Gut microbiota-derived hexa-acylated lipopolysaccharides enhance cancer immunotherapy responses

Main

Immune checkpoint inhibitors (ICI) unleash the power of a patient’s own immune system to combat cancer. Multiple landmark studies have identified a role for the gut microbiome in modifying ICI treatment response^1,2; however, there is little consensus between these studies as to which members of the gut microbiota are important for modifying treatment outcomes. More recent studies instead conclude that the relationship between the gut microbiome and ICI clinical response extends beyond the level of species composition³, pointing to associations between unfavourable immunotherapy outcomes and lipopolysaccharide (LPS)-producing bacteria in the gut microbiota⁴. However, beyond association, the effects of microbiota-derived LPS remain incompletely defined, and few studies consider the different structures of LPS produced by gut microbes.

Primarily characterized in pathogenic infections for its role in stimulating immune responses to Gram-negative bacteria, LPS activation of host toll-like receptor 4 (TLR4) has complex context-dependent roles in physiology. In the gut, LPS-TLR4 signalling has been shown to both dampen intestinal immune responses to the resident microbiota during states of health⁵ and modulate systemic autoimmune activity⁶. However, the immunostimulatory effect of LPS depends on its structure, which differs profoundly between bacterial species. Indeed, not all Gram-negative bacteria in the gut harbour all genes within the LPS biosynthetic pathway, resulting in production of a heterogenous pool of LPS structures by different gut microbes. The lipid A region of LPS that is recognized by TLR4 is composed of two glucosamine sugars linked to a variable number of acyl chains. In humans, hexa-acylated LPS potently activates TLR4 and host immunity, whereas hypo-acylated penta- and tetra-acylated LPS poorly activate TLR4 and can antagonize immune activation by hexa-acylated LPS^6,7,8,9. Recent studies have associated some LPS-encoding Gram-negative taxa, such as members of the Bacteroidaceae family, with non-response to ICI treatment^4,10,11. However, taxonomic associations with treatment response may not capture the diversity of LPS structures present and the resulting complexity of functional interactions occurring between gut-derived LPS species and host immune responses.

To determine the potential functional rather than taxonomic basis for Gram-negative bacteria-mediated modification of ICI response, we analysed LPS biosynthesis genes in faecal metagenomes of patients with cancer undergoing ICI with anti-PD-1 therapy. Using a multi-cohort meta-analysis combining taxonomic and functional annotation, we identify hexa-acylated LPS-encoding gut bacteria as a potential determinant of clinical response to anti-PD-1 therapy. Using both human and mouse in vitro assays and an in vivo mouse tumour model, we demonstrate a TLR4-dependent causative role for increased ratios of hexa-acylated LPS in the gut in enhancement of anti-PD-1-mediated anti-tumour responses. These findings further our understanding of the complex host–microbiome interactions that define response to ICI treatment and advise against the use of inhibitors of LPS-induced TLR4 signalling suggested by some previous studies^12,13.

Results

Patient microbiome LPS segregates responders to anti-PD-1

To explore the functional role of microbiota-derived LPS in responses to anti-PD-1 immunotherapy, we first analysed baseline gut microbiota composition before treatment using metagenomic sequences from a large-cohort study of anti-PD-1-treated patients with melanoma¹. Non-metric multidimensional scaling (NMDS) ordination based on taxonomy and abundance of Gram-negative bacterial genera did not significantly segregate responders from non-responders (Fig. 1a, left). However, NMDS ordination based on function, specifically abundance of genes involved in LPS biosynthesis, significantly segregated responders from non-responders and revealed differential abundance of lpxM, which mediates lipid A hexa-acylation (Fig. 1a, right). Among genes involved in bacterial LPS assembly, lpxA, B, C, D, H, K and wdtA/waaA encode genes required for basic LPS assembly and tetra-acylation of lipid A, whereas lpxL facilitates LPS penta-acylation, and lpxJ and lpxM hexa-acylate LPS with differing acyl chain lengths and positions^14,15.

Gut microbiota-derived hexa-acylated lipopolysaccharides enhance cancer immunotherapy responses — **Fig. 1: Patient responsiveness to anti-PD-1 therapy is associated with microbiome immunostimulatory hexa-acylated LPS.**

We next expanded our analyses by performing taxonomic and functional annotation on metagenomes from five published datasets encompassing 112 patients with melanoma treated with anti-PD-1 from seven clinical studies across four different countries^1,3,4,11,16, batch correcting for variance introduced by study and geographic differences (Extended Data Fig. 1 and Supplementary Tables 1 and 2). It is worth noting that only samples taken before the start of immunotherapy from patients not taking probiotics, antibiotics or medications such as proton pump inhibitors and H2 receptor antagonists were included, as these treatments are known to have significant effects on the microbiome, complicating assessment of the patient’s baseline microbiota³. On a taxonomic level, a majority of the commensal Gram-negative species in patient gut microbiotas encoded lpxL-mediated capacity for penta-acylated LPS biosynthesis, dominated by members of the Bacteroidota phylum (Fig. 1b), which has previously been demonstrated to be immunoinhibitory^7,8. While we observed considerable inter-patient variation on a taxonomic level (Fig. 1c), the abundance of total LPS biosynthesis genes was significantly higher in non-responders to anti-PD-1 therapy (Fig. 1d), as previously noted⁴. Among the genes involved in modification of lipid A, pagL (lipid A 3-O-deacylation), pagP (palmitoylation), lpxE and lpxF (dephosphorylation), lpxL (penta-acylation) and lpxJ and lpxM (different forms of hexa-acylation), only lpxM was significantly different between responders and non-responders (Fig. 1e–g), predominantly represented by Proteobacteria (homonym Pseudomonadota) (Fig. 1b). It is worth noting that these functional comparisons of microbiome LPS biosynthesis genes were consistent with those from pooled and per-study data before batch correction (Extended Data Fig. 2). These analyses confirmed a negative correlation between total gut microbiome LPS and immunotherapy response but uncovered a significant positive association between lpxM-mediated LPS hexa-acylation and favourable anti-PD-1 therapy outcomes.

