Genetic ancestry prediction for breast cancer samples from RNA-seq data

We analyzed RNA-seq data from tumors of 274 breast cancer patients and six healthy tissue controls, retrieved from 36 publicly available studies. Most of these studies were conducted in the United States and Spain (S1 File). Reads from all samples included in this study were aligned to the reference genome with a mapping rate exceeding 70%. To assess whether the genetic ancestry of each sample could be evaluated from RNA-seq data alone, we genotyped known variants in coding regions from the 1,000 Genomes Project, which included 1,600 reference female individuals in their project. This resulted in 9,578 single nucleotide variants (SNVs) genotyped in a total of 1,808 individuals with a missing data rate of 1.85%, attributed to RNA-seq differences in expression. The number of genotyped SNVs ranged between 4524 (47.23%) to 9578 (100%), with an average percentage of 88.18% (S1 File). Only 44 samples had a percentage of genotyped sites below 75%.

To infer the genetic population of origin of each sample, we performed a maximum likelihood estimation of individual ancestries using the previously called SNVs varying the number of clusters (k parameter) between 1 and 10, using the genomic variation database constructed from RNA-seq data as input. Ancestry was successfully predicted for samples from the 1,000 Genomes Project (Fig 1A). In particular, the African population was split in k = 2 given its higher diversity compared to other populations. Asian and European groups split in k = 4, and the probably Native American component (AMR, orange) appears in Admixed Americans in k = 5, separating them from the EAS population. We chose k = 5 to determine the ancestry of the breast cancer samples, considering the previous annotations on human populations. Most of the samples (n = 204) were classified as European ancestry (EUR), followed by East Asian (EAS) (n = 40), African (AFR) (n = 24), South Asian (SAS) (n = 7), and Admixed American (AMX) (n = 5) (Fig 1A), which is consistent with the geographic origin of the samples. Individuals from the USA, Israel, Canada, Germany, Spain, the UK, and the Netherlands showed predominantly European ancestry, whereas individuals of Asian origin, including samples from China and South Korea, were classified as EAS. Individuals from Singapore and India clustered predominantly within the SAS population (Fig 1B). Genetic admixture was predicted for individuals from New York, individuals of African origin living in Birmingham, and individuals of Mexican origin (Fig 1C). RNA-seq data for five admixed samples (HG01113 and HG01125 from Colombia, NA19648, NA19654, NA19657 from Los Angeles) was reanalyzed to validate the consistency of ancestry estimates obtained from DNA and RNA-seq data (S1 Fig in S6 File).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Genetic ancestry.

A. ADMIXTURE clustering of individuals from the 1,000 genomes project and breast cancer RNA-seq data. African (AFR), Admixed American (AMX), East Asian (EAS), European (EUR), and South Asian (SAS). B. Ancestry frequency for 280 breast tissue samples by origin. C. Ancestry prediction for the admixed individuals, each column represents an individual.

https://doi.org/10.1371/journal.pone.0352514.g001

Besides the five individuals classified as AMX, we observed 24 individuals with admixture patterns. Most of these individuals were originally classified as European. However, they had a membership probability lower than 0.6 to a single population so we re-classified them as: European–African (25%, 6/24), followed by European–South Asian, and European–Admixed Americans with an equal proportion of 21% each (5/24) and European–East Asian (4%, 1/24). Additionally, we found African–Europeans 21% (5/24), South Asian – European (4%, 1/24), Admixed American–African (4%, 1/24).

Uncertainty in breast cancer subtyping based on the PAM50 panel

Given that the selected samples were collected from 34 different studies, which used different protocols for RNA extraction and library preparation, it was likely that read counts were affected by different patterns of batch effects. A principal component analysis of a subset of 204 samples belonging to the 12 projects with the largest number of samples corroborated this situation (S2 Fig in S6 File). Taking into account that batch correction is more effective if the groups are well represented, and that subtype prediction based on GeneFu also required more than one sample, we restricted the analysis of expression data to these 204 samples. To assess the stability of GeneFu predictions using the PAM50 model, we performed the prediction of each sample with two different inputs: 1) corrected counts for the 204 samples; and 2) uncorrected counts for the samples within each project.

