Main

Probiotics are increasingly administered to hospitalized patients1. These supplementary products have shown benefit in acute infectious diarrhea, antibiotic-associated diarrhea, and ulcerative colitis2,10,11. In the intensive care unit (ICU), additional indications are being explored, including prevention of ventilator-associated pneumonia, pancreatitis, and sepsis12,13,14. However, studies on the efficacy and adverse effects of probiotic treatment in ICU patients show conflicting results, and their use remains controversial4,5,6,7,15. Adverse outcomes, including bacteremia, have been reported and may preclude their use in specific populations such as those with a compromised immune system or disorders of the gastrointestinal tract8,9,16,17.

Bacteremia that appears during the course of probiotic treatment can involve Lactobacillus species similar to those in probiotics, yet as these species are also common in the human gastrointestinal microbiome, pinpointing the source of these infections has been challenging18. Studies that use pulsed-field gel electrophoresis have previously revealed strain-level similarity between blood and probiotic isolates19, but higher genomic resolution is required to establish direct clonal ancestry and the possibility of direct transmission of probiotic bacteria to the blood.

In the context of bacterial pathogens, whole-genome methods have been powerful in identifying transmission links and within-host adaptation20,21,22,23. Constructing a single nucleotide polymorphism (SNP)-level phylogeny of isolates from patients can unravel ancestral links between lineages and likely paths of transmission. Whole-genome comparison of isolates can also reveal adaptive mutations important for the survival of the pathogen within the host20,22. Yet, despite their established power, the use of these whole-genome approaches for tracing of probiotic strains has so far been limited. Here, we apply whole-genome analysis and phenotyping to blood isolates and probiotic strains administered to ICU patients.

Analysis of cases of Lactobacillus bacteremia in ICU patients at Boston Children’s Hospital showed that patients receiving Lactobacillus rhamnosus strain GG (LGG) probiotics had a markedly higher risk of developing Lactobacillus bacteremia than those who received no probiotics. Over a period of 5.5 years, a total of 22,174 patients were treated in an ICU, and 522 of these patients received LGG-containing probiotic—typically through a feeding tube—as part of their treatment. Analysis of recorded Lactobacillus bacteremia among these patients showed a significantly greater risk for patients who received the LGG-containing probiotic than for those who did not; 6 of the 522 patients had Lactobacillus bacteremia (1.1%, patients R1–R6; Supplementary Table 1a) compared to only 2 of the 21,652 patients who did not receive the LGG probiotic (0.009%, patients N1–N2; P = 4.8 × 10−9, Fisher’s exact test; Supplementary Table 1a). Furthermore, all six of the ICU blood isolates from patients receiving the LGG probiotic were identified by matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry as Lactobacillus rhamnosus, while the two isolates from patients not receiving the LGG probiotic were identified as other Lactobacillus species (Supplementary Table 1a). The ICU patients receiving probiotics containing LGG are therefore at markedly higher risk of developing Lactobacillus rhamnosus bacteremia (6 out of 522 compared to 0 out of 21,652; P = 1.8 × 10−10, Fisher’s exact test). Further, the LGG probiotic bacteremia rate of 1.1% that we observed is also much higher than the annual rate of LGG probiotic bacteremia (0.00007%) reported in the general population19. Yet, L. rhamnosus bacteremia can occasionally also appear in patients not receiving these probiotics: during the study period, there were an additional ten cases of Lactobacillus bacteremia among approximately 93,000 non-ICU patients (Patients N3–N12, Supplementary Table 1b), and four of these ten isolates were identified by MALDI-TOF as Lactobacillus rhamnosus (Patients N5, N9–N11, Supplementary Table 1b). None of the ten non-ICU patients were receiving a probiotic at the time of the bacteremia. Taken together, these results suggest that ICU patients who receive probiotics containing LGG are at much higher risk of developing Lactobacillus bacteremia than patients who do not, but it is difficult to prove a direct causal relationship as occasional cases of L. rhamnosus bacteremia also appear in patients who do not receive these probiotics.

