Multiomics-based molecular subtyping based on the commensal microbiome predicts molecular characteristics and the therapeutic response in breast cancer

To investigate the microbiota-related metabolic pathways shared by cancers, we analysed 16 S rRNA sequencing data of the gut microbiota obtained from four public datasets (PRJNA86188, PRJNA817689, PRJNA639644, and PRJNA658160) (Fig. 1A and Supplementary Tables 1–3). PRJNA861885 contained data for 428 CRC specimens and 260 normal samples. PRJNA817689 and PRJNA639644 contained data for 124 GC specimens and 140 normal samples, and PRJNA658160 contained data for 350 BC specimens and 308 normal samples. Using the Wilcoxon test and a random forest model, we identified significantly differentially abundant bacterial genera between the normal groups and the cancer groups (Fig. 1B-D, Methods). Based on the differentially abundant bacterial genera, all patients were clustered into three subgroups by the self-organizing map (SOM) method. Each cancer cohort was divided into the G1, G2, and G3 subgroups (Fig. 1E-F and Supplementary Table 4), which exhibited distinct gut microbiota characteristics at the phylum and genus levels (Supplementary Fig. 1A-I). For all three cancer types, the differentially abundant genera were tumour specific and enriched in different subgroups. For example, the BC cohort had 25 unique differentially abundant genera, and the GC and CRC cohorts had 24 and 19 different genera, respectively. Only seven genera were shared among the three cancers (Fig. 1H). Escherichia Shigella was enriched in the G1 subgroup of CRC patients, G3 subgroup of GC patients, and G2 subgroup of BC patients. These results suggest that focusing on different bacterial genera has limited the understanding of the development of different cancer types. Therefore, using Phylogenetic Investigation of Communities by Reconstruction of Unobserved States (PICRUSt) software and one-way analysis of variance (ANOVA), we identified Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways that exhibited significant differential enrichment among the subgroups. We focused on 36 differentially enriched metabolic pathways shared by the three cancer types and found that their enrichment trends were consistent among the subgroups (Fig. 1I and Supplementary Table 5). For example, we observed significant alterations in cysteine and methionine metabolism in the G3 subgroups of the BC, CRC, and GC cohorts, consistent with previous reports [29, 30]. These results showed that although the microbiomes of different tumours have different microbial compositions, they have conserved effects on these 36 metabolic pathways, implying that these shared metabolic pathways may play important roles in tumour development.

The expression of genes associated with the shared microbial metabolic pathways was then assessed. We integrated multiomics data from The Cancer Genome Atlas breast cancer (TCGA-BRCA) dataset, such as gene expression profile, clinical phenotype, RNA-seq and clinical data, to develop a new BC subtyping system. A total of 700 genes associated with gut microbiota-related metabolic pathways and patient survival were identified and used for clustering the TCGA-BRCA patients into four clusters using distance-based k-means clustering [31] (Fig. 2A and Supplementary Table 6). Each cluster exhibited distinct gene expression patterns and hallmark pathways, with Cluster 4 showing the best prognosis and Cluster 1 showing the poorest prognosis (Supplementary Fig. 2A-B and Fig. 2B). Analysis revealed significant enrichment of immune-related pathways in Cluster 4, validating the accuracy of our subtyping method (Supplementary Fig. 2C-D). Furthermore, the clusters demonstrated differences in PAM50 molecular subtyping [32] (P = 4.95e-142, 95% CI [0.49, 1.00]), stage distribution (P = 0.02, 95% CI [0.00, 1.00]), and TNBC incidence (P = 4.87e-65, 95% CI [0.48, 1.00]). The four clusters also exhibited different clinical characteristics (Fig. 2E-G). Cluster 2 included patients with all PAM50 molecular subtypes, such as the LumA, LumB, Her-2, basal and normal-like subtypes. Cluster 3 predominantly included patients with the LumA and LumB subtypes, and Cluster 4 consisted primarily of patients with the LumA subtype (Fig. 2E). P (Fig. 2F). Cluster 2 was significantly enriched in TNBC patients (Fig. 2G). Additionally, at the genomic level, we evaluated the tumour mutation burden (TMB), aneuploidy score, fraction of genome alterations, and MSIsensor score. At the immune level, we calculated the abundances of CD4 + T cells, CD8 + T cells, neutrophils, and myeloid dendritic cells (Fig. 2H-O). Cluster 1 and Cluster 2 had the highest TMB values, while Cluster 3 and Cluster 4, especially Cluster 4, had the lowest TMB values, consistent with the results of our previous prognostic analysis. However, notably, Cluster 2 was associated with the highest TMB value but not the worst prognosis. This discrepancy may be related to the complex immune environment of Cluster 2, as demonstrated in our findings. Overall, our multiomics-based subtyping method captures distinct molecular and immune characteristics of BC.

