For a full explanation of approved Women’s Flat Track Derby Association rules, refer to www.wftda.com. Briefly, a bout consists of two 30-min periods, where two competing teams, each composed of up to 4 “blockers” and 1 “jammer”, circle a track with the goal of facilitating their own jammer in accumulating points. Players, both blockers and jammers, periodically rotate with players on the bench, so that few or no players actually play for the entire 60 min bout. Points are accrued when one team’s jammer makes her second, and subsequent, pass through the pack of blockers, in effect lapping the pack. Activity occurs in intervals called “jams”, and a single jam lasts for a maximum of 2 min. Flat track roller derby is a contact sport; blockers are allowed to initiate contact with another player to compete for track position using any of the following body parts: upper arm (shoulder to elbow), torso, hips, “booty” (official WFTDA nomenclature), and mid to upper thigh. Roller derby tournaments often involve multiple pairwise bouts in a single day between several teams, one home team and multiple visiting teams from different geographical locations. Players within a team practice together on a regular basis, and thus come into frequent physical contact, and live in or near the same city. Teams involved in this study were from Eugene, OR (Emerald City Roller Girls); Washington, DC (DC Roller Girls) and San Jose, CA (Silicon Valley Roller Girls).

Written consent forms were signed and collected from all participating subjects. The Institutional Review Board Initial Application Form for the study was reviewed and approved by the University of Oregon IRB with the Office for Protection of Human Subjects in January 2012 (protocol #10262011.038). The Willamalane Park and Recreation District Human Resources office granted written permission for the study to take place in their recreation facility. Written permission was acquired from the three teams’ coaches and administrators.

Microbial communities inhabiting skin vary greatly across the human body (Grice et al., 2008; Grice et al., 2009; Human Microbiome Project Consortium, 2012). We chose the upper arm (approximately the distal end of the lateral deltoid) as our focal skin sample site. The upper arm is the one skin region on roller derby skaters that is nearly universally exposed and frequently contacted during a bout. All players sampled had this area exposed during the entire bout. All samples were collected Feb. 10, 2012, at the “Big O” Tournament in Eugene, OR, USA, and all biological samples were taken by trained technicians using sterile technique. The two bouts that were sampled took place at 12:00pm (Emerald City vs. Silicon Valley) and 6:00pm (Emerald City vs. DC). DC had already played in one bout the same day at 10:00am (against Santa Cruz, not considered here), but Emerald City and Silicon Valley had not played that day prior to bout 1. For the purposes of this study, players were not monitored between bouts. Samples were collected by swabbing individual’s upper arms in a c. 4 cm by 5 cm area of skin with nylon-flocked swabs (COPAN Flock Technologies, Brescia, Italy) moistened with sterile buffer solution (0.15M NaCl, 0.1M Tween20). Both arms were swabbed on each player at each sampling point, and all samples were taken within 30 min of the beginning and end of each bout. Samples were stored at −20 °C until DNA extraction. Total number of jams was recorded for each player, and multiplied by 2 min (maximum jam length) to approximate total time played per person. Four swab samples were also taken from the floor of the facility (track) following the tournament using the same swabbing method and surface area as the arm samples.

Whole genomic DNA was extracted using the MO BIO PowerWater DNA Isolation Kit (MO BIO Laboratories, Carlsbad, CA) according to manufacturers instructions with the following modifications: swab tips were incubated with Solution PW1 in a 65 °C water bath for 15 min prior to bead beating; bead beating length was extended to 10 min since swab tips were included; and samples were eluted in 50 µL Solution PW6. Dual-arm samples from each player were combined for DNA extraction.