To achieve better ancestral resolution, we next used whole-genome sequencing to determine strain-level similarity among the blood and probiotic isolates. We performed whole-genome sequencing of all ten L. rhamnosus blood isolates (six from patients receiving probiotic and four from patients not receiving probiotic), as well as 16 isolates from each of three probiotic capsules of different lots (probiotic batches 1–3; Supplementary Table 2, Methods and Fig. 1a). To quantify strain-level relatedness among these isolates, we started by measuring their distance to all available L. rhamnosus genomes (GenBank, October 2017; Supplementary Table 3). Illumina reads of each isolate were aligned to each of these genomes, and the fraction of aligned reads, affected both by gene content similarity and SNP density, was quantified as a measure of similarity. We found that all six blood isolates and all probiotic product isolates shared the same closest reference genome—an LGG genome (GenBank chromosome ID: FM179322)—suggesting high relatedness between these two sets of isolates (Fig. 1b). By contrast, all four L. rhamnosus blood isolates from patients not receiving probiotics were more similar to other strains, indicating that they were not derived from the probiotic product (Fig. 1b).

Fig. 1: Genomic evidence for L. rhamnosus transmission from probiotic capsule to the blood of patients.
figure 1

a, Schematic for whole-genome sequencing of L. rhamnosus probiotic isolates, blood isolates from ICU patients (n = 6) receiving probiotics, and blood isolates from non-ICU patients (n = 4) who were not receiving probiotics. Black circles represent sequencing of multiple individual colonies for each probiotic batch but a single colony for each blood isolate. b, Similarity between L. rhamnosus isolates and available reference genomes are shown as the fraction of reads aligned to each reference. Isolates are identified by their source: four representative isolates from each of three probiotic product batches, the six blood isolates from patients receiving probiotics, and the four blood L. rhamnosus isolates from patients not receiving probiotics. c, Phylogenetic analysis of all 54 sequenced LGG isolates: 16 isolates from each of three separate probiotic batches (blue), and the six blood isolates from Patients R1 to R6 (magenta).

To further increase genomic resolution, we next compared the genomes of the blood and LGG probiotic isolates by alignment to the reference genome. Analysis of gene content of the isolates identified only a single deletion; one of the probiotic isolates of batch 2 had a large deletion of a region which included 82 genes of the reference genome FM179322 (genes 384–465, Fig. 2). Strains were also almost identical at the single-nucleotide level; analysis of SNP-level variations identified a total of only 23 SNPs among all isolates (Methods; for the list of SNPs, see Supplementary Table 4). Indeed, the greatest SNP distance between any isolate and the last common ancestor was not more than six SNPs. Two SNPs were shared by all isolates, separating them from the reference genome, and indicating that the blood and probiotic isolates share a more recent last common ancestor than the LGG clone deposited in GenBank (Fig. 1c). Moreover, the blood and probiotic isolates were phylogenetically inseparable; there was no mutation that strictly separated these two groups (Fig. 1c and Supplementary Table 4).

Fig. 2: Coverage of the LGG reference genome for the probiotic and blood Lactobacillus rhamnosus isolates.
figure 2

For each isolate (row in matrix) SNPs are marked as squares (magenta for blood isolates, blue for probiotic product isolates). Triangles (top panel) indicate all mutations identified in blood isolates (magenta triangles) and probiotic product (blue triangles) compared to the LGG reference genome FM179322. For the probiotic product, these are either high-quality SNPs in whole-genome sequencing (middle row) or diversity identified by deep sequencing of the product (bottom row, see Methods). Annotation is included for all SNPs identified in blood isolates. SNPs identified only in blood isolates are indicated with black open rectangles. Mb, megabase pairs.

