2.1 Study Design and Participants

Our prospective observational study recruited patients with IS between January 2020 and March 2022 at the Hospital de la Santa Creu i Sant Pau (Barcelona, Spain). Inclusion criteria were age ≥ 18 years and an IS diagnosis confirmed by an expert neurologist based on neuroimaging data.

Excluded were patients who had used antibiotics or probiotics in the month prior to IS onset, patients with recent infections, and patients with a prior disability. Pre-IS functional status was determined by a modified Rankin Scale (mRS) score ≥ 3, as assessed by an mRS-certified stroke neurologist via a simplified dependency questionnaire administered to either the patient or a first-degree relative [S1].

The study adhered to the ethical principles of the Declaration of Helsinki. The hospital's Ethics Committee approved the research protocol (IIBSP-MAE-2020-12), and all participants provided written informed consent to participate.

2.2 Variables

Detailed demographic, clinical, and epidemiological data were prospectively collected from each participant, as follows: age and sex; cardiovascular risk factors (hypertension, dyslipidemia, diabetes mellitus, atrial fibrillation, ischemic heart disease, previous IS, tobacco, and ethanol use); prior medication (antiplatelet agents, oral anticoagulants, statins); stroke etiology as classified according to the Trial of Org 10172 in Acute Stroke Treatment (TOAST) [S2]; baseline and 24-h post-IS neurological deficit severity as assessed using the National Institutes of Health Stroke Scale (NIHSS); type of reperfusion therapy; enema use; and stool consistency evaluated using the Bristol Stool Scale [S3].

On hospital admission, using a structured questionnaire, an mRS-certified stroke neurologist assessed the patient's baseline mRS score (pre-stroke) and 3-month mRS score during outpatient follow-up. Scores of 0–2 and 3–6 were categorized as favorable and unfavorable outcomes, respectively. For individuals unable to attend follow-up in person, the mRS score was assessed in a structured telephone interview.

2.3 Sample Collection

Stool samples collected from each patient following IS onset were processed according to a standardized protocol. The first sample, obtained by a healthcare provider, was immersed in a 96% ethanol solution, and stored at −80°C. Institutional guidelines dictated that oral laxatives be introduced on day 3 of hospitalization if no sample had been collected. If necessary, cleansing enemas were administered from day 5 onwards. As a result, all samples were obtained within five days or less from symptom onset.

2.4 DNA Extraction and Shotgun Sequencing

Stool samples were defrosted and 0.2 g of each was transferred to a 2-mL microcentrifuge tube for DNA extraction using the TIANamp Stool DNA Kit (TIANGEN BIOTECH). Sample concentration was measured with a Qubit fluorometer, and sample integrity and purity were determined with quantitative agarose gel electrophoresis. Genomic DNA was used to construct the libraries, which were pooled and sequenced using Illumina's NovaSeq 6000 System.

2.5 Bioinformatics

Metagenomic DNA sequences were processed and decontaminated from human genome reads using FastQC [S4] v0.11.8. Reads were aligned via BBMap [S5], trimmed, and filtered with Trimmomatic [S6] v0.39. SortMeRNA [S7] v4.2.0 was employed for 16S rRNA sequence extraction, while taxonomic classification was determined using MAPseq [S8] v2.0.1alpha, with the Silva v132 database used for prokaryotic gene assignment.

Microbial sequence read abundance was normalized to account for varying sequencing depths. Paired-end sequences were merged using BBMerge [S5] v38.84, with both merged and unmerged reads assembled into contigs via MegaHit [S9] v1.2.9. Contigs were then annotated using Prokka [S10] v1.14.6, leveraging multiple databases to assign functions to predicted coding sequences.

To quantify functional categories and pathways, original reads were mapped onto assembled contigs using Bowtie2 [S11] v2.3.5. Functional profiles derived from quantified metagenome results enabled comparisons between the favorable and unfavorable post-IS outcome groups. Prokka annotations were used to infer Enzyme Commission (EC) [S12] functions and Kyoto Encyclopedia of Genes and Genomes (KEGG) [S13] pathways.

2.6 Data Filtering and Statistical Analyses

Statistical and bioinformatic analyses were performed using R v4.2.0. Normally distributed and homoscedastic continuous variables were compared using the Student t-test, while non-normally distributed data were analyzed using the Mann–Whitney U-test. Dichotomous variables were evaluated using the chi-squared test.

