Respiratory System

We use two simulation strategies to evaluate the power of the generalized UniFrac distances under various conditions. The first strategy is a modification of the simulation method proposed by Schloss (2008) (link), where we draw points (16S rRNA sequences) from a 2D circle with known densities (Fig. 1A). This strategy facilitates simulations of different community characteristics such as species evenness and species richness. The Euclidean distance between points is analogous of the genetic distance between the sequences. The diameter of the circle represents the maximum genetic divergence between any pair of sequences within a sample. The area of the circle is proportional to the richness and the density distribution of the circle is proportional to the evenness. By varying the centroid positions (o) and their radius (r), it is possible to vary the fraction of shared membership and species richness within each sample (Fig. 1B and D). By varying the point distribution on the circle (density proportional to r^α, where α controls the degree of evenness and α = 0.5 for uniformly distribution), it is possible to change the species evenness (Fig. 1C). We also simulate scenarios where lineages of different abundance levels change by a k fold (Fig. 1E–G). These are achieved by simulating the community with point mass concentrated at the circle center (r^1.0) and varying the point density in different regions of the 2D circle corresponding to abundant lineages (0–0.2r from the center; Fig. 1E), moderately abundant lineages (0.4r–0.8r from the center; Fig. 1F) and rare lineages (0.8r–1.0r from the center; Fig. 1G). We further bin the sampled points into small hexagons as ‘OTU’s before calculating the UniFrac distance [‘hexbin’ function from the R package ‘hexbin’ (Carr et al., 2011 )]. The phylogenetic tree of these ‘OTU’s is built using NJ algorithm (Neighbor Joining, ‘nj’ function in R) and rooted by midpoint rooting method. Generalized UniFrac distances are then calculated based on the NJ tree and ‘OTU’ abundances. Each replication consists of drawing 400 points from each community, a bin size of 0.015 units to form ‘OTUs’ (∼ 300 OTUs per sample), and the maximum distance between any two points is 0.3 units (r = 0.15), corresponding to typical phylum level divergence of 30% for 16S rRNA gene. These conditions allow us to simulate the sampling intensity and biodiversity found within a typical 16S rRNA gene targeted sequencing experiment (Schloss, 2008 (link)).
Fig. 1

Two simulation strategies to evaluate the generalized UniFrac distances. (A–G), 2D circle-based simulation of microbial communities with different characteristics. (A) The microbial community is represented by a 2D circle. Points are drawn from the circle to simulate the 16S-based sampling process. These points are further binned into small hexagons as OTUs. UPGMA or NJ method is used to build the OTU phylogenetic tree. Six scenarios are investigated, where the difference occurs in: community membership (B), evenness (C), richness (D), most abundant lineages (E), moderately abundant lineages (F) and rare lineages (G). The affected lineages are indicated by a red circle or ring. H, tree-based simulation of microbial communities based on the phylogenetic tree and DM model. A real OTU phylogenetic tree from a throat microbial community dataset is used. These OTUs are roughly divided into 20 clusters (lineages) by performing PAM method using the OTU patristic distance matrix. Each cluster is subjected to abundance change in response to the environment. Counts are generated from a DM model.

The second set of simulations utilize a real upper respiratory tract microbiome dataset consisting of 60 samples and 856 OTUs from Charlson et al. (2010) (link) (Fig. 1H). A common way of modeling multivariate count data is to use the multinomial model. However, the multinomial model assumes fixed underlying proportions for each sample, which do not hold for real microbiome data due to high degree of heterogeneity among the samples. The real OTU count distribution (Supplementary Fig. S1A) exhibits more variance than expected from a multinomial model (Supplementary Fig. S1B). To realistically simulate the data, it is important to model extra-variation or overdispersion of the OTU counts. This can be achieved by using the Dirichlet-multinomial (DM) model (Mosimann, 1962 ), which assumes the underlying proportions of the multinomial model come from a Dirichlet distribution. The density function of a DM random variable N is given as

where is total count, k is the OTU number and proportion mean π = (π₁,π₂,⋯,π_k) and dispersion θ are parameters. When θ = 0, it is reduced to multinomial model. We estimate the DM parameters π,θ using maximum likelihood method (‘dirmult’ function from R package ‘dirmult’). We then generate OTU counts using the DM model with the estimated parameters and 1000 counts per sample. Supplementary Figure S1C shows an OTU heatmap generated by the DM model, in which the overdispersion is similar to that of the real data. To study the power of UniFrac variants for identifying potential environmental factors, we let the abundance of a certain OTU cluster change in response to environment. We use the UPGMA tree of the OTUs based on the OTU distance matrix calculated under the K80 nucleotide substitution model (Felsenstein, 2004 ), QIIME (FastTree algorithm (?)) and partition the 856 OTUs into 20 clusters using Partitioning Around Medoids (PAM) (‘pam’ function from R package ‘cluster’) based on patristic distances (the length of the shortest path linking two OTUs on the tree). These OTU clusters are highlighted in different colors in Figure 1H.
We call the first strategy 2D circle-based simulation and the second tree-based simulation. For power calculation, we use 2000 replications.

This retrospective and cross-sectional study was conducted in Trakya University Hospital Respiratory Intensive Care Units which was approved by the Trakya University Clinical Research Ethics Committee (TÜTF-BAEK 2021/275) and the Turkish Ministry of Health (2021-06-07T10_06_44). Patients diagnosed with ARF due to lung involvement of laboratory-confirmed (RT-PCR) COVID-19 and managed with HFNC at ICU admission were included in the study between April 2020 and January 2022.
As per the Turkish Ministry of Health COVID-19 management guideline,²¹ HFNC is indicated for patients with persistent hypoxemia or respiratory distress symptoms under low flow oxygen therapy systems. HFNC was administered in the ICU with HI-Flow Star^TM (Dragerwerk AG & Co., Germany), which is set to deliver a flow rate up to 50 l/min with FiO₂ to keep the patient’s SpO₂ above 90%.
If deterioration in the patient’s level of consciousness, worsening dyspnea, malign arrhythmia, or hemodynamic instability were detected or more than 60% FiO₂ under 50 l/min flow rate was required to keep the patient’s PaO₂/FiO₂ over 150 mmHg, it was considered a treatment failure. Non-invasive ventilation (NIV) or IMV was initiated as rescue therapy.
Data were abstracted from the hospital records and nurse charts. Patients’ demographics, body mass indices, comorbidities, Charlson Comorbidity Indices,^{22 (link)} disease severity scores [Acute Physiology and Chronic Health Assessment (APACHE),^{23 (link)} Sequential Organ Failure Assessment (SOFA)^{24 (link)}] and laboratory findings (hemogram, d-dimer, ferritin, C-reactive protein, procalcitonin, arterial blood gas parameters within 2 hours thereafter HFNC initiation) at ICU admission; ROX indices at initiation, 2^nd, 8^th, 12^th, 24^th and 48^th hours of HFNC; and out-comes (ICU and hospital length of stay, in 28-day mortality) were recorded (Figure 2). ROX index was calculated using the formula (SpO₂/FiO₂)/respiratory rate.^{18 (link)} Patients were excluded who were younger than 18 years old and HFNC failed within 2 hours of the therapy.