Functional metagenomics reveal hexa-LPS signature of response

To validate our findings using additional unbiased clustering methods, we identified three functional enterotypes (clusters of bacterial communities in the gut) using partitioning around medoids (PAM) clustering with a Jensen–Shannon divergence (JSD) distance metric calculated from the functional annotation profiles of the metagenomes of the 112 patients with melanoma (Fig. 2a). A majority of patients with functional enterotype 2 were responders to anti-PD-1 therapy, while most patients with functional enterotype 3 were non-responders (Fig. 2b). Consistent with our other analyses, while functional enterotype 3 was characterized by a significantly higher abundance of total LPS biosynthesis genes (Fig. 2c), functional enterotype 2 was characterized by a significantly higher abundance of lpxM (Fig. 2d). Although direct measurement of lipid A acylation in complex mixtures such as faeces remains challenging, as a proof of principle, we used isolates from our culture collection¹⁷ and confirmed by liquid chromatography–mass spectrometry that gut commensal Gammaproteobacteria that encode lpxM can produce hexa-acylated lipid A (Extended Data Fig. 3).

**Fig. 2: Functional metagenomic analyses indicate elevated hexa-acylated LPS ratios are favourable for patient responses to anti-PD-1 therapy.**

Using taxonomic classification paired with pangenome annotation of LPS biosynthesis genes, we also quantified the abundance of microbial taxa that encode lpxL penta-acylated versus lpxJ and lpxM hexa-acylated LPS. On a taxonomic level, both total LPS-encoding and penta-acylated LPS-encoding taxa were significantly higher in non-responders to anti-PD-1 (Fig. 2e,f). However, differential abundance of hexa-acylation was not dominantly mediated by a single genus or species of bacteria but rather a functional convergence on total LPS hexa-acylation capacity via multiple lpxJ- and lpxM-encoding taxa in the gut microbiota of responding patients (Fig. 2g). This resulted in ratios of hexa- to penta-acylated LPS-producing taxa that were more than doubled in responders (mean lpxM-to-penta ratios are Responder = 0.9238, Non-responder = 0.4217, and mean total hexa-to-penta ratios are Responder = 1.255, Non-responder = 0.6244) (Fig. 2h,i) with significantly higher lpxM-to-lpxL ratios at a functional (gene or open reading frame (ORF)) level (Fig. 2j). It is worth noting that we also functionally annotated and examined abundance of LPS biosynthesis genes in patients with renal cell carcinoma (RCC) from a single previous study². Although lpxM also trended higher in this smaller cohort, we found in patients with RCC that lpxJ, which also mediates LPS hexa-acylation, was significantly more abundant in responders (Fig. 2k), in line with the increased abundance of Akkermansia muciniphila previously noted in this patient cohort². This also resulted in an increased ratio of total hexa-to-penta-acylated LPS genes in patients responding to anti-PD-1 (Fig. 2k), raising the possibility that similar microbiome LPS signatures are associated with anti-PD-1 efficacy in other cancer types. Collectively, these data suggested that having an increased proportion of gut bacteria that encode hexa-acylated LPS is favourable for response to anti-PD-1 immunotherapy.

Microbiome LPS and TLR4 are required for anti-PD-1 efficacy

To investigate potential mechanisms by which the acylation of microbiota-derived LPS might impact anti-PD-1 responses, we explored targeted oral antibiotic treatments to alter populations of Gram-negative gut bacteria. We treated mice with control drinking water, a cocktail of broad-spectrum antibiotics, or polymyxin B (PMB), a narrow-spectrum antibiotic known to both inhibit LPS signalling and deplete Gram-negative bacteria by binding the lipid A region of LPS¹⁸. We then performed shotgun metagenomic sequencing of faecal samples before and after treatment to assess effects on the relative abundance of different taxa in the gut microbiota of treated mice. Overall, PMB treatment significantly preserved microbial abundance and microbiota composition compared to treatment with broad-spectrum antibiotics (Fig. 3a). However, while both antibiotic regimens were equally effective at depleting Gammaproteobacteria (Fig. 3b), PMB did not significantly deplete other LPS-encoding Gram-negative taxa, with the exception of lpxJ-encoding taxa, compared to control drinking water (Fig. 3c and Extended Data Fig. 4a). By contrast, broad-spectrum antibiotics significantly depleted all LPS-encoding bacteria (Fig. 3c and Extended Data Fig. 4a), in line with their total reduction of bacterial reads (Fig. 3a). Using this selective depletion of hexa-acylated LPS-encoding Gammaproteobacteria by PMB, we investigated how altering LPS acylation ratios in the gut microbiome affects anti-PD-1 efficacy in a mouse model of tumour immunotherapy. As a proof of principle, we used the well-established anti-PD-1-responsive, syngeneic MC38 colorectal adenocarcinoma tumour model¹⁹ and pre-treated mice with PMB or a broad-spectrum antibiotic cocktail in their drinking water starting 2 weeks before subcutaneous tumour implantation (Extended Data Fig. 4b). As previously reported^1,2, broad-spectrum antibiotics significantly abrogated anti-PD-1-induced tumour reduction; however, targeted microbiome depletion with PMB equally abolished anti-PD-1 efficacy (Fig. 3d). These observations together revealed that selective reduction of gut bacteria capable of producing immunostimulatory hexa-acylated LPS was sufficient to disrupt anti-PD-1-mediated anti-tumour responses.