Much of the genetic diversity among blood isolates mirrored pre-existing genetic diversity within the probiotic capsules. We identified 11 genomic positions that were polymorphic across blood isolates (Fig. 2). Three of these mutations, all of which were non-synonymous, were recurring mutations, observed in more than one blood isolate (H294Q in CamS, H248Y in GlvA, and Q1827R in SpcB; Fig. 2 and Supplementary Table 4). These repeatedly occurring blood isolate mutations were all shown to pre-exist in the probiotic product (these same loci were diverse within each of the three batches of the probiotic product, Fig. 2). Furthermore, the camS and glvA SNPs always appeared together in both the blood and probiotic isolates. One other polymorphic locus, identified in a single blood isolate, was also found to pre-exist in the probiotic product (a D220G mutation in the ABC transporter CcmA). Overall, correspondence between blood isolate mutations and pre-existing diversity within the probiotic product further supports the likelihood of transmission of bacteria from probiotic to blood.

In addition to the six blood isolate mutations that were found to pre-exist in the probiotic product, we identified five blood isolate mutations that did not appear in the isolates from the probiotic product, suggesting de novo evolution within the patient (Fig. 2, Supplementary Table 5). These 5 mutations were not found in any of the 16 genomes isolated from each of the 3 capsules. To further test for their possible existence in the product, we deep-sequenced capsules from five different batches, obtained both from the hospital and from a commercial pharmacy, and identified diverse loci (batches 2–6, Methods; Fig. 2, Extended Data Fig. 1, Supplementary Table 6). No pre-existing genotypic diversity was found at the loci of the five blood-isolate-specific mutations (Fig. 2, Supplementary Table 7). One of these five mutations was in an intergenic promoter mutation, two were non-synonymous coding mutations (H487D in the RNA polymerase RpoB (Fig. 3a, Extended Data Fig. 2) and A259D near the active site of the RbsK ribokinase (Extended Data Fig. 3)), and two were synonymous mutations (at G44 of the YhfS transferase and at V132 of phosphoglucomutase). These mutations, existing in the blood but not identified in the probiotic product, could represent de novo mutations selected within the patient.

Fig. 3: The L. rhamnosus blood-isolate-specific rpoB SNP occurs at the rifampin-binding site and confers rifampin resistance.
figure 3

a, Predicted structure of LGG RNA polymerase β-subunit RpoB showing the rifampin-binding site (white) with H487 of the probiotic (blue, left) and D487 of the blood isolate from Patient R1 (magenta, right). b, Rifampin susceptibility testing of blood isolates of each patient (R1–R6) compared to a probiotic isolate with no SNPs (P3-2). Bars depict the medians of three independent experiments, and error bars show the interquartile ranges. *P = 0.0021 for R1 compared to P3-2 by Kruskal-Wallis test followed by Dunn’s multiple comparisons test. The blood isolate from Patient R1 was resistant based on zone cutoffs for Staphylococcus aureus (Supplementary Table 8).

The blood-isolate-specific mutation in the rpoB RNA polymerase gene (H487D) appeared in an isolate from patient R1, who had received L. rhamnosus GG and the rifampin derivative rifaximin concurrently during the 3 months prior to bacteremia. This mutation, which changes a specific residue in the cleft of the RpoB DNA-binding site, is known to provide resistance to rifampin (Fig. 3a, Extended Data Fig. 2)24,25,26. Antibiotic susceptibility measurements showed that this blood isolate was indeed resistant to rifampin, whereas all other blood isolates as well as a probiotic isolate containing no SNPs were sensitive to this antibiotic (Fig. 3b). By contrast, susceptibilities to other antibiotics were nearly identical among the blood and probiotic isolates (Supplementary Table 8), suggesting that the R1 isolate adapted specifically to resist rifampin. Interestingly, although rifampin resistance mutations at other rpoB positions typically confer decreased fitness, mutations at the H487 position can retain fitness similar to that of the wild type27. Indeed, the R1 isolate carrying the rpoB mutation showed no significant fitness cost compared with the probiotic strain containing no SNPs (Supplementary Table 9). The specificity of the rpoB mutation to the patient receiving rifampin, together with its associated resistance and growth phenotypes, further suggest that the probiotic strains may acquire adaptive mutations that increase their fitness in the host environment.

