Probe identification, selection, in silico and in vitro testing

To establish and optimise the most applicable bacterial probeset, data from previous IBD and IBS intestinal microbiota research was compiled based on pre‐defined search criteria (Data S3) to provide >500 bacterial observations associated with the occurrence of IBD and IBS. From a combined dataset of 496 16S rRNA gene sequences (consensus sequence[s] for each species, chosen from all available long 16S rRNA sequences and purified to avoid sequences errors) from 269 bacterial species, probes were designed to cover major bacterial observations made from the literature. All probes were designed according to Vebø et al.22 with a minimum melting temperature (T _m) of 60 °C by the nearest‐neighbour method23 for the target group where the nucleotide 3′ end of the probe is a cytosine; nontarget group probe requirements were a T _m of 30 °C or absence of a cytosine as the nucleotide adjacent to the 3′ end of the probe. Each probe was designed to target a bacterial species or group, i.e. Faecalibacterium prausnitzii (species), Lactobacillus (genus), Clostridia (class) and Proteobacteria (phylum), based on their 16S rRNA sequence (V3–V9). Probes that satisfied target detection and nontarget exclusion in silico were evaluated for cross‐labelling, self‐labelling and cross hybridisation before final validation was performed against bacterial strains in vitro.

After in vitro testing, a panel of 124 optimal probes was further selected using variable selection methods: variable importance in projection, selectivity ratio and interval partial least squares using data from a selection of healthy and IBS samples (data not shown). The variables (probes) were selected based on their ability to distinguish between samples isolated from healthy individuals and IBS patients. A final panel of 54 probes was selected covering the sites across V3 to V7 on the 16S rRNA sequence (Figure 1). Bacterial target specificity, tested with the 54‐probe set against 368 available single bacterial strains (Data S4), was performed to define the target bacteria for each probe. The probes detect bacteria within the six phyla; Firmicutes, Proteobacteria, Bacteroidetes, Actinobacteria, Tenericutes and Verrucomicrobia, covering 10 taxonomic bacterial classes and 36 genera (for more details on the bacterial targets for the 54 probes see Data S5).

Sample preparation and detection

The GA‐test is based on regular molecular biology techniques, comprising human faecal sample homogenisation and mechanical bacterial cell disruption; automated total bacterial gDNA extraction using magnetic beads; 16S rRNA PCR DNA amplification covering V3–V9; probe labelling by single nucleotide extension; hybridisation to complementary probes coupled to magnetic beads; and signal detection using BioCode 1000A 128‐Plex Analyzer (Applied BioCode, Santa Fe Springs, CA, USA). The method is described in detail in Data S2, and an overview to the whole process from sample preparation to result is shown in Figure 2.

Open in a separate window

Figure 2

Flow chart illustrating the GA‐map Dysbiosis Test development, starting with in silico development of bacterial probe set, standardisation of laboratory analysis process, model calibration and verification in healthy individuals (normobiotic reference collection), and validation in healthy, IBS and IBD individuals. Derivation of a DI based on bacterial 16S rRNA DNA analysis in faecal samples demonstrates that a DI score >2 confirms microbiota profile deviations from the normobiotic reference collection.

Data pre‐processing

To ensure high quality assurance, several quality control criteria were applied to the detection data for each sample: (i) a bead count >2 for each probe; (ii) the hybridisation control (HYC) median signal >13 000; (iii) a median background signal <500 and (iv) a universal control median signal >4500. Normalisation was applied by first dividing the signal intensity of each probe in each sample by the signal intensity for HYC for that sample, and multiplying by 1000. This was done to adjust for sample differences due to pipetting or hybridisation. Subsequently, normalisation to adjust for run differences was applied by dividing the HYC‐normalised signal of each probe in each sample by the median HYC‐normalised signal of each probe for replicates of a synthetic DNA control (Data S2; Table S1), and multiplying by 1000. Prior to normobiotic microbiota profile calibration, normalised signal intensities below 15 were set to 0 to remove for low background noise and data was mean centred. Test and validation samples were normalised, and normalised signals below 15 were set to 0 before data were mean centred using mean probe signals from the normobiotic reference cohort.