A fragment of the 16S rRNA gene including the V4 region was amplified using a modified F515/R806 primer combination (5′-GTGCCAGCMGCCGCGGTAA-3′, 5′-TACNVGGGTATCTAATCC-3′) (Caporaso et al., 2011b; Claesson et al., 2010). Amplification proceeded in two steps using a custom Illumina preparation protocol, where PCR1 was performed with forward primers that contained partial unique barcodes and partial Illumina adapters. The remaining ends of the Illumina adapters were attached during PCR2, and barcodes were recombined in silico using paired-end reads. Adapter sequences are detailed in Supplemental Data. All extracted samples were amplified in triplicate for PCR1 and triplicates were pooled before PCR2. PCR1 (25 µL total volume per reaction) consisted of the following ingredients: 5 µL GC buffer (Thermo Fisher Scientific, U.S.A.), 0.5 µL dNTPs (10 mM, Invitrogen), 0.25 µL Phusion Hotstart II polymerase (Thermo Fisher Scientific, U.S.A.), 13.25 µL certified nucleic-acid free water, 0.5 µL forward primer, 0.5 µL reverse primer, and 5 µL template DNA. The PCR1 conditions were as follows: initial denaturation for 2 min at 98 °C; 22 cycles of 20 s at 98 °C, 30 s at 50 °C and 20 s at 72 °C; and 72 °C for 2 min for final extension. After PCR1, the triplicate reactions were pooled and cleaned with the QIAGEN Minelute PCR Purification Kit according to the manufacturers protocol (QIAGEN, Germantown, MD). Ten µL of 3M NaOAc (pH 5.2) was added to decrease the pH of the pooled reactions and facilitate efficient binding to the spin column during cleanup. Samples were eluted in 11.5 µL of Buffer EB. For PCR2, a single primer pair was used to add the remaining Illumina adaptor segments to the ends of the concentrated amplicons of PCR1. The PCR2 (25 µL volume per reaction) consisted of the same combination of reagents that was used in PCR1, along with 5 µL concentrated PCR1 product as template. The PCR 2 conditions were as follows: 2 min denaturation at 98 °C; 12 cycles of 20 s at 98 °C, 30 s at 66 °C and 20 s at 72 °C; and 2 min at 72 °C for final extension. Amplicons were size-selected by gel electrophoresis: gel bands at c. 440bp were extracted and concentrated, using the ZR-96 Zymoclean Gel DNA Recovery Kit (ZYMO Research, Irvine, CA), following manufacturer’s instructions, quantified using a Qubit Fluoromoeter (Invitrogen, NY), and pooled in equimolar concentrations for library preparation for sequencing. Samples were sent to the Georgia Genomics Facility at the University of Georgia (Athens, GA; www.dna.uga.edu), and sequenced on the Illumina MiSeq platform as paired-end reads.

Raw sequences were processed using the FastX Toolkit (http://hannonlab.cshl.edu/fastx_toolkit) and the QIIME pipeline (Caporaso et al., 2010). Barcodes were recombined from paired-end reads, and forward reads were used for downstream analysis. All sequences were trimmed to 112 bp, including a 12 bp barcode, and low quality sequences were removed. Quality filtering settings were as follows: minimum 30 quality score over at least 75% of the sequence read; no ambiguous bases allowed; 1 primer mismatch allowed. After quality control and barcode assignment, the remaining 1,368,938 sequences were binned into operational taxonomic units (OTUs) at a 97% sequence similarity cutoff using uclust (Edgar, 2010). The highest-quality sequences from each OTU cluster were taxonomically identified using reference sequences from Greengenes (DeSantis et al., 2006). Plant-chloroplast and mitochondrial OTUs were removed. Not all samples returned the same number of sequences. Following rarefaction precedents (e.g., Human Microbiome Project Consortium, 2012; Kuczynski et al., 2010) we rarefied all samples to 500 sequences per sample. Samples with fewer than 500 sequences were not used in subsequent analyses (Table 1); since some low-yield samples were removed from all team groups, players were not considered as paired samples before vs. after a bout. Samples analyzed for this study were compiled from two separate MiSeq runs, and additional aspects of this study were included in the MiSeq runs but are not considered here, so the returned sequence count does not reflect the full volume of the runs. Three of the track samples contained enough sequences to be considered in analysis, and were processed exactly as the rest of the samples, but were not used in any ordination analysis. Sequence files and metadata for all samples used in this study have been deposited in MG-RAST (ID 4506457.3–4506498.3).

All statistical analyses were performed in R. Community variation among samples, or β-diversity, was calculated using the quantitative, taxonomy-based Canberra distance, implemented in the vegan package (Oksanen et al., 2011). Non-metric multidimensional scaling (NMDS) was performed using the bestnmds function in the labdsv package (Roberts, 2010), using 20 random starts. Discriminant analysis of within-group similarity was conducted using permutational MANOVA with the adonis function in vegan. To determine whether skin microbial communities became more similar to one another after playing in a bout, we used a β-dispersion test with the betadisper function in vegan. This test is a multivariate analog of Levene’s test for homogeneity of variances (Anderson, Ellingsen & McArdle, 2006), and it tests for a significant difference in sample heterogeneity between groups (i.e. the spread of data points in ordination space). Indicator analysis (Dufrene & Legendre, 1997), using indval in labdsv, was conducted on each team group, and all players combined before vs. after respective bouts, to identify OTUs responsible for observed β-diversity results. P-values for significant indicators were adjusted for multiple comparisons using Holm’s correction (Holm, 1979). The relationship between time played and change in community composition was assessed with Pearson’s correlation test by comparing individual players’ pairwise community distances with their estimated cumulative times during a bout.

