Integration of multiomics data shows down regulation of mismatch repair and tubulin pathways in triple-negative chemotherapy-resistant breast tumors

TNBC patients enrolled in the prospective BEAUTY clinical trial neoadjuvant study were treated with NAC and were followed up for at least 4 years. We have previously compared pCR and resistant disease in pre-NAC TNBC tumors [6]. We identified several pathways significantly associated with pCR in the TNBC cohort, including the glucocorticoid receptor (GR) regulatory network; FOXA1, FOXA2, and FOXA3 networks; integrins; SMAD2/3; and androgen receptor signaling. In this study, we focused on the molecular analysis of patients who did not respond to NAC treatment (non-pCR patients—who failed to respond to paclitaxel and anthracycline/cyclophosphamide combination therapy). We aimed to identify genomic signatures associated with recurrence using the pre- and post-NAC treatment data. From the BEAUTY breast cancer NAC study, we analyzed RNA-Seq data from 18 paired (pre-NAC and post-NAC) triple-negative breast tumors. Our previous investigation and existing studies have demonstrated that the luminal androgen receptor (LAR) TNBC tumors are biologically and clinically distinct from non-LAR tumors. Hence, we removed the three TNBC LAR tumor pairs identified by LAR-Sig from further analysis. In addition, tumor bed analysis was performed using an in-silico deconvolution method, indicating that three post-NAC biopsies contained low luminal and basal epithelial compositions. Spearman’s correlation to the pathological assessment of in situ cancer cellularity was observed to be 0.53 among basal epithelial cells and 0.50 when considered together (summed) with mature luminal epithelial cells. Similarly, the correlation between our deconvolution model assessment of immune cells and the immune assessment score from the ESTIMATE package was 0.59 and 0.41 for the pre-NAC and post-NAC samples, respectively. While the correlation between our stromal assessment and the ESTIMATE package's stromal assessment score was 0.83 and 0.81 for the pre-NAC and post-NAC, respectively. Hence, we removed those three samples from the analysis. We investigated 12 paired non-LAR TNBC samples with matched pre- and post-NAC gene expression data (Table 1). Of these 12 patients, four had an early recurrence (less than two years after the completion of NAC), and eight remained recurrence-free for more than four years. Furthermore, for these 12 non-LAR TNBC patients, all patients had pre-NAC WES data, and 9/12 had paired protein and phosphoprotein (RPPA) data.

We evaluated 20,543 genes assayed from both pre-and post-NAC biopsies. A co-inertia plot using the first two principal components of paired samples is presented in Fig. 1A. In the plot, we noted a slight separation trend between patients in the early-recurrence (ERC; n = 4) group and the nonrecurrent (NRC; n = 8) group for the first principal component. This component explains 20.3% of the variance observed in the pooled dataset. We present the distribution of correlations in Fig. 1B. We observed a higher intra-sample correlation of paired pre- and post-NAC data than the inter-sample correlation within pre- and post- NAC data. Additionally, we see a small notable decrease in intra-sample correlations with the NRC cohort as opposed to the ERC cohort. Our evaluation of the pre-NAC transcriptome gene expression between the ERC and NRC groups identified only 11 genes to be differentially expressed after correcting for multiple testing with a false discovery rate (FDR) < 0.05.

Conversely, the comparison of early-recurrent and nonrecurrent tumors in the post-NAC gene expression data identified 660 genes differentially expressed with at least twofold change and FDR < 0.05 (Fig. 1C, Additional file 2: Table s3). Among these genes, 478 (72.4%) were up regulated in ERC tumors (Fig. 1C). The list of the top 23 significant genes based upon an FDR < 1e-4 is shown in Fig. 1D. Further over-representation analysis (ORA) identified 10 cytobands that were significantly enriched in the 660 DE genes (Fig. 1E), including 17q (17q25.3, 33 genes, p-value less than 10^–15 and 17q25.1, 15 genes, p-value 1.01 × 10^–1), 6p (6p22.2, 14 genes, p-value 1.40 × 10^–10), 10q (10q22.3, 9 genes, p-value 6.51 × 10^–5), and 1q (1q, 8 genes, p-value 9.39 × 10^–5; 1q24.3, 3 genes, p-value 0.007; and 1q23.2, 4 genes, p-value 0.01).

Since WES data of post-NAC tumors were unavailable, we examined copy number alterations in the pre-NAC patient tumors between the ERC and NRC groups. We identified a common gain/amplification in large regions of the chr1q (q21.3-q24.2,395 altered genes) and chr17q (q25.3, 97 altered genes) in at least 3 out of 4 recurrent patients; however, they were not detected in any of the eight nonrecurrent patients (p-value <  = 0.012 by Fisher's exact test). The 1q and 17q gain events involved 492 genes (Additional file 2: Table s4). Similarly, we also identified a deletion of 4 genes (in 4p) among the ERC group. A circos plot depicting the altered copy number regions is presented in Fig. 2. We also present histogram views of the median copy gains in the early-recurrent samples for the 17q and 1q regions (particularly 1q24), along with the elevated expression levels observed with these copy gains using pre-NAC and post-NAC RNA-Seq data (lower panels of Fig. 2, respectively).