Dysbiosis test development and validation

Principal component analysis (PCA)24 was used to build a normobiotic microbiota profile (model). The boundary between nondysbiotic and dysbiotic was determined by calculating confidence regions for the values of Hotelling's T‐square and Q statistics given by PCA scores in the model. Geometrically this corresponds to a rectangle with one corner located at the origin which classifies samples located within the rectangle as nondysbiotic and samples located outside as dysbiotic. Analysis of T‐square and Q statistics scaled by the confidence limit showed that the Euclidian distance from the origin had a log‐normal like distribution (data not shown). Euclidian distance from the origin was used to merge the two dimensions, and weighting was performed to capture the effect of T‐squared and Q statistics as appropriate. A single numeric representation of the degree of dysbiosis, defined as the Dysbiosis Index (DI), was derived from a log‐normal distribution by assigning estimated portions of the distribution to different values on a scale set from 0 to 5. A DI value of 2 was defined as class separation represented by the identified confidence limits; a DI of 2 or lower being the nondysbiotic region and a DI of 3 or higher being the dysbiotic region. The higher the DI above 2, the more the sample is considered to deviate from normobiosis, e.g. sample A with DI = 4 is farther away from the normobiotic reference cohort in the Euclidian space than sample B with DI = 3, thus A is more dysbiotic than B. The scale was optimised with emphasis on reducing technical variation between replicates, meaning that the integer part of the numeric output is decided by pre‐determined levels of the Euclidian distance.

To create the GA‐test, 211 healthy individuals were selected and randomly split into a training set (n = 165) designed to build models and a test set (n = 46) designed to tune parameters. Duplicate samples were run, and mean normalised signal was used for training and testing. Sample demographics for the two groups were similar (Table 2). In addition, a set of IBS patients were included in the test set (n = 127). A number of models were developed and evaluated, and the frequency of dysbiosis in the test set was used as measure of model performance. For the final PCA model, 15 principal components were used, and a 98% confidence limit was determined for T‐squared and Q statistics to define class separation. When the model is used to score other samples, values outside these limits are defined as dysbiotic.

External validation using an independent test set comprising healthy, IBS and IBD subjects (n = 287) was used to assess the clinical diagnostic performance of the model (Table 3). The validation set subjects were all from unique donors who had not been included in the healthy reference collection used for normobiotic profile calibration or in parameter tuning. Each sample was processed using the finalised algorithm which converts data for each sample into a single integer, i.e. the DI, which represents the degree of dysbiosis based on bacterial abundance and profile within a sample relative to the established normobiotic profile. A DI > 2 represents a potentially clinically relevant deviation in microbiotic profile from that of the normobiotic reference collection. Finally, the dysbiosis frequency was calculated. In addition, PCA was performed on the validation set to investigate differences in microbiota profile between the three subject groups.

Technical performance

The EU directive for in vitro diagnostic tests was followed to ensure compliance with a CE‐marked test.25 The main technical parameters evaluated were precision and quantitative range of the test; both at probe signal level and at final output level (i.e. DI). At probe level, precision of signals [coefficient of variation [CV], %) varied with raw signal intensity. Signals below 500 IU were regarded as background noise; therefore, measurement of variance was not applicable. For signals above 500 IU precision was estimated to be 8.4%, using repeated runs for six donors over six faecal extractions per donor over 2 days (n = 328). A CV below 10% was set as a criterion in development of the DI algorithm. Based on repetitive measurements of 139 dysbiotic samples, 94% of the samples showed CVs below 10%. In addition, several in‐process test steps were evaluated (data not shown).

Faecal microbiota variation over time

Variation in microbiota over time was investigated both for normalised data across the selected probe set, and for the test result (DI). Faecal samples were collected from five donors (aged 24–38; 80% women) at a 1‐week interval for up to 14 weeks. PCA of normalised data was performed, and statistical assessment of variation in the signals for donor and sampling time (weekly) was conducted using R package ffmanova, an implementation of 50–50 multivariate analysis of variance.26

Comparison to Illumina deep sequencing

To compare the performance of deep sequencing and the GA‐test data for the gut microbiota, a total of 188 samples from 162 subjects (89 healthy subjects and 73 IBS subjects; from the training and test cohorts described in Table 2) were randomly selected. Sequencing was performed using the paired‐end 250 bp sequencing on the Illumina MiSeq platform27 at the Norwegian High Throughput Sequencing Center (UiO, Oslo, Norway). Demultiplexed Illumina readings were clustered into Operational Taxonomic Units (OTUs) using Qiime pipeline (v.1.7), StarCluster (http://star.mit.edu/cluster) and Amazon Web Services (https://aws.amazon.com, virtual machine identifier ami‐9bc9a7f2) at 97% sequence similarity. Standard tools and parameters for Qiime downstream analysis were used, such as uclust for OTU picking and Ribosomal Database Classifier for taxonomy assignment. A pre‐defined taxonomy map of reference sequence OTU to taxonomy was used rather than open‐reference picking and assignment, as the reference database of 16S rRNA sequences found in human gut is comprehensive. Thereafter, one representative read for each OTU group was extracted and aligned to create a phylogenetic tree and an OTU Biological Observation Matrix table was constructed (data not shown). The OTU table was rarefied to 5000 sequences to remove sample heterogeneity. Four samples which had less read count than the set threshold were excluded from further analysis.28

To compare the MiSeq sequence reads to the GA‐map Technology (GA‐Technology, Genetic Analysis AS, Oslo, Norway) probe signals, we identified probes that were specific for a maximum of two species or genera, and compared the normalised signals from the probes for each sample to the number of sequences of the corresponding sample and closest matching taxonomic bins found by MiSeq sequencing. If a genus found by sequencing corresponded to several probes, the sum of the probes was used in the comparison, and if a probe represented two genera the sum of sequences from both genera was used. Finally, correlation between deep sequencing data and GA‐technology data was calculated using Pearson correlation.