**Fig. 3: Microbiota-derived hexa-acylated LPS and host TLR4 are required for anti-PD-1-mediated anti-tumour immunity.**

As TLR4 is the canonical host receptor for LPS and is required for a majority of its endotoxic and immunomodulatory effects²⁰, we next investigated the contribution of host TLR4 signalling to anti-PD-1 responses. We treated tumour-bearing wild-type mice with the small-molecule TLR4 inhibitor TAK-242 only during anti-PD-1 therapy and examined the effect of treatment on tumour growth (Extended Data Fig. 4c). It is worth noting that temporally controlled inhibition of TLR4 during anti-PD-1 treatment alone resulted in significant loss of treatment efficacy (Fig. 3e,f). This was accompanied by reduced infiltration of IFNγ⁺ and TNF⁺ CD8⁺ T cells and proliferating Ki67⁺ effector T cells within tumours of TAK-242-treated animals (Fig. 3g and Extended Data Fig. 5) without significant effects on other immune cell populations (Extended Data Fig. 6). Taken together, these data indicated that stimulation of host TLR4 by microbiota-derived hexa-acylated LPS was required for effective anti-tumour immunity in response to anti-PD-1.

Hexa-LPS stimulates host immunity and enhances immunotherapy

Given that the acylation state of LPS lipid A has been shown to directly impact immune signalling through TLR4^6,21, we quantified the relative potency of hexa-acylated and penta-acylated LPS at stimulating NF-kB activation via TLR4 in both human monocyte/macrophage and mouse macrophage cell lines. As expected, in human monocytes we found that TLR4, but not TLR2, was required for LPS-mediated NF-kB activation, and stimulation with hexa-acylated LPS resulted in substantially more activation than penta-acylated LPS (Fig. 4a). By contrast, expression of CD14, a receptor that also binds LPS but does not discriminate lipid A acylation states, did not affect differential responses to hexa- and penta-acylated LPS (Extended Data Fig. 7a). Hexa-acylated LPS purified from a Gammaproteobacterium was 10,000 times more potent at stimulating NF-kB activation than penta-acylated LPS from a Bacteroidota species (Fig. 4b), resulting in higher cytokine secretion (Fig. 4c). This was blocked by PMB in a dose-dependent manner (Extended Data Fig. 7b), and although the LPS response between human and mouse TLR4/MD2 has been shown to differ⁸, similar effects on cytokine secretion (as a proxy of NF-kB activation) were also observed in a mouse macrophage cell line (Extended Data Fig. 7c,d). It is worth noting that in human monocyte/macrophages, altering the ratio of hexa- and penta-acylated LPS via the addition of low, non-stimulatory concentrations of penta-acylated LPS significantly decreased the ability of hexa-acylated LPS to activate NF-kB (Fig. 4d), confirming that ratios of stimulatory to inhibitory LPS can dictate immune activation.

**Fig. 4: Oral administration of hexa-acylated LPS enhances the efficacy of anti-PD-1 therapy.**

To examine the effects of increasing the ratio of immunostimulatory hexa-acylated LPS in the gut, tumour-bearing mice were supplemented with a low dose of hexa- or penta-acylated LPS in their drinking water during anti-PD-1 treatment. Boosting levels of hexa-acylated LPS significantly improved the efficacy of anti-PD-1 therapy, resulting in decreased tumour burden (Fig. 4e), while oral supplementation with penta-acylated LPS did not significantly affect anti-PD-1 efficacy. In addition, when examining the tumour-infiltrating immune cells from these mice, we observed a significant increase in effector cytokine-secreting CD8⁺ T cells infiltrating the tumours of hexa-acylated LPS-supplemented mice (Fig. 4f), with a trend towards decreased immune activation in penta-acylated LPS-treated mice (Fig. 4f and Extended Data Fig. 8a). Oral LPS did not result in any signs of inflammation or systemic immune activation, as no significant differences were observed in spleen or peripheral lymph nodes in mice treated with oral LPS and anti-PD-1 compared to anti-PD-1 alone (Extended Data Fig. 8b,c). This suggests that hexa-acylated LPS augmentation of anti-PD-1-mediated immunity is restricted to the tumour. Neither hexa- nor penta-acylated LPS affected tumour growth in the absence of anti-PD-1 (Extended Data Fig. 8d). Collectively, these findings show that an increased relative abundance of hexa-acylated LPS in the gut can support the immune activation required for robust anti-PD-1 efficacy in both patients and mouse models.

Discussion

Improving response rates to ICI is essential for improving survival of patients. In this initial study, we functionally annotated a large cohort of patient microbiota profiles to uncover the role of gut bacteria in ICI efficacy. Although differences in patient population based on geographic location significantly contribute to variance at a taxonomic level (even after batch correction), this source of variation contributed far less at a functional level (Extended Data Fig. 1 and Supplementary Table 2), supporting the use of our functional metagenomics approach to see past taxonomic heterogeneity to an underlying mechanism involving LPS acylation. By exploring this mechanism in animal models, we were able to demonstrate that hexa-acylated LPS increases the efficacy of anti-PD-1 ICI treatment in a TLR4-dependent manner, cautioning against use of PMB and related antibiotics in patients undergoing immunotherapy treatments and highlighting actionable potential avenues to therapeutically predict and pharmacologically improve patient outcomes.

It is important to note that inferred LPS biosynthesis capacity based on taxonomic identification and pangenome annotations did not always match our taxonomy-independent functional analyses at the level of LPS biosynthesis genes. This highlights potential limitations of taxonomy-based analyses for inferring function and may explain how meta-analyses in previous studies did not identify differential contributions of LPS hexa- and penta-acylation to immunotherapy responses. In addition, current databases and methods of functional annotation of microbiome genes are far from complete and comprehensive, necessitating our use of multiple approaches for both taxonomic and functional analyses to derive biological meaning. Methods to accurately measure different lipid A acylation states in complex environments such as patient stool samples also merit further development and optimization to determine the degree to which metagenomic profiles can reflect the heterogenous mixture of lipid A variants in the gut.