A new subtype of BC, termed “challenging BC”, was identified using the novel BC subtyping method developed in this study. Cluster 2 exhibited more genetic mutations and a more complex immune microenvironment than did the other clusters, leading to its designation as the “challenging BC” subtype. Cluster 2 contained patients with all of the traditional subtypes, including the LumA, LumB, Her-2-positive, basal, normal-like, and TNBC subtypes (Fig. 2C and E). Notably, TNBC patients were significantly overrepresented in Cluster 2. The inherent complexity of treatment for Cluster 2 patients poses substantial challenges and underscores the clinical significance of this subtype. To further analyse the “challenging BC” subtype, a score index was proposed based on gene expression and its independent prognostic coefficient (Methods). Each patient was assigned a score, and Cluster 2 exhibited the highest degree of score dispersion (Fig. 3A and Supplementary Table 7). Patients were then divided into the high-score and low-score groups based on the median score, and analysis revealed significantly poorer survival outcomes in the high-score group (Fig. 3B). Moreover, the score index emerged as an independent prognostic factor, with area under the receiver operating characteristic (ROC) curve (AUC) values of 0.634 and 0.671 for 365-day and 3-year overall survival (OS), respectively (Fig. 3C and Supplementary Fig. 3A). Subsequent analysis demonstrated significant enrichment of cancer-related and immune-related pathways in the high-score group within Cluster 2, confirming the association of Cluster 2 with poorer survival (Fig. 3D-E). Furthermore, notable differences between the high- and low-score groups in Cluster 2 were observed at both the immune and genomic levels, surpassing the differences observed in the other clusters (Fig. 3F-I and Supplementary Fig. 3B-G). These findings underscore the utility of the score index as an independent prognostic factor for the “challenging BC” subtype.

To explore immune cells linked to the prognosis of “challenging BC”, we analysed single-cell sequencing data from two patients with “challenging BC”, both of whom were pathologically diagnosed with TNBC (Fig. 4A and Supplementary Table 8). Initially, we constructed a classification model by leveraging the random forest algorithm, integrating the expression profiles of 700 genes with the classification data shown in Fig. 2. We then meticulously screened the expression of these 700 genes in each cell through single-cell sequencing and determined the average expression level of each gene. With this model, we were able to predict the subtype of the samples and assign a score to each cell based on the patient’s prognosis. The patients were then stratified into the high-score and low-score groups, and the model proficiently identified them as having the “challenging BC” subtype (Fig. 4A, Methods). We categorized the 16,282 cells that passed quality control into five major cell types (Fig. 4B and Supplementary Fig. 4A-C): tumour cells, three types of immune cells (natural killer T [TNK] lymphocytes, B cells, and myeloid cells), and stromal cells. Notably, a significant proportion of tumour cells exhibited higher scores than did cells of the other types (P = 0.00, 95% CI [0.69, 1.00], Fig. 4C). Each cell type exhibited the expression of its well-known marker genes with high specificity (Fig. 4D). Pathway enrichment analysis based on the high- and low-score groups revealed significant differences in immune-related pathways, particularly T-cell-related pathways (Fig. 4E). Further comparison of the pathways of TNK cells revealed enrichment of immune-related pathways in the low-score group (Fig. 4F-G). TNK cells were clustered into eight distinct subtypes. CD8 + CCL5 cells, which are cytotoxic T cells, were more abundant in the low-score group and were potentially associated with a better prognosis [33, 34] (P = 1.44e-31, 95% CI [0.16, 1.00], Fig. 4H-J). Conversely, CD4 + FOXP3 cells, representing Treg cells, and CD8 + CXCL13 + ITGAE cells, representing tissue-resident T cells, were more prevalent in the high-score group, possibly contributing to the poorer prognosis observed in this group [35] (Fig. 4J). The results of the multicolour immunofluorescence experiments confirmed these findings (Fig. 4K), suggesting that specific immune cell populations are associated with the prognosis of “challenging BC”.