Fig 2A shows the distribution of subtype assignment probabilities for each sample if corrected counts for the complete set of samples are taken as input. For 73 samples the maximum probability was lower than 0.6 and for 21 of these samples the probability was lower than 0.5. Consequently, 24 samples (11.76%) had discordant subtyping between the two prediction methods. From the samples with consistent assignments between methods, 66 were classified as Basal, 48 as Luminal A, 26 as Luminal B, 20 as HER2+ and the remaining 20 as Normal like. Within these samples, 10 classified as Luminal A (20.8%) had a probability larger than 0.4 of belonging to another group. A similar situation was observed for 5 samples subtype as Luminal B and 5 samples subtyped as HER2 + .

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Subtype predictions.

A. Subtype frequency obtained with the PAM50 algorithm for 204 RNA-seq breast cancer tumor samples. (LumA = Luminal A, LumB = Luminal B). B. Percentage of 180 samples with consistent subtype predictions discriminated by genetic population. C. Percentage of 140 samples of European origin with consistent subtype predictions discriminated by TNBC status. Numbers in parenthesis correspond to the total number of samples.

https://doi.org/10.1371/journal.pone.0352514.g002

We investigated the relationship between the subtypes predicted using the PAM50 algorithm to the genetic populations inferred by ADMIXTURE (Fig 2B). Consistent with the distribution of the complete dataset, 140 (77.8%) of the 180 samples with consistent subtype predictions had European ancestry (EUR), while only 5 samples were admixed american (AMX). The group was completed by a nearly even distribution between African (AFR) and east asian (EAS) ancestry. The most noticeable difference in subtyping among ancestry groups was observed for the Basal subtype. While only 30% of the samples with consistent subtyping and European ancestry were subtyped as Basal, the same percentage was higher (>55%) for other population groups, although the difference was not considered significant after correction for multiple testing (p-value > 0.01 for a one sided Fisher exact test). Regarding the correlation between the intrinsic subtype predicted by PAM50 and TNBC status, although more than half of the samples did not have TNBC as reported metadata, we could assess that 47.22% of the 36 individuals cataloged as TNBC had Basal as predicted subtype. This percentage is significantly higher than the 14.29% of the 42 non-TNBC samples having Basal subtype (p-value = 3.50 × 10⁻⁴ for a one sided Fisher exact test). The difference between percentages increases if only the 140 samples with consistent subtypes and European ancestry are taken into account (Fig 2C, 50% vs 16.22%, p-value = 7.05 × 10⁻³).

Clustering of breast cancer tumors based on complete expression profiles

Considering the potential uncertainty of predictions obtained with the PAM50 algorithm, we investigated analysis alternatives that could lead to a more clear differentiation among subtypes from complete expression profiles. First, we selected a group of 53 samples from the four projects with the largest numbers of samples, having European origin and having consistent subtype predictions (see methods for details). A set of 475 differentially expressed genes (DEGs) were selected performing the 10 possible pairwise comparisons between each pair of subtypes and taking the union of DEGs obtained in each comparison (S2 File). A principal component analysis (PCA) of the corrected counts within these genes visually places samples from different subtypes in consistent regions of the space (Fig 3A). The first principal component explains 37% of the variation and differentiates basal samples from most non-basal samples. The second component explains 13% of the variance and seems to differentiate Normal-like samples from the Luminal subtypes. The third component explains an additional 8% of the variance and differentiates HER2 + samples from most samples with other subtype assignments. Functional enrichment of the selected genes highlighted biological processes related to chromosome segregation, regulation of cell cycle, and programmed cell death (S3 File and S3 Fig in S6 File).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Unsupervised clustering.