The ability of human TLR4/MD2 to discriminate LPS of distinct acylation states differs from that of mouse TLR4/MD2⁸, and it has been previously shown that hexa-acylated LPS potently activates human TLR4/MD2, whereas hypo-acylated tetra- and penta-acylated LPS poorly activate or antagonize human TLR4/MD2 activation driven by hexa-acylated LPS^22,23. Thus, while the mouse model was critical to demonstrate a requirement for hexa-acylated LPS/TLR4 signalling in anti-tumour PD-1 responses, and the ability of orally administered hexa-acylated LPS to augment such responses, we complemented our metagenomic analyses of human anti-PD-1 responses with human macrophage experiments to explore the relative activating versus immunosilencing properties of hexa- and penta-acylated LPS in vitro. Further optimization of ‘humanized’ mouse models holds promise for better in vivo characterization of the possible mechanisms of hexa-acylated LPS-mediated immunotherapy enhancement. For example, the NSG TLR4^−/− mouse model reconstituted with human haematopoietic stem cells²⁴ can be used to test effects of hexa-acylated LPS on immune cells expressing human TLR4, although this system will not account for possible roles of TLR4 activation in intestinal epithelial cells and other stromal cell types. Mice expressing only human TLR4/MD-2 have also been generated, but only male mice can be used due to integration on the Y chromosome²⁵.

Other microbiota-derived metabolites, including Bifidobacterium inosine²⁶, Enterococcus muropeptides²⁷ and bacterial flagellin²⁸, have also been implicated in modulating ICI treatment response. While it is currently unknown to what degree gut LPS signalling interfaces with signals from other microbiome-derived immunostimulatory metabolites, simultaneous activation of different pattern recognition receptors has been shown to result in synergistic immunostimulation^29,30. It is therefore likely that multiple microbiota-mediated mechanisms that modulate treatment response coexist and interact to define patient outcomes, perhaps contributing to the divergent taxonomic associations described by previous studies. In addition, host factors are likely to play a role in this mechanism of immunotherapy modulation. For instance, differential expression of host intestinal alkaline phosphatase has been shown to affect the strength of the TLR4 signal³¹, and this represents just one of many currently uncharacterized patient factors that could be explored in the future.

Numerous other aspects of the microbiome LPS–immunotherapy interaction also warrant further studies, including the contribution of TLR4 in different cell types to anti-PD-1 efficacy. Additional tumour models, such as the YUMM2.1 melanoma model and the CT26 microsatellite-stable colorectal carcinoma model, may shed light on the broad applicability of hexa-acylated LPS for enhancing anti-PD-1 efficacy in other cancer types. In addition, as shown by the heterogeneity of Proteobacteria abundance seen in human patients in Fig. 1c, the levels of LPS tolerated in the gut can be quite high. This is in stark contrast to the relatively low concentrations (1–4 ng per kg body weight) of LPS needed in the circulation to cause pathology. Several studies in rodents have indeed demonstrated that oral LPS administration at doses up to 1 mg kg⁻¹ does not cause any inflammation, pathology or associated toxicity and may instead have neuroprotective effects³². A better understanding of how and where gut-derived microbial metabolites engage host signalling pathways to determine clinical response to ICI immunotherapies will enable design of new therapeutic approaches to overcome microbiota-driven heterogeneity in immunotherapy outcomes and allow better prognostication to benefit patients with cancer.

Methods

We confirm that our research complies with all relevant ethical regulations.

Curation of publicly available datasets

We retrieved human metagenomic data from six publicly available datasets^{1,2,3,4,11,16} through the National Center for Biotechnology Information (NCBI) Sequence Read Archive (SRA) using the accession numbers ERP127050, SRP339782, SRP197281, SRP116709 and SRP115355 (ERP104577 for RCC) using NCBI sra-tools (v2.11.2). These publicly available cohorts are shown in Supplementary Table 1 (Supplementary Table 3 for RCC). We included only those samples in our study that were collected at baseline, that is, before starting the therapy. We excluded any samples from patients who received combination therapy (that is, anti-CTLA4 along with anti-PD1 or IFNγ with anti-PD1) and/or had a recent history of proton pump inhibitor, H₂ receptor antagonist or antibiotic usage whenever metadata for these parameters was available. We classified patients into responder and non-responder groups according to the outcomes reported in the respective studies above. Patients classified as having stable disease, partial response or complete response were considered responders, and patients with progressive disease were considered non-responders for the purposes of our analyses.

We are aware that working with public datasets may introduce biases and artificial variability into the analyses due to nonuniformity of the data structure. It is noteworthy that in our study we did not directly use the processed count or abundance matrices, which are known to contribute to batch effects. Along with the baseline filtering criteria mentioned above, we also implemented a uniform stringent quality control and downstream analysis method for all the samples at individual raw sequencing read levels, as detailed below. Thus, all the pooled sequencing samples were treated as part of one experiment, and the source of the sample (that is, study and country) was used as a covariate, similar to the original studies.

Quality control and preprocessing

The pre-processing pipeline consists of three main stages: (1) initial quality control by removing low-quality reads (quality score <Q30), reads <75 bp and Illumina adapters from the forward and reverse reads using Trimmomatic software (v0.39)³³; (2) DNA contamination removal using Bowtie 2 (v2.3.5.1)³⁴ in sensitive mode, removing host DNA from humans (GRCh38) and mice (GRCm39) and non-host phiX174 Illumina spike-in; and (3) checking the FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/) results and removing any samples that had >50% read duplication from the human-associated metagenomes. FastQC results were aggregated with multiqc (v1.11) for easy manual checking.