Furthermore, we applied the approach for defining the healthy reference collection, as described in the Dysbiosis test development section, to the Illumina sequence data set. In order to compare the results from this model with GA‐technology data, we constructed a new model using GA‐technology data limited to the same 188 samples. The samples were classified either as dysbiotic or nondysbiotic using both models, and the results compared by counting the number of samples that were classified equally across the models.

Bacterial profile in dysbiosis

Applying PCA to the validation cohort using normalised data for all 54 probes demonstrated a relative clustering of samples by disease cohorts. The scores for the first two principal components (PC), accounting for 48% of the variance in the data, showed a tighter cluster for healthy subjects in the bottom right corner compared with a more diverse spread for subjects with IBD and IBS (Figure 4a). The sample distribution in the scores plot was found to be linked to the degree of dysbiosis, with a central cluster of nondysbiotic samples surrounded by samples with weak dysbiosis (DI = 3), and the samples with the most severe dysbiosis (DI = 5) scattered outside this cluster (Figure 4b). Both the first and second PC each separate the samples from healthy volunteers from IBS and IBD samples to a certain degree. The scatter of DI values implies that different bacteria dominate dysbiosis for different samples. To further investigate which bacterial groups were the main contributors to dysbiosis in IBD and IBS, differences in overall mean normalised signal between dysbiotic and nondysbiotic status for each of the 54 probes were calculated. The pre‐dominant bacteria contributing to dysbiosis within the IBS cohort were Firmicutes (Bacilli), Proteobacteria (Shigella/Escherichia), Actinobacteria and Ruminococcus gnavus (Figure 5a). Similarly, the pre‐dominant bacteria within the IBD cohort were Proteobacteria (Shigella/Escherichia), Firmicutes, specifically F. prausnitzii, and Bacteroidetes (Bacteroides and Prevotella) (Figure 5b). Interestingly, Proteobacteria (Shigella/Escherichia) was among the top five dysbiosis‐contributing bacterial groups for both IBS and IBD, implying similarities in dysbiosis between IBS and IBD. However, all bacterial groups that contributed most to dysbiosis in the IBS cohort showed increased probe signal intensity compared to nondysbiotic patients, while for the IBD cohort, both reduced (F. prausnitzii) and increased probe signal intensities were the main contributors to dysbiosis.

Open in a separate window

Figure 4

PCA scores for the first two principal components for validation cohort (n = 287) based on 54 probes. The two PCs account for 48% of the variation, and points are coloured according to (a) cohort: yellow – healthy, blue – IBS, and red – IBD; and (B) DI: grey = 1–2, orange = 3, red = 4, dark red = 5.

Open in a separate window

Figure 5

Mean normalised signal for top five probes sorted by absolute relative difference between dysbiotic (red) and nondysbiotic (grey) as determined by the GA‐map Dysbiosis Test for (a) IBS patients (n = 109), and (b) IBD patients (n = 135). Act, Actinobacteria; B/Prev, Bacteroides/Prevotella; Firm(b), Firmicutes (Bacilli); Firm (c), Firmicutes (Clostridia); F. prau, Faecalibacterium prausnitzii; Pb, Proteobacteria; Rum.g, Ruminococcus gnavus; Sh/Es, Shigella/Escherichia.

We found a single probe with a differential signal between samples from the Spanish and Scandinavian cohorts (P < 0.01; Benjamini–Hochberg correction). The probe targets Firmicutes (Streptococcus), and this signal was found to be elevated in the Spanish samples compared to the Scandinavian cohort. Figure 6 shows the pre‐dominant bacteria contributing to dysbiosis within the Spanish samples. As expected, Proteobacteria (Shigella/Escherichia) is again found to be a contributing bacteria in dysbiosis. In addition, Bacteroides stercoris and Bifidobacterium contribute to dysbiosis, which potentially could be linked to differences, in e.g. diet between Scandinavian countries and the Mediterranean region.

Open in a separate window

Figure 6

Mean normalised signal for probes sorted by absolute relative difference between dysbiotic (red) and nondysbiotic (grey) as determined by the GA‐map Dysbiosis Test for Spanish cohort (n = 24). Bf; Bifidobacterium, B. ster; Bacteroides stercoris, Parab; Parabacteroides, Pb; Proteobacteria, Sh/Es; Shigella/Escherichia.