Spatial information plays a crucial role in the understanding of transcriptional heterogeneity and the cellular spatial distribution. In this study, we utilized ST-seq to acquire in situ gene expression profiles from four patients with “challenging BC”. The quality control results indicated significant differences among the 15 identified clusters (Supplementary 5A-C and Fig. 5A). Using a classification model and calculation methods, all four patients were identified as “challenging BC” patients who were pathologically diagnosed with TNBC. Two patients (SXR_1 and SXR_2) belonged to the low-score group, and the other two patients (YZL_1 and YZL_2) belonged to the high-score group (Fig. 5A). Although all 15 cell types were detected in each patient, their proportions and marker gene expression levels varied greatly (Fig. 5B and Supplementary Fig. 5D-F). Malignant tumour cells were more abundant in the high-score group, whereas the low-score group exhibited greater abundances of immune cells, especially cytotoxic immune cells, such as cytotoxic (CD8+) T cells, consistent with the results of prognostic analysis and single-cell sequencing (Fig. 5C, Supplementary Fig. 4D-E and Supplementary Fig. 5G-H). Specifically, the proportion of CD8 + CCL5 cells was greater in the low-score group, and CD4 + FOXP3 and CD8 + CXCL13 + ITGAE cells were more prevalent in the high-score group. These findings are consistent with previous findings (Figs. 4J-K and 5C). According to the single-cell sequencing results, most of the tumour cells in both the low- and high-score groups were ARHGAP15 and TPK1 cells, respectively (Supplementary Fig. 4F). Moreover, significant colocalization of tumour cells with specific types of immune cells was observed (ARHGAP15 cells and CD8 + CCL5 cells in the low-score group; TPK1 and both CD4 + FOXP3 and CD8 + CXCL13 + ITGAE cells in the high-score group), demonstrating spatial consistency with the ST-seq data (Fig. 5D). Pearson correlation analysis revealed positive correlations between gene expression in these cells (CD8 + CCL5 and ARHGAP15, p value = 0 [Fig. 5D]; TPK1 and both CD4 + FOXP3 and CD8 + ITGAE, p values = 1.6e-10 and 6.8e-08 [Fig. 5E]), further suggesting the influence of tumour cell populations on T-cell enrichment and the immune microenvironment. Analysis of the shared immune signalling pathways associated with prognosis identified potential pathways related to good (NOD-like receptor signalling) and poor (Hedgehog and mTOR signalling) prognoses in the low- and high-score groups, respectively (Fig. 5F-H). To validate these findings, we established a PDX model using tumour tissues from a patient in the high-score group and treated the mice with inhibitors targeting the identified signalling pathways (Fig. 5I). The inhibitors used were sonidegib and rapamycin, which are a clinically approved SMO inhibitor that inhibits Hedgehog signalling pathway activity [36‐38] and an immunosuppressive mTOR inhibitor [39, 40], respectively. Significant inhibition of tumour growth was observed (Fig. 5J-K), and the identified pathways were suppressed, confirming the importance of the immune signalling pathways associated with poor prognosis identified by the score index.

Our subtyping method and score index were also validated in a neoadjuvant therapy dataset containing data for 221 patients who received anthracycline and/or taxane-based therapy [41]. We used the classification model to determine molecular subtypes using gene expression profiles, integrating these data with survival data to derive the corresponding scores (Methods). Our objective was to assess the impact of molecular subtype on the treatment response and treatment efficacy (Supplementary Table 9). Although we observed no significant difference in the score distribution between the residual disease (RD) and pathologic complete response (pCR) groups (Supplementary Fig. 6A), the molecular subtyping method and the evaluation of molecular features yielded consistent results in this dataset (Fig. 6A-E). The four clusters exhibited varied responses to treatment, with the RD samples in Clusters 1, 2, and 4 (but not those in Cluster 3) having significantly higher scores (Fig. 6F-I). This pattern underscores the utility of our novel molecular subtyping method for BC, as it correlates the score index with treatment efficacy, indicating that higher scores are associated with lower treatment efficacy. Similar results were observed for the TNBC subset (Fig. 6J and Supplementary Fig. 6B-C). Furthermore, we analysed differential gene expression and pathway enrichment between the pCR and RD groups of TNBC patients in Cluster 2 (Fig. 6K-L). Immune activation-related pathways, such as Cytokine − cytokine receptor interaction, T-cell receptor signalling pathway, and Natural killer cell-mediated cytotoxicity, were significantly enriched in the pCR group (Fig. 6L). Correlation analysis between the highly expressed genes in the pCR group and the scores revealed a significant negative correlation, suggesting that the expression of immune-activating genes, such as NKG7, CD3E, CD247, GZMA, and IL6R, increased with decreasing score (Fig. 6M). These results indicate that the score index calculated using our subtyping method can serve as an indicator of the immune microenvironment and predict the treatment efficacy and response in BC patients receiving neoadjuvant therapy.