Contig assembly, gene-calling and functional annotation

Quality-trimmed host decontaminated paired-end metagenomic reads were assembled individually using metaSPAdes (v3.15.4)³⁵ to build contigs. Single-end samples (RCC and cohorts; Supplementary Table 3) were assembled using MEGAHIT (v1.2.9)³⁶. Contigs from individual samples were concatenated into a single file, and contigs smaller than 500 bp were discarded. To make a non-redundant contig database, 100% identical contigs were clustered, and a representative longer contig was kept using cd-hit-est from CD-HIT program (v4.8.1)^37,38. Prodigal (v2.6.3)³⁹ was used in ‘meta’ mode for the ORF and translated protein prediction from the non-redundant contig collection. The script sqm_annot.pl from SqueezeMeta software (v1.5.2)⁴⁰ was used to assign KEGG (Kyoto Encyclopedia of Genes and Genomes) orthology and corresponding microbial taxa to the predicted protein sequences. The databases were downloaded, using the script make_databases.pl from SqueezeMeta. In brief, functions were assigned using Diamond (v2.0.15.153)⁴¹ blastp alignments of the proteins against the KEGG⁴² and NCBI non-redundant⁴³ databases using the lowest common ancestor⁴⁴ methods. This way each protein sequence from the metagenomes was assigned to a function (KEGG) and bacterial taxa, and thus a connection between function and taxonomy was established. To quantify the predicted genes in the samples across the metagenomes, we used the Salmon tool (v1.8.00)⁴⁵. First, the predicted ORFs were indexed, and then reads (paired end or single end) from individual samples were mapped onto the indexed Salmon database.

Analyses of patient metagenomes

Metagenomes from patients with melanoma or RCC were curated and quality controlled as described above. For gene-level relative abundance, undetected (zero value) samples were excluded, and all non-zero values were reported. In patients with RCC, the sample numbers for hexa-acylated LPS were as follows: lpxM, n = 44; lpxJ, n = 50; and total (lpxM + lpxJ), n = 51. A KEGG orthologous group present in less than five samples was considered as sporulating and/or misannotation and removed from the analysis. Relative abundance of the predicted LPS biosynthetic genes (ORFs) from the patient metagenomes was compared between ‘Responder’ and ‘Non-responder’ through a single comparison, and therefore there was no need for a multiple testing correction. Before functional analysis, we used ComBat-Seq (v3.48.0)⁴⁶ to correct the batch effects from the functional profile of the cohort of patients with melanoma. ComBat-Seq was run on the read count matrix with study metadata as batch and treatment response as group (full model).

Taxonomic profiling was performed with Kraken v2.1.2⁴⁷ and Bracken v2.6.2 (https://doi.org/10.7717/peerj-cs.104) using a custom database built from representative genomes of the human gut microbiota that were re-annotated using the Genome Taxonomy Database v2.1. This database is available via Zenodo at https://doi.org/10.5281/zenodo.7319344 (ref. ⁴⁸). After analysis with Kraken2/Bracken2, we used ConQuR (v1.2.0)⁴⁹ to correct for batch effects between studies. ConQuR was run on the OTU read count matrix using a penalized fitting strategy (logistic_lasso=T, quantile_type=“lasso”, interplt=T) with study metadata as the batch variable and the response as a covariate. The resulting batch-corrected read count matrix was analysed in R (v4.3.1).

Genomic prediction of bacterial LPS structure

To predict which bacterial species of the human gut microbiota produce LPS, lpx operon genes were identified from functional pangenomes of 3,006 bacterial species using the ‘feature_search’ module of the MGBC-Toolkit v1.0¹⁷. lpx genes were identified using both the InterPro and EggNOG functional schemata. Accessions for searched lpx genes are shown in Supplementary Table 4.

For each pangenome, genes that were encoded by at least 50% of the constituent genomes were considered to be encoded by that bacterial species. As most genes could be annotated using both InterPro and eggNOG, species that were identified as encoding a gene by at least one schema were included in analyses. All annotations were manually reviewed in context of the published literature and known structures for LPS. Following this review, only the InterPro database was used for lpxM annotation as eggNOG over-identified these genes in taxa which are known not to encode them.

LPS structure was then predicted using these annotation data according to the KEGG reference pathway for LPS biosynthesis. Species encoding lpxA, C, D, B and K were considered to produce lipid IVA and were thus defined as total Gram-negative LPS-encoding taxa (n = 884), while taxa that additionally encoded kdtA/waaA were considered to produce KDO2-lipid IVA (n = 827). Taxa that further encoded lpxL and either lpxM or lpxJ were considered to have the capacity to produce hexa-acylated LPS (lpxLM, n = 127; lpxLJ, n = 112), while taxa that further encoded lpxL but not lpxM or lpxJ were defined as penta-acylated LPS producers (n = 538). LPS-encoding taxa that did not encode lpxL were defined as tetra-acylated LPS producers (n = 107).

Data normalization

Gene abundance of the single melanoma¹, and RCC cohort was expressed in the form of gene count per million (GCPM).

The GCPM within the sample was calculated as follows:

$${{rm{GCPM}}}_{i}=frac{left({q}_{i}/{l}_{i}right)}{{sum }_{j}left({q}_{j}/{l}_{j}right)}times {10}^{6}$$

where q_i denotes the number of raw reads mapped to the predicted gene, l_i is the gene length, and ∑_j(q_j/l_j) corresponds to the sum of mapped reads to the predicted gene normalized by gene length. Thus, GCPM accounts for gene length and sequencing depth and facilitates comparisons across samples. As the batch correction was performed on the raw read counts of the pooled cohort of patients with melanoma, we used a slightly different method to normalize and scale our data. The batch-corrected count matrix was log-transformed and added with a pseudo count of 10⁻⁶ followed by scaling to count per million (CPM) value.