A. PCA plot of the normalized counts for 476 DEGs selected from 53 RNA-seq samples with consistent subtype assignments. B. Jaccard similarity between clusters obtained by k-means and by hierarchical clustering from different sets of genes expressed in 204 samples. C. Weighted jaccard similarity between clusters obtained from different sets of genes expressed in 204 samples and clusters inferred from predicted subtypes. The jaccard similarity is weighted by the probability of subtype assignment. Continuous lines correspond to k-means and dotted lines correspond to hierarchical clustering.

https://doi.org/10.1371/journal.pone.0352514.g003

Based on this initial result, two unsupervised clustering methods (k-means and hierarchical) were executed on the counts obtained for the 475 DEGs on the larger set of 204 samples belonging to the 12 studies with at least 9 samples per study. The number of clusters (k parameter) varied from 1 to 10. A weighted jaccard distance was used to evaluate clusters obtained in each experiment, comparing the consistency between methods for the same k value and calculating the agreement with the clustering induced by the predicted subtypes (see methods for details). The clusters obtained from the 475 DEGs were also compared with clusters obtained from the 500 genes with highest variance and the clusters obtained using only the PAM50 genes.

The clusters obtained from the 475 DEGs were more consistent between clustering algorithms across k values, compared to clusters obtained with other sets of genes (Fig 3B). They were also more consistent with the predicted subtypes, compared to clusters obtained from the PAM50 genes (Fig 3C). The best overall consistency (0.43) was achieved running the k-means method on the 475 DEGs with k = 4. This was closely followed by the same method and the same genes with k = 5. Considering that the subtype prediction includes 5 possible subtypes, we selected the latter clustering for further downstream analysis.

Immune and stromal cell composition and in silico estimation of drug sensitivities in clustered transcriptomic profiles

To explore potential differences in immune cell infiltration across the three identified clusters, we employed the xCell deconvolution method [52], which estimates the frequencies of 67 unique immune and stromal cell types from bulk transcriptomic profiles. From these, 40 subtypes showed significant differences between groups (adjusted p-value <0.01 for a Kruskal Wallis test, S4 File). The cell types with lowest p-values were adipocytes and Th-cells (Fig 4A). Adipocytes were more represented in the clusters 2 and 4, which have most samples subtyped as Normal-like and Luminal A respectively. Conversely, Th1 and Th2 cells were more represented in clusters 3 and 5, which have most samples subtype as Basal and Luminal B respectively. Biologically relevant differences across clusters were observed in Th2 cells, M2 macrophages, and endothelial cells. Th2 cells showed enrichment in cluster 3, suggesting a skewing toward a type 2 immune response in this group. Similarly, M2 macrophages displayed significant variation across clusters, with higher scores in clusters 1 and 5, while cluster 3 consistently showed lower infiltration levels. This pattern may reflect differences in immunosuppressive microenvironmental features among clusters. Endothelial cells were also differentially distributed, with an enrichment in cluster 2 compared to other clusters, indicating potential differences in angiogenic activity across transcriptomic groups. These differences were consistently observed in an analysis of the groups directly inferred from subtype assignments, but with larger p-values overall and lack of significance for the case of macrophages and endothelial cells (S4 Fig in S6 File).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Downstream analysis of clusters of breast cancer samples.