We further considered other adaptive phenotypes. Survival in serum or human whole blood was similar among the probiotic and blood isolates (Supplementary Fig. 1), whereas the level of biofilm formation (Extended Data Fig. 4), which may lead to increased adhesion to a central venous line (CVL) and/or enhanced survival in the gastrointestinal tract, was significantly higher in the LGG blood and probiotic isolates compared to the non-LGG blood isolates (from Patients N5, N9–N11) and to the probiotic isolate P2-1 containing an 82-gene deletion, which includes the spaCBA pilus genes critical for biofilm28 (Supplementary Table 10). These results suggest that biofilm is not required for bacteremia and that the LGG probiotic products may contain mutants with markedly different biofilm phenotypes.

Our patient population was critically ill, but the patients who developed LGG bacteremia while receiving probiotics did not have the typical risk factors for Lactobacillus bacteremia such as severe immune compromise or bowel disintegrity. Furthermore, in a case-control study to compare potential risk factors for bacteremia in these 6 cases with 16 matched control ICU patients who received probiotics but did not have bacteremia (Methods), we found no significant differences in device utilization, vasopressor support, recent surgery, diarrhea, parenteral nutrition, or antibiotic exposure (Supplementary Table 11). Although the low number of patients in this case-control study may limit statistical power, our inclusion of a control group with case-control methodology represents a significant improvement over prior descriptive studies in understanding specific risk factors within the ICU. The lack of strong differences between the patients who had bacteremia and the control group that did not have bacteremia suggests that the ICU patients at risk for transmission of probiotics from product to blood may not be easily identifiable.

The exact mechanism of transmission from probiotic to blood is unclear. Nearly all of these patients had a central line, and direct contamination of the central line with a probiotic strain or with stool containing the probiotic strain could lead to the observed probiotic bacteremia. Alternatively, the probiotic bacteria could have translocated across the bowel wall. The antibiotic-resistance-related adaptation we observed could appear either prior to or immediately following the transmission of the bacteria to the blood. Our results suggest that these adaptive mutations are absent in the probiotic capsule and therefore evolved within the host environment. Nevertheless, given possible genomic variations among batches of the probiotic product, we cannot exclude the possibility that some of these presumably blood-specific mutations were present in the specific capsules given to each patient. In any case, whether they appear through rare mutations that already exist in the product or de novo during treatment, these emerging antibiotic-resistant probiotic bacteria could potentially undermine treatment efficacy. It would be interesting in future studies to explore the importance of other de novo mutations with additional in vitro phenotyping, or in animal models.

In summary, our epidemiological analysis uncovers a statistically and clinically significant risk for bacteremia associated with the use of probiotic Lactobacilli in the ICU, and genome-level analysis identified six independent cases of transmission of probiotics from capsule to blood in ICU patients treated with probiotics. Our results also provide evidence of within-host evolution of the probiotic, including acquisition of antibiotic resistance. Probiotics have shown significant benefits for acute infectious diarrhea, antibiotic-associated diarrhea, and ulcerative colitis2,10,11. However, our findings highlight that as ICU patients have increased risk for probiotic-associated bacteremia, these potential benefits must be weighed against this risk when considering the continued use of probiotics in the ICU.