The CPM within the sample was calculated as follows:

$${{rm{CPM}}}_{i}=frac{{A}_{i}}{{sum }_{i=1}^{n}{A}_{i}}times {10}^{6}$$

where A_i is the batch-corrected number of reads mapped to the predicted gene and n is the total number of predicted genes.

Clustering and functional enterotype analysis

Functional enterotyping and between-class analyses were performed according to the procedure described by Arumugam et al.⁵⁰ Data generated by functional annotation with the KEGG database (KEGG orthologs) were used to calculate the JSD among samples. The PAM clustering algorithm was applied to cluster the relative abundance of functional profiles of genus or orthologous groups. The number of optimal clusters was estimated using the Calinski–Harabasz index using the clusterSim package in R according to a previously described method⁵¹. The between-class analysis was performed to identify the major drivers of the metagenome functions and support clustering the orthologous groups abundance profiles into clusters. The between-class analysis is a type of ordination with instrumental variables with the highest effect sizes that maximize the separation between classes of this variable. In this study, the instrumental variable was cluster classification using PAM clustering, obtained from KEGG orthologous groups, which contributed the maximum to the principal coordinates, identified based on their eigenvalues. The analysis was performed using the ade4 package in R.

Mice

In vivo tumour experiments were performed using 9-week-old female C57BL/6 mice purchased from Charles River. Experiments were conducted in the University of Cambridge UBS Anne McLaren facility in accordance with UK Home Office guidelines and were approved by the University of Cambridge Animal Welfare and Ethical Review Board (PPL number PP3638130). Mice were kept under a consistent 12 h light–dark cycle (7 a.m. to 7 p.m.), ambient temperature (20–24 °C) and humidity 45–65%, as per the UK Home Office guidelines. All mice were fed Scientific Diets SAFE 105 rodent diet ad libitum.

Treatment of mice with antibiotics

Mice were treated with PMB (500 mg l⁻¹) (APExBio, C3090-APE) or a broad-spectrum antibiotic solution (ABX: ampicillin (1 mg ml⁻¹) (Cayman Chemical, 14417-25g-CAY), streptomycin (3 mg ml⁻¹) (Sigma-Aldrich, S9137-25G), and colistin (1 mg ml⁻¹) (BioServ UK Limited, BS-9867A-5G)) in sterile drinking water. Treatment was started 2 weeks before tumour implantation, and antibiotic water was changed twice a week.

Shotgun sequencing of mouse metagenomes

Mouse faecal pellets were collected immediately after arrival from the supplier, termed ‘Pre’ samples. Additional faecal samples were collected after 2 weeks of treatment with broad-spectrum antibiotics, PMB or control drinking water. For broad-spectrum antibiotic-treated mice, treatment was discontinued for 1 day before sample collection to allow for some bacterial DNA to be recovered. A detailed description of the mouse metagenomic samples has been provided in Supplementary Table 5. Mouse faecal genomic DNA was extracted using the FastDNA SPIN kit for soil (MP Biomedicals, 116560200) per the manufacturer’s instructions and stored in a −80 °C freezer. Shotgun metagenomic sequencing was carried out at GENEWIZ from Azenta Life Sciences on an Illumina NovaSeq -X platform. Each sample was sequenced at a minimum depth of 20 million reads (minimum 10 million paired-end reads) with a 150-nucleotide read length. Sequences are deposited in NCBI SRA under Bioproject ID PRJNA1171992.

Taxonomic classification of mouse gut metagenomes

Filtered paired-end metagenomic reads were used for taxonomic profiling of mouse gut metagenomes with Kraken v2.1.2 and Bracken v2.6.2 using the Mouse Gastrointestinal Bacteria Catalogue (MGBC) database¹⁷, as described above. To predict the LPS-producing bacterial species in the mouse metagenomes, lpx operon genes were identified using the MGBC protein database¹⁷. The abundance data were normalized by the weight of the faecal pellet used for the DNA extraction. Data were analysed and visualized in R.

Mass spectrometry of acylation of lipid A from commensal gut bacteria

Mass spectrometry analyses of lipid A were performed using an Agilent 1290 Infinity II LC system coupled to an Agilent 6546 Q-TOF instrument. Lipid A-1P standard (Sigma-Aldrich catalogue number L6638-1MG) was prepared in chloroform/methanol/water (74:23:3) at 1 mg ml⁻¹. About 5 µl of 0.5 mg ml⁻¹ of lipid A-1P was injected onto a Waters XBridge C18 column (50 mm × 2.1 mm, 3.5 µm particle size), and the mobile phase was a linear gradient of 0–95% isopropyl alcohol with 200 mM ammonium hydroxide and 5 µM medronic acid (mobile phase B) in methanol/water (80:20) with 200 mM ammonium hydroxide and 5 µM medronic acid (mobile phase A) over 17 min (before holding at 100% mobile phase A for 3 min). The flow rate was 0.25 ml min⁻¹, and the column was operated at 45 °C. Post-column addition of 10% DMSO in acetone at a rate of 0.25 ml min⁻¹ was used. The published extracted ion chromatograms by Sándor et al.⁵² were used to characterize the lipid A molecule. Increases in acyl chain length led to later chromatographic elution.