We investigated the somatic mutations in the pre-NAC sample biopsies using WES tumor and blood DNA data. We identified 392 high/moderate impact somatic mutations corresponding to 366 genes in the four early-recurrent patient tumors and 545 high/moderate impact mutations in 519 genes in the eight nonrecurrent patient tumors (Additional file 2: Table s5). Due to the small sample size, low mutation frequency across genes and low overlap among the genes were observed between the ERC and NRC groups. It was noted that the most frequently mutated genes across our cohort were TP53 (all four ERC samples and five of eight NRC samples) and MUC4 (two of the four ERC samples and three of eight NRC samples), consistent with a previous publication [47].

We evaluated protein expression (232 proteins) and phosphorylation (63 phosphorylation sites) data obtained from the pre-NAC and post-NAC tumors using RPPA. The pre-NAC RPPA data were compared between the ERC and NRC groups. We observed that 30 proteins were significantly different (p-value < 0.05), including ten phosphorylation sites (10/63, 15.9%) and 20 total proteins (20/232, 8.6%) among the pre-NAC samples. In contrast, there were only four protein level variations, and no phosphorylation site differences in the post-NAC RPPA data. See Additional file 2: Tables s6-s7 for the full results.

While clinically significant, the investigative niche of our study presented us with a small sampling size challenge. In order to address this challenge, we first sought to validate results from the BEAUTY study with the I-SPY1 clinical trial (the only publicly available study with both pre-NAC and post-NAC gene expression data). However, the biological variation and technical variability of these two studies utilizing two high-throughput platforms (BEAUTY RNA-Seq and I-SPY1 microarray) provided a small number of concordantly differentially expressed genes (see Additional file 1: Fig. 1). In order to overcome the limitations of a small-sample-size study and get greater insight into the groupings of genes (biological pathways/gene sets) from the gene expression data, a parallel single-sample gene set analysis was carried out rather than a single gene expression analysis. Gene set analysis enabled us to minimize the false discovery rate by reducing the n >  > p by a factor of 6, while capturing the global changes rather than specific univariate differences. We conducted Gene Set Variation Analysis (GSVA) and obtained individual gene set scores for all non-LAR TNBC tumors with the pre-NAC and post-NAC RNA-Seq data. We performed three comparisons: 1) post-NAC difference between ERC vs. NRC, 2) paired pre-NAC and post-NAC differences in ERC, and 3) paired pre-NAC and post-NAC differences in NRC for both BEAUTY and I-SPY1. Among the total 2282 gene sets, we identified 251 gene sets (Additional file 2: Table s8a) that were significantly altered by comparing the ERC and NRC groups in the post-NAC tumors (p-value < 0.05).

In addition, we performed a hierarchical clustering analysis of the log2 fold changes for 251 gene sets (Fig. 3). The log fold changes represented the changes in post-NAC data and the differences between the tumors for the ERC and NRC groups. The average silhouette width, a metric of within-class similarity and between-class dissimilarity, was 0.90 for two gene set clusters and was markedly reduced for three gene set clusters (0.61). We observed that the first cluster of 61 gene sets were down regulated in the post-NAC nonrecurrent group of patients, compared to the post-NAC early recurrent group as well as the paired pre-NAC nonrecurrent biopsies (Fig. 3). Tubulin beta 4A (TUBB4A) was the most frequent gene in these gene sets, present in at least eleven (18.0%) gene sets, whereas the pre-NAC expression levels for TUBB4A were not significantly different. Conversely, the second cluster contains 190 gene sets that were up regulated in the post-NAC nonrecurrent tumors. Furthermore, this cluster was downregulated in the post-NAC ERC cohort compared to the paired pre-NAC ERC sample biopsies. The heatmap image is presented with the labeled gene clusters, and the gene sets' most commonly represented genes are shown in the word clouds in Fig. 3. In summary, Cluster 1 gene sets downregulated in the post-NAC NRC patients were predominately associated with tubulin. In Cluster2, the up-regulated gene sets included growth receptors such as EGFR, TGFBR2, and GHR, as well as tumor suppressor genes found in apoptotic gene sets. Among the 660 DE genes between ERC and NRC at post-NAC, 97 were also identified in the 251 gene sets. Although the overlap is relatively small, gene sets identified (n = 6 with p-value < 0.05) by the GESA analysis on the 660 DE genes showed consistent directionality with those in the GSVA analysis (Additional file 2: Table s8b). It is also noted GSVA analysis identified more global changes compared to the GSEA analysis based solely on the DE genes.