Methods

Patient inclusion criteria and clinical data

Eighteen cases of Lactobacillus bacteremia were identified through usual surveillance activities of the Infection Prevention and Control program at Boston Children’s Hospital from January 2009 to June 2014 (6 patients receiving probiotics, 12 patients not receiving probiotics; Supplementary Table 1). Based on pharmacy records from January 2009 to June 2014, there were 15,736 probiotic doses administered to 645 ICU patients, including 5,859 (37%) in a medical ICU; 4,080 (26%) in an intermediate care program (ICP); 3,560 (23%) in a medical-surgical ICU; 2,114 (13%) in a cardiac ICU; and 123 (0.8%) in a neonatal ICU, all at a single center (Boston Children’s Hospital). The majority (522 of 645, 81%) of the ICU patients who received probiotics, including all 6 patients who developed bacteremia, received a probiotic containing LGG. ICU patients who were prescribed probiotics received a median of 8 (IQR 3–23) doses, with a range of 1–347 doses. Twenty-four percent of doses were given by mouth, 62% by gastrostomy or jejunostomy tube, and 14% by nasogastric or nasojejunal tube. The average numbers of doses per month did not change significantly over time, with 217 doses per month in 2009, 216 doses per month in 2010, 244 doses per month in 2011, 268 doses per month in 2012, 239 doses per month in 2013, and 249 doses per month in 2014. Probiotics were administered to only 3% of ICU patients (645 ICU patients among a total of 22,174 patients admitted to these ICUs during the study period). Probiotics were most commonly prescribed because patients had been receiving them prior to ICU admission. There were no ICU-specific guidelines for probiotic administration. Nearly all patients had a CVL at the time of the bacteremia (only Patients R2, N1, and N5 did not), so nearly all of these bacteremias met CDC criteria for central line-associated bloodstream infection (CLABSI). We did not examine the details of probiotic doses administered to non-ICU patients. The study was approved by the Boston Children’s Hospital Institutional Review Board.

Our study was not designed to assess the clinical impact of bacteremia, although we do note that these episodes of bacteremia manifested initially as clinically active infection and that nearly all of the patients were treated with intravenous antibiotics directed at Lactobacillus. CVLs were removed during treatment from two of the five patients receiving probiotics who had a CVL and from nine of the ten patients not receiving probiotics who had a CVL. Notably, two of the six cases of LGG bacteremia (patients R2 and R4) and one of the four cases of non-LGG L. rhamnosus bacteremia (patient N5) were considered by their treating physicians as potential contaminants or transient bacteremias and were not specifically treated with a long course of antibiotics. However, as patient R4 had a CVL, the bacteremia was classified as a CLABSI. None of the patients had endocarditis, and none died within 7 days of bacteremia.

Isolation of bacteria from probiotic capsules and blood

Blood isolates were frozen at the time of isolation by the clinical microbiology laboratory and were later streaked on MRS-agar plates, and the Lactobacillus species were identified by MALDI-TOF. To isolate DNA for whole-genome sequencing, single colonies were picked (16 colonies per probiotic product batch and a single colony for blood samples) to inoculate MRS broth (BD Biosciences, 288130), and overnight cultures were frozen. To isolate individual bacteria from probiotic capsules, we used two complementary techniques. For probiotic product batch 1, a capsule was streaked on a CDC Anaerobe Blood Agar plate, bacterial lawn was scraped off the plate, frozen (−80 °C, glycerol) and then streaked to single colonies on MRS-agar plates (BD Biosciences, 288210) incubated at 37 °C. For product batches 2 and 3, capsules were resuspended and thoroughly vortexed in PBS, streaked on MRS-agar plates incubated at 37 °C until colonies showed, and then re-streaked to purity.

Whole-genome sequencing of individual isolates and deep sequencing

For single isolates, DNA was extracted from frozen overnight bacterial cultures derived from single colonies picked as described above (Macherey-Nagel, NucleoSpin 96 Tissue). For capsule deep sequencing (probiotic product batches 2–6, Supplementary Table 2), DNA was extracted from 100 μl (>108 cells) of resuspended capsule (same kit as above). Batch 1 was not available for deep sequencing. Nextera sequencing libraries were prepared29 and sequenced in an Illumina HiSeq 2500 machine in rapid mode to produce 125base paired-end reads. DNA extraction and library preparation for single-isolate deep sequencing used as control (see “Genomic analysis of capsule deep sequencing” below) was carried out in a similar manner to other single isolates, and sequencing was done in the same Illumina run as the capsule DNA deep sequencing.