(A) Differences in cell subtypes among clusters based on xCell deconvolution analysis. The xCell scores represent the estimated frequencies of cell types within samples. (B) Drug sensitivity predictions, measured as IC50 values, for six FDA-approved drugs across transcriptomic clusters using the pRRophetic R package. Paired comparisons between clusters were performed using the Dunn test, with p-values < 0.05 considered significant. Error bars represent the median and interquartile range.

https://doi.org/10.1371/journal.pone.0352514.g004

Breast cancer is a highly heterogeneous disease with varying responses to standardized treatments across molecular subtypes. While chemotherapy is effective for many patients, challenges persist in cases of refractory disease and particularly in triple-negative breast cancer (TNBC), where therapeutic options remain limited, and there is a critical need for novel treatments. In this study, we explored potential drug sensitivities through an in silico prediction of response to 137 drugs based on the transcriptomic profiles from individual patient clusters. Differential IC50 values were predicted for 113 drugs (adjusted p-value <0.01 for a Kruskal Wallis test, S5 File). Particularly, cluster 3 (mostly composed of samples with basal subtyping) had higher sensitivity for Imatinib and Sorafenib, and lower sensitivity for Etoposide, Midostaurin, Mitomycin C, and Sunitinib, drugs that are already approved to treat other cancers (Fig 4B). The analysis of groups inferred from the predicted subtypes was also consistent in this case, but also with higher p-values and lack of significance of some pairwise comparisons, especially for Etopocide and Mitomycin C (S5 Fig in S6 File).

Data collection and initial processing

The Human reference genome hg38 was downloaded from the Broad Institute public database (https://console.cloud.google.com/storage/browser/gcp-public-data--broad-references;tab=objects?prefix=&forceOnObjectsSortingFiltering=false). The complete dataset analyzed in this study consists of publicly available RNA-seq data obtained from breast cancer samples of anonymized patients. The data was recovered from the Sequence Read Archive (SRA) database, using the search term “(breast cancer) AND “Homo sapiens”[orgn:txid9606] NOT cell”, then adding as a filter “fastq” in the File type section, “RNA” in Source section and “Public” in Access section. The results were filtered by a minimum number of 10 million reads. We selected a total of 274 samples based on percentage of mapped reads and the geographical origin of the study. We also included 6 samples of healthy tissue that were used as controls, for a dataset of 280 samples in total. RNA-seq reads were mapped to the human reference genome using HISAT2 (v.2.2.1) [69] obtaining aligned reads in SAM format. SAM files were sorted using Picard (v2.27.4). To assess batch effects, we analyzed the distribution of samples from the same study within the PCA space.

Ancestry estimation

Regions covered by RNA-seq reads were calculated from BAM files using samtools depth (v.1.16.1) [70] and were joined to generate the set of common regions between all the samples. Genomic variation data from the human reference populations was retrieved from the 1,000 Genomes Project [71] (http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/1000G_2504_high_coverage/working/20201028_3202_raw_GT_with_annot). The VCF file was filtered retaining the variants covered by RNA-seq common regions, only women individuals, a minimum MAF value of 0.01, and a minimum number of genotyped samples of 1,850, using the VCFFilter functionality of the NGSEP software [72]. The resulting VCF was merged with that obtained from the breast cancer samples into a single VCF.

The merged VCF file was converted to plink format using vcftools (v.0.1.16) [73], ADMIXTURE input was generated using plink (v1.9) [74]. ADMIXTURE (v.1.3.0) [75] analysis was performed for k values between 1 and 10 with a coefficient of variation (CV) value of 20. We used k = 5 to determine the population based on the major groups sampled in the 1,000 Genomes Project, allowing the model to predict the native american component in latin american admixed individuals. Individuals having the largest membership probability to a single population lower than 0.7 were classified as admixed.

Extraction of read counts from RNA-seq data and correction of batch effects

Gene counts per gene per sample were calculated with Stringtie (v.1.3.5) [76] from the aligned sequences. A matrix of raw counts was obtained from individual sample files running the script prepDe.py available with the Stringtie distribution. Data for 204 samples belonging to the twelve projects having at least nine samples was selected for downstream analysis. A principal component analysis (PCA) of the uncorrected counts, normalized using the VST method of the DESeq2 R package [77], revealed biases related to the study of origin (S2 Fig in S6 File). Consequently, the raw counts were corrected running CombatSeq [78].