Bacterial communities detected on players’ upper arms from different teams were significantly different before playing a bout, as well as after playing a bout (Table 2). In other words, the skin microbiome of an individual player was predicted by team membership. Teams clustered together in ordination space using a non-metric multidimensional scaling representation of players’ skin microbiomes both before (Fig. 1A) and after (Fig. 1B) playing a bout, based on Canberra taxonomic distances. Though team clustering is significant in both cases (before and after a bout), there is a greater degree of overlap between the teams following bouts. Emerald City was considered as two different teams in two different bouts during analysis.

The home team’s pre-bout bacterial communities (EC) were more similar on average to communities detected on the track than the two visiting teams when considered together (p < 0.001; from a Welch two-sample t-test; Fig. 2), and when considered separately (p = 0.001 & 0.007; for bouts 1 & 2, respectively). All players were more dissimilar on average from track samples (mean Canberra distance = 0.89) than from all other players (mean Canberra distance = 0.83; p < 0.001; from a Welch two-sample t-test). Player bacterial communities did not become more similar to the track after a bout, and in fact both bout 2 teams (EC & DC) became less similar to the track following a bout (p = 0.008 & 0.003, respectively; from Welch two-sample t-tests).

When teams were considered separately, bacterial communities detected on players’ upper arms before a bout were significantly different than those detected after the bout in all cases (Table 2; Fig. 3). We also detected a signal of already having played in a bout. Two teams, Emerald City and DC, had already played in a bout the morning of the tournament, and that was a significant predictor of community composition before the second bout (F-statistic = 2.16; p-value <0.001; EC  & DC  in Fig. 1A).

All players’ skin microbiomes were more similar to one another after competing in a bout. To test this we conducted a β-dispersion test, which compared all players before and after bouting. A significant reduction in β-dispersion between groups (before vs. after) confirmed that communities became more similar (based on Canberra distances; F = 11.79; p < 0.001; Table 4; Fig. 4) when all players were considered together, and when players were grouped by the bout in which they played (F = 12.41 & 4.11; p = 0.002 & 0.048; Table 4). A greater proportion of OTUs was shared by competing teams after bouting compared to before (Table 5), but, interestingly, the opposite is true in 3 out of 4 cases when comparing non-competing teams. When each team was considered separately, both teams in bout 1 experienced a significant β-dispersion reduction as did EC after playing in bout 2, while DC did not (Table 4; Fig. 4). None of the four teams experienced a significant shift in Shannon-Wiener OTU diversity or evenness after playing in a bout, nor was there a difference when all players were considered together (all p-values > 0.2). Both teams in bout 2 had already played a bout previously in the day; neither team in bout 1 had played during the same day. Changes in bacterial communities before and after a bout were not correlated with each players time spent in a bout (Pearsons correlation test; ρ = 0.12; p = 0.45). Jammers and blockers play different roles on a team, and thus engage in different amounts of contact and time played; however, since players are not limited to a single position during a bout, we did not differentiate between the positions during analysis.

We conducted indicator analysis to identify OTUs responsible for the observed difference between teams and for the convergence of bacterial communities after playing. Forty-nine OTUs were significant indicators for single pre-bout teams (p-value <0.05) before multiple comparison adjustment, and 11 were significant after adjustment (Table 3); indicator taxa are presented at the finest taxonomic resolution provided by Greengenes assignment. Though the significant indicators include bacterial genera commonly associated with soil and plants (e.g. Dietzia & Xanthomonas), notable human-associated groups also stand out (e.g. Coprococcus, Streptococcus, Lachnospiraceae & Brevibacterium). This list of indicators includes both rare OTUs that account for <1% of sequences, and relatively common OTUs (Alicyclobacillus made up >25% of pre-bout SI sequences). We also identified OTUs that were shared between competing teams in post-bout samples but not in pre-bout samples. Six OTUs fit this description for both bouts, belonging to six different bacterial genera: Streptococcus, Sphingomonas, Eubacterium, Porphyromonas, Aerococcus and Methylobacterium. It is worth noting that although these OTUs showed a shared response after bouting, all six OTUs were relatively rare throughout the study; none accounted for more than 1% of total sequences for any player.