We investigated the I-SPY1 study to validate our pathway analysis findings [13]. Like previously described, due to the platform differences between the I-SPY1 (microarray) and BEAUTY (RNA-Seq) datasets, we calculated the gene set scores for 188/251 (74.9%) gene sets using the ISPY-1 microarray data and GSVA method. Of the 188 gene sets, 56 gene sets (29.8%) demonstrated high similarity between the I-SPY1 and BEAUTY post-NAC studies (Additional file 2: Table s9 of 56 gene sets). Among these 56 gene sets, 17 gene sets (Cluster 1, Fig. 3) were down-regulated, and 39 (Cluster 2, Fig. 3) were up-regulated in the post-NAC I-SPY1 early-recurrence tumors. Cluster 1 could not recapitulate the same set of genes as TUBB4A was not assayed on the microarray platform used for the I-SPY trial. We observed five consistent genes in Cluster 2, including ERLIN2, FCER1A, TGFBR2, PER2, and EGFR. These genes are commonly included in multiple gene sets, such as MYC targets, TGFβ, FOXO, and PI3K.

We identified 113 differentially expressed genes that were up/down regulated in the same direction between the ERC and NRC patients in the I-SPY1 and BEAUTY studies related to 56 gene sets (listed in Additional file 2: Table s9). We assessed the predictive value of these 113 genes in a publicly available independent TNBC cohort (n = 392) using the Kaplan–Meier (KM) plotter online breast cancer database [38]. Our analysis shows that 17 genes were significantly associated (log-rank test p-value < 0.05) with relapse-free survival (RFS) in the TNBC cohort (Fig. 4, Additional file 1: Figure s2). Further investigation into this sub-cohort with positive lymph node disease (n = 189) showed that 10 of the 17 genes remained significant (Table 2). Additionally, we evaluated the proportional hazards for each gene in the BEAUTY and I-SPY1 trials. The hazard ratios between the two studies showed a significant correlation (Pearson correlation coefficient = 0.78) between the two studies (Additional file 2: Table s10).

Since there are no independent neoadjuvant breast cancer studies with publicly available data to validate our signature, transcriptomic data from the BEAUTY and I-SPY1 studies were pooled together to evaluate the 17-gene signature using machine learning models. We examined six classification algorithms, including a generalized linear model (glmnet), k nearest neighbors (k-NN), neural networks, random forests, support vector machine (SVM), and kernel SVM. The models were implemented as wrappers in the R package SuperLearner using gene down selection methods (Pearson's correlation, Spearman's rank correlation, and random forest). Given the lack of sampling power, we implemented a threefold cross-validation strategy, which we believe provided the most reasonable cross-validation strategy without subsampling the data into too trivial of a sample representation. We monitored the area under the receiver operating curve as an indicator of the model's classification performance, as shown in Fig. 5. The rank-based correlation (deep brown diamond) consistently underperformed other feature selection approaches (Additional file 1: Figure s3).

Most importantly, the ensemble of seventeen genes (labeled as 'All' in Fig. 5, tan diamond) was consistently the best-performing approach. We present the distribution of AUCs observed for each classification wrapper in Fig. 5. The models are ordered by increasing median AUC values, and kernel support vector machines, k nearest-neighbor (kNN), and random forest were the best performing modeling approaches, with the ensemble of all 17 genes achieving AUCs of 0.84, 0.88, and 0.88, respectively. A similar assessment of a model based on the available clinical features (as shown in Table 1) was also performed to compare with the transcriptome data-based 17-gene signature model. Due to the limited clinical data available for the I-SPY1 cohort, this comparison was limited to the BEAUTY TNBC cohort. The best performance of the clinical feature-based models was AUC = 0.69, and clinical feature selection failed to improve the performance of the clinical feature-based models. Overall, the 17-gene signature provided classification model that outperformed all clinical feature models.

We investigated several datasets to confirm that the dysregulation of our signature's 17 genes was 1) specific to non-LAR tumors and 2) specific to post-NAC non-LAR tumor biopsies (Table 3). We evaluated the expression among eleven paired TNBC primary tumors (in LARs and non-LARs) and normal adjacent samples from TCGA and observed dysregulation among the tumor samples, with 9 out of the 17 (52.9%) surveyed being differentially expressed and 7 of 9 genes confirmed to be DE when the analysis was refined to the nine paired non-LAR TCGA tumors (Additional file 1: Figures s4-s5). Moreover, among these seven genes, we confirmed dysregulation in a core set of six genes (COL4A6, AGGF1, TANK, CITED2, PDE2A, and MCM3) in an independent cohort of n = 444 TNBC samples (Table 3, Additional file 1: Figure s6). These publicly available microarray datasets collated from the Thompson et al. 2022 study were primarily assayed with the Affymetrix U33 platform [17].

Furthermore, we evaluated whether dysregulated expression of these 17 genes existed prior to NAC treatment. As shown in Fig. 6, we verified that the differential expression of these 17 genes was specific to post-NAC tumors. In contrast, none of the genes were observed to be differentially expressed among the pre-NAC tumor biopsies. We subsequently investigated the 288 (of 388) non-LAR primary tumor TNBC samples from the independent cohort (124 pCR, 164 non-pCR, 100 missing). We observed that only 2 (12.5%) of the 16 genes (RSPO3 was not assayed on the microarray platform) were differentially expressed (Additional file 1: Figure s7). In summary, the dysregulation of the 17 genes identified in this study was associated with post-NAC recurrence only in non-LAR TNBCs with residual disease after NAC (Additional file 1: Figure s8).