Gut microbiota diversity was assessed through α-diversity, which evaluates microbial diversity within a single sample, and β-diversity, which compares microbiota composition between samples. α-diversity was estimated using the Shannon, Simpson, observed species, Chao1, and abundance-based coverage estimator (ACE) indexes. The observed species index accounts for the total number of features, while Chao1 and ACE indexes estimate richness (species count). Shannon and Simpson indexes assess both species diversity and distribution evenness. β-diversity was analyzed using principal coordinates analysis (PCoA), with group variation estimated via permutational multivariate analysis of variance (PERMANOVA).

Linear discriminant analysis (LDA) effect size (LEfSe) [S14] was used to identify the species that differed significantly between the two outcome groups; a logarithmic LDA score ≥ 2 and an adjusted p < 0.05 were considered significant. The LEfSe algorithm employs the Kruskal–Wallis rank sum test to detect bacteria with differential abundance, while LDA evaluates their relevance.

A linear model, adjusted for age, sex, IS subtype (TOAST classification), and stroke severity (NIHSS), was used to discriminate taxa and metabolic pathways that differed significantly between the two outcome groups. These clinical variables, selected via logistic regression, analyzed variable relationships with 3-month outcomes (Table S1).

To identify disease-associated taxonomic signatures, we used Taxon Set Enrichment Analysis (TSEA) via the MicrobiomeAnalyst module [S15], a variation of Gene Set Enrichment Analysis (GSEA) adapted for microbiome sequencing data. Similar to GSEA, TSEA performs hypergeometric tests against a taxon set library to identify relevant microbial signatures.

Sequencing errors and low-level contaminants were managed by removing rare taxa, defined as those present in < 25% of samples [S16]. To mitigate false-positive associations from multiple testing, p-values were adjusted using the Bonferroni correction (adjusted p < 0.05).

2.7 Mendelian Randomization Analyses

To evaluate potential causal relationships between significant gut bacteria, metabolic pathways, and post-IS outcomes, Mendelian randomization (MR) analyses were conducted using publicly available genome-wide association study (GWAS) summary statistics (Table S2). Exposure GWAS included abundance data for multiple bacterial taxa (Pseudomonadales [S16], Enterococcaceae [S16], Enterococcus [S16], Bacteroides fragilis [S17], Faecalibacterium prausnitzii [S17], Paraprevotella clara [S17], Sutterella wadsworthensis [S17], Eubacterium eligens [S17]), as well as pyruvate [S18] and S-adenosyl-L-homocysteine [S19] (SAH). Outcome GWAS data were sourced from GISCOME [S20], a meta-analysis of 12 IS recovery cohorts (6165 cases from across Europe, the USA, and Australia) that evaluated functional status 60–190 days post-IS using mRS dichotomous scoring (0–2 vs. 3–6), adjusted for age, sex, and IS severity.

MR analyses were conducted using the TwoSampleMR [S21] package. Summary statistics were filtered (p < 5 × 10⁻⁶) and clumped using the European 1000 Genomes Project reference panel (r² < 0.01, clump distance > 10,000 kb) [S22]. Ambiguous and palindromic single nucleotide variants (SNVs) were excluded during harmonization.

MR methods applied were inverse-variance weighted (IVW), MR-Egger, weighted median, penalized weighted median, and weighted mode, with IVW serving as the primary approach. Horizontal pleiotropy and heterogeneity were assessed using Egger regression [S23] and the Cochran Q statistic [S24], respectively. To detect potential outliers influencing results, leave-one-out analysis was performed for significant MR results.

The minimum causal relationship required for 80% MR power (α = 0.05) was estimated using Burgess' online calculator, considering sample size conditions and the proportion of variance explained by the instruments [S25].

2.8 Reporting Guidelines

This article follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) and STROBE-MR [S26] reporting guidelines.

3.1 Study Participants

Of a total of 731 eligible subjects, metagenomic sequencing was performed on fecal samples from 128 patients with IS, 84 (65.6%) with favorable and 44 (34.4%) with unfavorable 3-month functional outcomes (Figure S1). Stool samples were collected between days 1 and 4 in 82% of patients.

Patient baseline characteristics are reported in Table 1. Regarding variability between the two outcome groups, patients with unfavorable outcomes were older, were more likely to have a history of hypertension, experienced greater stroke severity, and more frequently underwent mechanical thrombectomy procedures; they were also more frequently administered enemas, which would explain stool consistency differences found between the two groups.

3.2 Gut Microbiota Diversity

We identified 46 classes, 78 orders, 131 families, 279 genera, and 240 species, whose distributions were measured within the samples. Class-level α-diversity analyses of the microbial community showed that bacterial richness (observed species index, p_class = 0.017; Chao1 index, p_class = 0.026; ACE index, p_class = 0.017) was statistically higher in patients with favorable 3-month post-IS outcomes (Figure 1A). For orders, families, genera, and species, however, α-diversity differences were not statistically significant (Table S3).