Genomic data analysis of isolates

Illumina reads were filtered to remove reads contaminated by the Nextera adapter or low-quality bases (>2 bases with a Phred Score of <20), yielding an average of 1.13 M reads per sample (s.d. = 2.9 x 105). These reads were aligned to indicated reference genomes using Bowtie 1.2.1.1, allowing a maximum of 3 mismatches per read. The fraction of aligned reads was used to determine the distance between isolates and reference genomes. Alignment to the closest reference genome (Genbank FM179322) was analysed further. Base calling was done using SAMtools and BCFtools 0.1.19. A genome position was identified as a SNP if more than a single allele was identified across isolates using a quality threshold of FQ < −80. Phylogeny was based on the identified SNPs and was determined by the PHYLIP dnapars algorithm, which carries out unrooted parsimony.

Gene content analysis

For each isolate, a ‘raw copy number’ for each gene was calculated as the median base coverage across the gene divided by the median coverage across the genome of the isolate. To remove gene-specific biases, this raw copy number was further normalized by the median raw value of the gene across all isolates, yielding the gene copy number used to identify deleted genes. For Fig. 2, a similar analysis of genomic coverage was performed, in which for each 6 kb region the mean read coverage was divided by the median coverage across the genome and normalized by the median of this region-specific value across isolates.

Genomic analysis of capsule deep sequencing

Reads were filtered and aligned to the reference genome as described for single-isolate analysis above (GenBank: FM179322). This resulted in coverage of >97% of the reference genome. Per batch, median coverage of these positions was 372–1,268× (Supplementary Table 2). As a control, a single-isolate colony (batch 2, isolate 15) was also sequenced at high depth (median coverage 897×). To identify variable loci in the probiotic batches, for each probiotic batch and genomic position, we performed a Fisher’s exact test to compare the number of reads calling the reference with the alternative base in the probiotics versus the single-isolate control. To control for multiple comparisons, a P value of (0.05)/(genome length) = 1.66 × 10−8 was used for calling within-batch diversity.

Bacterial strains and culture conditions for in vitro assays

L. rhamnosus bacteria were grown at 37 °C with 5% CO2 for 48 h on Trypticase Soy Agar II with 5% sheep blood (BD Biosciences). Liquid culture was performed using MRS broth (Sigma-Aldrich) supplemented with 0.001% Tween 80 (MRST) (Sigma-Aldrich) at 37 °C with 5% CO2 for 24 h statically. Pseudomonas aeruginosa PAO1 galU mutant30 were grown at 37 °C overnight on Trypticase Soy Agar II (BD Biosciences). Liquid culture was performed using LB broth Miller (Fisher BioReagents) at 37 °C, at 200 rpm overnight. Modified TSB (mTSB) medium consisted of 15 g l−1 TSB (BD Biosciences) and 20 g l−1of Bacto-proteose peptone no. 3 (BD Biosciences) was used for the biofilm assay.

Biofilm assay

The assay for biofilm formation was based on a previous report, with minor modifications31. In brief, 3 × 107 colony-forming units (c.f.u.) were added to 200 μl of mTSB in three replicates in flat bottom polystyrene 96-well plates (Costar) at 37 °C with 5% CO2 for 72 h. Bacteria were removed by inverting the plate. The plate was then washed with water, and attached bacteria were stained for 30 min with 200 μl 0.1% (wt/vol) crystal violet in an isopropanol–methanol–PBS solution (volume ratio of 1:1:18). Plates were washed with water, left to dry for 15 min, and then 150 μl of 33% glacial acetic acid was added to each well. Biofilm was measured at 570 nm (Versa max, Molecular devices).

Antibiotic resistance testing

Disk diffusion susceptibility testing was performed by the Boston Children’s Hospital Infectious Diseases Diagnostic Laboratory using the standard methods for Staphylococcus aureus (specific disk diffusion methods for Lactobacillus are not available)32.

Competition assay

The single culture and competition assays were based on a previous report, with modifications33. In brief, bacteria of the probiotic strain containing no SNPs (P3-2) and of the blood isolate R1, from overnight MRST liquid cultures were adjusted to an OD600 of 0.05, either in single culture or mixed in a 1:1 ratio. During growth in MRST broth at 37 °C with 5% CO2 for 24 h CFUs were determined every 2 h by serial dilutions on MRST-agar plates for the single cultures and on both MRST-agar and MRST-agar with 1 μg ml−1 rifampicin (Research Products International) for the competition cultures.

