9 × 10−9), showing that dysbiosis measurements are highly consistent across measurement types.
To test the sensitivity of the dysbiosis classification to the choice of reference data set, we also performed the dysbiosis classification using the HMP1-II stool samples10 (link) as the reference sample set instead of the non-IBD samples. The resulting dysbiosis scores (Extended Data Fig.3e ) were highly concordant (Spearman ρ = 0.86; P < 2.2 × 10−16), as were the dysbiosis classifications (odds ratio of 56; Fisher’s exact P < 2.2 × 10−16). This shows that, despite the inclusion of subjects with other conditions in the non-IBD group here, as well as large differences in measurement technologies between the data sets, the dysbiosis classification is highly robust. Furthermore, 43 out of 426 (10.1%) of non-IBD samples were classified as dysbiotic using the HMP1-II samples as reference, falling remarkably close to the 10% expected by the definition and showing that the enrichment of IBD samples in the dysbiotic set is not simply a consequence of the definition.
To identify samples that were highly divergent from the reference set, we thresholded the dysbiosis score at the 90th percentile of this score for non-IBD samples. This therefore identifies samples with a feature configuration that has a less than 10% probability of occurring in a participant without IBD. By this measure, 272 metagenomes were classified as dysbiotic. Samples from participants with CD or UC were overrepresented in the dysbiotic set, with 24.3% and 11.6% of their samples classified as dysbiotic, respectively. As expected, these samples also tended to locate in the extremes of the taxonomic ordination based on metagenomes (Extended Data Fig.
To lend additional support to the definition of dysbiosis (that is, as outliers by one type of microbiome profile), we tested the concordance between dysbiosis classifications made using the same statistical definition, but applied to metabolomic rather than taxonomic profiles. That is, we defined a metabolomic dysbiosis score as the median Bray–Curtis dissimilarity of one metabolomic profile to the non-IBD metabolomic profiles (after the 20th week), and defined the dysbiosis threshold as the 90th percentile of this distribution among non-IBD metabolomic profiles. We then compared these dysbiosis classifications with those from the nearest metagenomic sample (up to two weeks, see ‘Cross-measurement type temporal matching’) and found that dysbiotic samples identified by metagenomics were 4.6 times more likely to be dysbiotic by metabolomics (Fisher’s exact P = 5.9 × 10−9), showing that dysbiosis measurements are highly consistent across measurement types.
To test the sensitivity of the dysbiosis classification to the choice of reference data set, we also performed the dysbiosis classification using the HMP1-II stool samples10 (link) as the reference sample set instead of the non-IBD samples. The resulting dysbiosis scores (Extended Data Fig.
Free full text: Click here