Mouse commensal Escherichia coli (isolate A7_7.EC.CDM; NCBI assembly GCA_910574425.1) and Klebsiella pneumoniae (isolate A5_5.KP.CDM; NCBI assembly GCA_910574035.1) from our mouse culture collection were confirmed to encode lpxM using the MGBC ToolKit with ‘feature_search’ function. These isolates were inoculated in 5 ml of yeast casitone fatty acid (YCFA) broth using a single colony from streaked YCFA agar plates. The cultures were incubated anaerobically overnight at 37 °C. The 5 ml cultures were used to inoculate 1 l of YCFA liquid media. One-litre liquid cultures were incubated anaerobically overnight at 37 °C to an optical density at 600 nm of approximately 1.000. Cultures were separated into 200 ml tubes and centrifuged at 3,500 × g for 20 min. Media was discarded, and bacterial cell pellets were washed once with PBS. Dry pellets were stored at −80 °C. The culture pellet was prepared at 100 mg ml⁻¹ in water, hydrolysed with a mild acid hydrolysis buffer (50 mM sodium acetate), and Bligh-Dyer extracted (proportion of chloroform/methanol/water, 1:2:0.8). Final resuspension was in chloroform/methanol/water (74:23:3), and 5 µl was injected onto the Waters XBridge C18 column (50 mm × 2.1 mm, 3.5 µm particle size), using the same chromatographic conditions as described above. The acyl chain lengths for the monophosphorylated and diphosphorylated forms of lipid A were annotated based on Sándor et al.⁵², and the proportional composition of these acyl chains was calculated for each isolate.

Tumour challenge and treatment

MC38 colon carcinoma cells were purchased from Kerafast. Cells were passaged in DMEM (Sigma-Aldrich, D5796) supplemented with 10% fetal bovine serum (FBS) (Sigma-Aldrich BCCC3714), 1% Glutamax (Thermo Fisher Scientific, 35050-038), 1% non-essential amino acids (Thermo Fisher Scientific, 11140-035), 1% sodium pyruvate (Thermo Fisher Scientific, 11360-039), 1% penicillin–streptomycin (Thermo Fisher Scientific, 15140-122), 0.1% of 2-mercaptoethanol (Thermo Fisher Scientific, 21985023), amphotericin B (Thermo Fisher Scientific, 15290-026) and gentamycin (Thermo Fisher Scientific, 15750-045). Mice under isoflurane anaesthesia were injected subcutaneously with 10⁶ cells in 100 µl of sterile 1× PBS in the right flank. Tumours were measured using callipers at days 7, 10, 14, 17 and 20 after implantation. To control for incomplete tumour engraftment, mice with a tumour volume of less than 27 mm³ at day 7 after implantation were removed from the study. Mice were injected intraperitoneally (i.p.) with 200 µg anti-PD-1 (clone 29F.1A12) (BioXCell, BE0273) or IgG2a isotype control (BioXcell, BE0089) twice a week. Hexa- and penta-acylated LPS were purchased from Sigma-Aldrich (L2630-100MG and L9143-100MG). Mice were administered LPS (25 mg l⁻¹) in sterile drinking water starting 10 days after tumour implantation. In line with our animal ethics guidelines, mice were culled if their tumour reached 1.5 cm in diameter or if the tumour became ulcerated. LPS-containing water was changed twice a week. The TLR4 inhibitor Resatorvid (TAK-242) (APExBIO, A3850) was administered i.p. on days 10, 12, 14, 17 and 19 post tumour implantation (60 μg per injection).

Sample processing of murine tumours, spleens, mesenteric lymph nodes and tumour-draining lymph nodes for immunophenotyping

Tumour and spleen samples were digested in Roswell Park Memorial Institute (RPMI) medium containing collagenase D and DNase I (Sigma-Aldrich, 11088882001 and 10104159001) at 37 °C for 30 min before dissociation through 70 µm cell strainers. A 40/80 Percoll gradient was used to isolate lymphocytes from tumours, and resulting cell suspensions were filtered using 40 µm cell strainers. For spleen samples, Red Blood Cell Lysis Buffer (Sigma-Aldrich, R7757-100ML) was applied at room temperature for 5 min to selectively lyse red blood cells. Mesenteric lymph nodes and tumour-draining lymph nodes were digested in complete HBSS with 2% FBS containing collagenase D at 37 °C for 30 min and dissociated through 70 µm cell strainers.

Flow cytometry of tumour-infiltrating and lymphoid tissue lymphocytes

Isolated lymphocytes were re-stimulated with 1 µg ml⁻¹ Brefeldin A (eBioscience, 00-4506-51), 50 ng ml⁻¹ phorbol 12-myristate 13-acetate (Sigma-Aldrich, P8139-1MG) and 1 µg ml⁻¹ ionomycin (Sigma-Aldrich, I0634-1MG) in complete RPMI with 10% FBS in a 37 °C cell culture incubator with 5% CO₂ for 4.5 h. Dead cells were then stained using the LIVE/DEAD Fixable Aqua Dead Cell Stain Kit (Thermo Fisher Scientific, L34965). Cells were then stained with extracellular antibodies for 20 min on ice and permeabilized for 30 min using the BD Cytofix/Cytoperm Fixation/Permeabilization Kit (BD Biosciences, 554714) according to the manufacturer’s protocol. The intracellular antibodies listed in Supplementary Table 6 were then added and incubated for 30 min on ice. Cells were post-fixed in 1% PFA at 4 °C until analysis on a Cytek Aurora flow cytometer. Samples with cell viability below 30% were excluded from further analyses. Quantification of cell numbers was achieved using CountBright Absolute Counting Beads (Thermo Fisher Scientific, C36950), in accordance with the protocol provided by the manufacturer. Data were exported as FCS files using Cytek Aurora software (v3.3.0) and analysed using FlowJo software (v10.8.1, Tree Star). For lymphocyte flow cytometry antibody panel, see Supplementary Table 6.