BLASTP of RpoB protein

The protein accession numbers of the RpoB protein from the bacteria LGG (CAR88393.1), S. aureus M1112 (EWR31828.1), Stahylococcus epidermidis RP62A (AAW53580.1), Enterococcus faecium 343-3 (AAO00728.1), E. faecium 38-15 (AAO00731.1), E. faecium 40-4 (AAO00730.1), Escherichia coli K-12 substr. MG1655 (NP_418414.1), Bacillus velezensis CC09 (ANB47365.1) were used in COBALT for amino acid alignment from NCBI (https://www.st-va.ncbi.nlm.nih.gov/tools/cobalt).

Whole blood killing assay

Bacteria were grown on trypticase soy agar (TSA) with 5% sheep blood (BD Biosciences) and incubated for 48 h at 37 °C with 5% CO2. Overnight cultures in MRST medium were washed once in PBS (Boston Bio-products) and adjusted to give 106 CFU per 50 μl. Next, 50 μl of each strain were added to 450 μl of heparinized blood from a healthy donor. Inoculum c.f.u. values were determined by serial dilutions on TSA with 5% sheep blood. After 1 h and 3 h of rotation at 37 °C, serial dilutions were plated to determine the number of surviving c.f.u. In parallel, static tubes were held at 37 °C as a non-phagocytosis control for all time points (0 h, 1 h, and 3 h).

Serum sensitivity assays

Bacteria were grown on TSA with 5% sheep blood (BD Biosciences) and incubated 48 h at 37 °C with 5% CO2. Static overnight cultures in MRST medium at 37 °C with 5% CO2 were washed in PBS once and then diluted in PBS plus 1 mM CaCl2 and 1 mM MgCl2, and 100 μl aliquots were placed in a sterile 96-well plate to give a final inoculum of approximately 5 × 106 c.f.u. per well. Pooled male, type AB human serum (Sigma-Aldrich) was diluted in PBS plus 1 mM CaCl2 and 1 mM MgCl2 to give twice the desired final concentration. Final serum concentrations that were used were 50% and 25%. Human serum (50%) that was heat-inactivated by incubation at 56 °C for 30 min, and 0% serum served as controls. Equal volumes (100 μl) of sera and bacterial suspensions were mixed and incubated at 37 °C for 1 h with gentle shaking. An aliquot from each well was serially diluted and then plated on TSA with 5% sheep blood after incubation for 48 h at 37 °C with 5% CO2 for enumeration. A serum-sensitive, rough lipopolysaccharide (LPS) strain of P. aeruginosa (PAO1 galU30), grown on TSA at 37 °C overnight and then in liquid culture in LB at 37 °C overnight, was used as a positive control.

Case-control study methods

As all cases of Lactobacillus bacteremia in patients receiving probiotics occurred in an ICU, cases were matched with up to 3 control patients who had received probiotics in an ICU within 90 d of the case and had similar or longer ICU exposure prior to censoring. Controls were selected randomly using incidence density sampling. Censoring occurred on the date of bacteremia diagnosis or, for controls, on the date of discharge, death, or transfer from the ICU. Five cases had 3 controls identified, whereas 1 case with a particularly long ICU stay, had only 1 possible control identified. As a result, 6 cases were compared with 16 controls in our analysis.

Patient data for the 6 cases and 16 controls were collected retrospectively by chart review onto a standardized case report form. Temperature, white blood cell (WBC) count, C-reactive protein (CRP), and 30 d mortality were collected from the date of bacteremia or censoring. Clinical variables previously associated with either Lactobacillus bacteremia or with central line-associated bloodstream infections in general were collected for the 7 days prior to bacteremia, death or discharge34,35,36,37.