Faecal microbiota variation over time

Faecal samples were collected from five individuals at 1‐week intervals for up to 14 weeks. PCA of the normalised data (n = 64) revealed that most variability in the longitudinal faecal microbiota analysis was related to inter‐individual variability; donors could clearly be distinguished by the three‐first and most important PCs in the score plot (Figure 7). The samples were clustered according to faecal donor independently of sample collection time. The three first PCs described 65% of the total variability in the faecal microbiota data.

Open in a separate window

Figure 7

Scores for the first three principal components from PCA of normalised data from five healthy subjects collected weekly for up to 14 weeks (n = 64). One point is one sample for donor x taken at time point y. The first three PCs account for 65% of the variation, and points are coloured according to donor.

The significance of the PCs was analysed by ffmanova and performed using normalised data with only the main effects of donor and sampling time (weekly) included in the model. The results show that the average amount of variation between donors was greater than that within a donor (P < 0.001) with explained variances based on sums of squares of 0.48. The variation between sampling time points was not significant (P = 0.26), with explained variances based on sums of squares of 0.11. The low level of variation within one individual over time is crucial in utilising the test for monitoring changes during treatment for altering the microbiota profile.

We thank N Pedersen, F‐A Halvorsen, J Sauar, A Røseth and L Böhn for clinical samples for IBS validation, P Ricanek and S Brackmann for clinical samples for IBD validation, and L Agreus, L Kjellstrom and A Andreasson for clinical samples for normobiotic validation, and Jordi Guardiola Capón for samples from healthy volunteers from Spain. We thank the GA Scientific Advisory Board (M H Vatn [Chair], L Engstrand, W Kruis, R Löfberg, O H Nielsen, G Rogler, and L Öhman) for contribution to discussion surrounding development of the dysbiosis test.

Declaration of personal interests: GR has been a consultant for Abbot, Abbvie, Augurix, Boehringer, Calypso, FALK, Genentech, Genetic Analysis, Essex, MSD, Novartis, Pfizer, Roche, UCB, Takeda, Tillots and Vifor; has received speaker honoraria from AstraZeneca, Abbott, Abbvie, FALK, MSD, Phadia, Tillots, UCB, and Vifor; and has received educational and research grants from Abbot, Abbvie, Ardeypharm, Essex/MSD, FALK, Flamentera, Novartis, Roche, Tillots, UCB and Zeller. PM has been a consultant for AstraZeneca, MSD, Tillots and Calpro. MHV has been an advisor for Genetic Analysis and organizer of the International Advisory Board of Genetic Analysis, a member of the Advisory Board of Tillots, and has received speaker honoraria from AstraZeneca, Abbott, MSD and Falk. OHN has been a member of the International Advisory Board of Genetic Analysis. KR is a board member of Genetic Analysis. LE has been an adviser for Genetic Analysis. LÖ has received speaker honoraria from Abbvie, UCB and Takeda, and is a member of the Advisory Board of Genetic Analysis, and has received an unrestricted research grant from AstraZeneca. MS has received unrestricted research grants from Danone, and served as a Consultant/Advisory Board member for Danone, Nestlé, Chr Hansen, Almirall, Albireo, AstraZeneca, Sucampo, and Shire. C Casén, HC Vebø, M Sekelja, FT Hegge, MK Karlsson, E Ciemniejewska, S Dzankovic, C Frøyland and R Nestestog are employees of Genetic Analysis AS. C Casén, HC Vebø, M Sekelja, FT Hegge, S Dzankovic, R Nestestog and K Rudi owns stocks and shares in Genetic Analysis AS. Genetic Analysis AS owns patent invented by HC Vebø and K Rudi: Methods of amplifying a target sequence of a 16S rRNA or 16S rDNA in a prokaryotic species, US patent No. 8889358, the 16S oligonucleotide primers used in this work. Genetic Analysis AS owns patents invented by M Sekelja, K Rudi and HC Vebø: Oligonucleotide probe set and methods of microbiota profiling, US application No 13/919056 describing a set of oligonucleotide probes used for profiling the microbiotia in the GI tract used in this work. Genetic Analysis AS owns patent invented by K Rudi: Nucleic acid detection method with US patent No. 6617138, describing the method of selective binding, labeling and hybridization of a oligonucelotide probe used in this work.

Declaration of funding interests: This study was funded by an unrestricted grant from Genetic Analysis AS and the EU FP7‐Health‐2012 project; IBD‐Charter Grant agreement no: 305676. Medical writing support was provided by Carl Felton of Prime Medica Ltd (Knutsford, Cheshire, UK) and funded by Genetic Analysis AS.