Flow cytometry of myeloid cells

The TruStain FcX Antibody (BioLegend, 101320) was used to block cells for 5 min at 4 °C. Dead cells were then stained using the LIVE/DEAD Fixable Aqua Dead Cell Stain Kit. After washing, cells were stained with the surface antibodies listed in Supplementary Table 6 in 1× PBS for 20 min on ice. Cells were fixed in 1% PFA at 4 °C until analysis on a Cytek Aurora flow cytometer. Data were exported as FCS files using Cytek Aurora software (v3.3.0) and analysed using FlowJo software (v10.8.1, Tree Star). The myeloid flow cytometry antibody panel is shown in Supplementary Table 6.

In vitro NF-kB activation in monocyte/macrophages

THP-1 cells (THP1-Dual Reporter Cells, TLR2 KO Dual Reporter THP-1 Cells, TLR4 KO Dual Reporter THP-1 Cells and THP1-Dual MD2-CD14-TLR4 over-expressing THP-1 cells) were purchased from InvivoGen. Cells were cultured in RPMI 1640 (Gibco) with 2 mM l-glutamine (Sigma-Aldrich), 25 mM HEPES (Gibco), 10% FBS (Sigma-Aldrich), 100 µg ml⁻¹ Normocine (InvivoGen), 100 U ml⁻¹ penicillin–streptomycin (Gibco), 10 µg ml⁻¹ blasticidin (InvivoGen) and 100 µg ml⁻¹ Zeocin (InvivoGen) according to the manufacturer’s instructions. For NF-kB activation assays, THP-1 cells were resuspended in Test medium (RPMI 1640 with 2 mM l-glutamine, 25 mM HEPES, 10% FBS and 100 U ml⁻¹ penicillin–streptomycin) and cultured at 10⁵ cells per well for 24 h in a final volume of 200 µl of Test medium with indicated concentrations of ultrapure LPS from E. coli (hexa LPS, InvivoGen) and/or ultrapure LPS from Porphyromonas gingivalis (penta LPS, InvivoGen) with or without PMB (APExBio) in CELLSTAR 96-well plates (Greiner). About 20 µl⁻¹ of culture supernatants was mixed with 180 µl of QUANTI-Blue Solution (InvivoGen) in 96-well enzyme-linked immunosorbent assay microplates (Greiner) according to the manufacturer’s instructions. The plates were incubated at 37 °C for 1–6 h. Optical density at 630 nm was measured on a FLUOstar Omega (BMG Labtech).

In vitro cytokine secretion in monocyte/macrophages

Human THP-1 cells were cultured as indicated above, and supernatants were collected at 24 h. The indicated cytokine cytometric beads and paired-end-conjugated detection antibodies from the BD Cytometric Bead Array Human Inflammatory Cytokines Kit (BD Bioscience) were mixed with these supernatants according to the manufacturer’s instructions. Mouse RAW 264.7 cells (Merck, catalogue number 91062702-1VL) were cultured in high-glucose DMEM (Sigma-Aldrich) with 10% FBS (Sigma-Aldrich) and 100 U ml⁻¹ penicillin–streptomycin (Gibco). About 10⁵ cells per well were cultured for 24 h in a final volume of 200 µl culture medium with indicated concentrations of ultrapure LPS from E. coli (hexa LPS, InvivoGen) and/or ultrapure LPS from P. gingivalis (penta LPS, InvivoGen) with or without PMB (APExBio) in CELLSTAR 96-well plates (Greiner). Culture supernatants were mixed with the indicated cytokine beads and paired-end-conjugated detection antibodies from the Mouse Th1/Th2/Th17 CBA Kit (BD Bioscience) according to the manufacturer’s instructions. All samples were analysed on a BD LSRFortessa Cell Analyzer (BD Bioscience).

Statistical analysis

Statistical analyses were performed using either Graphpad Prism software (v9.3.1) or R (v4.3.1). Non-parametric Kruskal–Wallis tests and parametric Brown–Forsythe tests were used as indicated to calculate statistical significance of the difference in sample medians and means, respectively, and post hoc Dunn P values with Bonferroni false discovery rate adjustment were calculated for the pairwise comparisons among the enterotypes. The proportions of clinical responders were compared between enterotypes using one-tailed two proportions Z-tests without continuity correction. Two-way analysis of variance (ANOVA) was used to evaluate statistical significance in experiments with tumour growth measurements. We used NMDS, a distance-based ordination technique^53,54 in R (vegan::metaMDS()) with default engine = “monoMDS”. In the process of calculation of goodness of fit, a single NMDS run identifies a local minimum; therefore, we iterated the runs 1,000 times, or the best solution achieved to reach the global minimum (convergence) using the function Procrustes in R (vegan:: procrustes()). The final stress score was used to assess the optimal model and goodness of fit as suggested previously^55,56. In addition to the stress score, we also performed a two-tailed t-test on the NMDS coordinates as shown previously⁵⁷ to highlight the difference between the ‘Responder’ and ‘Non-responder’ groups. P values of less than 0.05 were considered statistically significant. Statistical tests used are specified in the figure legends. Data distribution was assumed to be normal/parametric or non-parametric based on the data type, but this was not formally tested. P values correlate with symbols as follows: NS, not significant; *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001; ****P ≤ 0.0001. No statistical methods were used to pre-determine sample sizes, but where relevant, sample sizes were determined based on variability observed in published experiments of a similar kind². In some experiments prior experience of sample size requirement was used to design experimental group sizes. For experiments where technical limitations prevented adequate statistical power to be obtained from single experiments, results from multiple experiments were pooled to provide sufficient statistical power. Data reported are in most cases non-subjective and did not require randomization or blinding at measurement. Where possible, data collection and/or the organization of the experimental conditions were randomized. Investigators were not formally blinded. However, in randomized experiments it was difficult for investigators and technicians to readily determine treatment groups from animal IDs at the bench.

Reporting summary

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