Molecular subtyping prediction and differential expression analysis

Molecular subtypes were predicted from z-scores calculated from gene count matrices normalized using the VST method. The R package geneFu (v.2.28.0) was used to run the PAM50 algorithm [35,79]. Two separate subtype predictions were performed for each sample, the first using corrected counts from the set of 204 samples, and the second using uncorrected counts from the project of origin of the sample. Moreover five additional predictions were performed including random subsets of 25% of the samples within the project of origin of the sample, ensuring that at least four samples were included in each experiment. The number of times that the predicted subtype agreed with the subtype predicted from the dataset of 204 samples were recorded (S1 File).

The different alternatives of subtype predictions were combined to select a reduced set of samples for differential expression analysis between predicted subtypes. An initial set of 38 of samples was selected with the following criteria: a) Belong to one of the four projects with largest number of samples; b) European ancestry; c) Consistency between project specific subtype and the subtype predicted using the 204 samples; d) Complete agreement of predictions in the subset experiments and assignment probability larger than 0.6 or 80% agreement of predictions in the subset experiments and assignment probability larger than 0.8. Because this initial dataset had no representation of Her2 and Normal-like subtypes and only four samples of Luminal B, a second set of 15 samples was selected for these subtypes relaxing condition d) to allow samples with assignment probability larger than 0.5 and at least 60% of agreement in the experiments with subsets. This increased the representation of the underrepresented subtypes to six, five and eight samples respectively.

Differential expression analysis was performed between samples belonging to each possible pair of subtypes to identify differentially expressed genes (DEGs). The analysis was performed by running the DESeq2 package, which estimates log2 fold changes and calculates adjusted p-values (Benjamini-Hochberg correction) to control for false discovery rates (FDR). Different thresholds of log fold change were selected for each comparison to reduce the overrepresentation of genes differentiating specific pairs of subtypes (S1 Table in S6 File). Genes with absolute log fold change significantly higher than the threshold (adjusted p-value < 0.05) in at least one comparison were added to the set of genes for unsupervised analysis. Functional enrichment analysis of the 475 DEGs was performed using the gProfiler software [80] available at https://biit.cs.ut.ee/gprofiler/gost.

Unsupervised machine learning of gene counts

Z-scores calculated from corrected counts normalized using the VST method of the DESeq2 were used as input for unsupervised clustering. Different experiments were performed varying the input genes (500 with highest variance, 475 selected DEGs and PAM50 genes) and clustering methods. For each set of genes, a PCA of the z-scores was obtained by running the prcomp function of R. Hierarchical clustering and k-means were executed from the PCA latent space. In both cases the number of clusters varied from 1 to 10. The clusters were compared to each other and to the clustering inferred by the predicted subtypes calculating the jaccard similarity between the cluster assignments. For comparisons against the subtypes a weighted version was implemented in which the weight of each pair of samples is calculated as the average of subtype assignment probabilities. Visualization of PCAs was performed using the ggplot2 package [81].

In Silico drug sensitivity analysis

Drug sensitivity analysis was performed using the R package pRRophetic [82], which predicts the half-maximal inhibitory concentration (IC50) of chemotherapeutic agents based on RNA-seq normalized expression data. Sensitivity differences between the K-means clusters were assessed using the Kruskal-Wallis test, followed by the pairwise Dunn test, All p-values were corrected for multiple testing using the Benjamini-Hochberg (BH) method and we established p > 0.05 as the significance threshold.

Immune and stromal cell infiltration analysis

Immune and stromal cell infiltration was estimated using the xCell algorithm implemented in R (https://github.com/dviraran/xCell). xCell infers the relative enrichment of 64 immune and stromal cell types based on curated gene expression signatures. Differences in cell-type enrichment across clusters were first assessed using the Kruskal–Wallis test. For cell types showing significant global differences, pairwise comparisons between clusters were performed using Dunn’s test, A significance threshold of adjusted p < 0.05 was applied.