Immunodeficiency was defined as active cancer diagnosis, solid organ or stem cell transplant, primary immunodeficiency, receipt of immunosuppressant within 6 weeks prior to censoring, or neutropenia (absolute neutrophil count (ANC) or total WBC count of <500 occurring for at least two days and within three calendar days before or after the date of censoring). Medical device data included endovascular prosthetic material, CVL, tracheostomy tube, gastrostomy tube, and urinary catheter. Gastrointestinal breakdown included documentation of mucositis, diarrhea, or skin breakdown around the gastrostomy or jejunostomy insertion site. Diarrhea was identified by documentation in the physician or nursing notes or by stool output of >20 ml kg−1 in a 24 h period as per the CDC definition of mucosal barrier injury38. Antibiotic data included information on oral and intravenous antibiotics regardless of indication.

For the case-control study, odds ratios for continuous and categorical variables were generated by exact conditional logistic regression using SAS 9.4.

Statistical analysis

Fig. 3b: Kruskal–Wallis test (P = 0.0297, Kruskal–Wallis statistic = 13.99) followed by Dunn’s multiple comparisons test to P3-2 were performed. α = 0.05. Statisical results relating to Fig. 3b: See Supplementary Table 12a.

Extended Data Fig. 4: 3 independent experiments were performed on different days. In each experiment, each bacterial isolate had 3 technical replicates. P < 0.0001 by ANOVA with Tukey’s multiple comparisons test for the pairwise comparison of any of the isolates making biofilm (defined as OD570 > 1) compared to either P2-1, N5, N9, N10, N11, or medium control. There were no statistically significant differences among the isolates making biofilm or among the isolates not making biofilm. F = 38.93. d.f. = 42. For statistical results relating to Extended Data Fig. 4: see Supplementary Table 12b.

Supplementary Fig. 1a: 2 independent experiments on different days were performed. Error bars show the interquartile range of 3 technical replicates for all apart from 50% heat-inactivated (h.i.) PAO1 galU control, which had 2 technical replicates. *P = 0.0448 for PAO1 galU 50% serum versus 50% h.i. by Kruskal–Wallis test (P = 0.0297, Kruskal–Wallis statistic = 26.88) followed by Dunn’s multiple comparisons test.

Supplementary Fig. 1b: 2 independent experiments were performed on different days. Error bars show s.d. of 3 technical replicates. The ratios t1h/t0h and t3h/t0h, (t = time), were used for statistical analysis. For t1h/t0h, P = 0.1893 by one-way ANOVA followed by Dunnett’s multiple comparisons test. F = 1.677. There were no statistically significant differences upon multiple pairwise comparisons against P3-2. P = 0.7058 for P3-2 versus P1-1, P = 0.9998 for P3-2 versus R1, P = 0.5002 for P3-2 versus R2, P = 0.9020 for P3-2 versus R3, P = 0.9547 for P3-2 versus R4, P = 0.8192 for P3-2 versus R5, P = 0.2698 for P3-2 versus R6. d.f. = 15. For t3h/t0h: P = 0.1901 by one-way ANOVA followed by Dunnett’s multiple comparisons test. F = 1.658. There were no statistically significant differences upon multiple pairwise comparisons against P3-2. P = 0.8893 for P3-2 versus P1-1, P = 0.9998 P3-2 versus R1, P = 0.8571 for P3-2 versus R2, P = 0.9998 for P3-2 versus R3, P = 0.8957 for P3-2 versus R4, P = 0.2034 for P3-2 versus R5, P = 0.9353 P3-2 versus R6. d.f. = 16.

Extended Data Table 9: 3 independent experiments on different days were performed. In each experiment, 3 independent bacterial cultures of each strain were used. Values shown are the median with 25% and 75% percentiles. For single culture experiments, doubling time: P > 0.9999, Mann–Whitney U = 4, number of divisions: P > 0.9999, Mann–Whitney U = 4 by unpaired two-tailed Mann–Whitney test. For the competition culture experiment, doubling time: P = 0.1, Mann–Whitney U = 0, no. of divisions: P = 0.1, Mann–Whitney U = 0 by unpaired two-tailed Mann–Whitney test.

Please refer to the Nature Research Reporting Summary for additional details.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.