Regarding taxa distribution in the samples, we plotted PCoA based on Bray-Curtis dissimilarities and measured the β-diversity using PERMANOVA. A strong separation was observed in microbial structure at the family (p_family = 0.030), genus (p_genus = 0.035), and species (p_species = 0.018) levels between patients with favorable and unfavorable outcomes (Figure 1B).

3.3 Gut Microbiota Taxonomic Differences

Analysis using LEfSe revealed that the unfavorable outcomes group had a greater abundance of pathogenic bacteria, including Enterobacter sp. UPIB, Erysipelotrichaceae bacterium 6_1_45, Clostridium clostridioforme, Clostridium bolteae, Bacteroides fragilis, Eisenbergiella tayi, Clostridium symbiosum, Hungatella hathewayi, and Neglecta timonensis, as well as a lesser abundance of the commensal bacteria Faecalibacterium prausnitzii, Paraprevotella clara, Blautia stercoris, Sutterella wadsworthensis, Roseburia faecis, Coprobacillus sp. 8_1_38FAA, and Eubacterium eligens (Figure 2A, Table S4).

Further statistical analysis adjusting for age, sex, stroke severity, and stroke subtype revealed different relative taxa abundance between the two outcome groups (Table 2). For the unfavorable outcomes group, Synergistia was increased at the class level; Synergistales and Pseudomonadales were increased at the order level; Synergistaceae, Pseudomonadaceae, and Enterococcaceae were increased at the family level; Pseudomonas, Finegoldia, Porphyromonas, and Enterococcus were increased at the genus level; and finally, Butyrate-producing bacterium SL6/1/1 was increased at the species level (Figure 2B). TSEA, using the different relative taxa abundance between the two groups, showed an association between our taxon set library and stroke (p = 0.036). Nevertheless, the MR analysis did not point to any causal relationship between these bacteria and post-IS functional outcomes (Table S5). MR results were not biased by heterogeneity nor pleiotropy (Table S6), although MR power calculations suggested that the MR analysis was underpowered (Table S7).

3.4 Gut Microbiota Functional Differences

Mapping of metagenomic sequences against the KEGG pathway [S13] database revealed functional profiles for the samples. A total of 165 KEGG pathways were annotated, revealing significant differences between the outcome groups in the ethylbenzene degradation pathway (map00642), which was increased in the unfavorable outcomes group (p = 4.9 × 10⁻², β = −0.561, 95% CI: −0.893, −0.237) (Figure 3A).

Metagenomic sequences were also mapped against the EC Number [S12] database and classified into 1225 genes. Enzyme 2.1.1.193, also known as 16S rRNA (uracil1498-N3)-methyltransferase, was reduced in the unfavorable outcomes group (p = 5.0 × 10⁻², β = 0.101, 95% CI: 0.052, 0.150) (Figure 3B).

The MR analysis showed a causal relationship between pyruvate, one of the products of ethylbenzene degradation, and post-IS functional outcomes (p_IVW = 0.014, β = −0.502). Furthermore, although it did not reach significance, the MR analysis pointed to an association between levels of SAH, as the main product of 16S rRNA (uracil1498-N3)-methyltransferase, and post-IS functional outcomes (p_IVW = 0.150, β = 0.182) (Table 3). MR results were not biased by heterogeneity, pleiotropy, or outliers (Table S8).

4.1 Study Strengths and Limitations

This study leveraging shotgun metagenomic sequencing to identify microbiota signatures linked to 3-month post-IS outcomes has several limitations. Our findings may lack broad generalizability given that the study was conducted in a single center, although this ensured a more homogeneous sample with reduced external variability. Patient groups differed, although bias was mitigated by adjusting for multiple clinical variables. Notably, the proportion of patients undergoing mechanical thrombectomy—an important prognostic factor—varied between groups. Dietary and physical activity data were not collected, although hospitalized patients received similar diets, and the first fecal sample was collected at stroke onset, meaning that microbiota profiles may have been altered as time passed. A limitation that we plan to address in future studies was that we could not conduct additional assays or metabolomic profiling due to budget constraints. Despite adjusted analyses, baseline differences—such as vascular risk factors—may limit results validity. Finally, as an observational study, causal links between microbiota and 3-month functional outcomes could not be established, and although we identified relevant associations, the statistical power for the MR analysis was insufficient. To refine understanding of the gut microbiota's role in IS recovery, future research should focus on larger sample sizes, more comprehensive patient data, and mechanistic investigations.