Clinically actionable alterations in Indian breast cancer patients derived through whole transcriptome sequencing

Study design, sample biobanking, and next generation sequencing

Patient recruitment, sample biobanking, RNA extraction, and whole transcriptome sequencing (RNA-Seq) were carried out at the Clinician Scientist laboratory, ACTREC, as previously described.^5,10,11

RNA-Seq was conducted to generate 30-60M paired-end reads per sample, as described earlier.^5,10,11 RNA-Seq data were re-analysed to identify genomic (DNA) alterations and classify the tumours using gene expression-based classifiers. This RNA-Seq data was mined for genomic alterations using a bioinformatics pipeline specific for variant calling from transcriptomic data. When RNA-Seq data from multiple samples from the same patient tumour were available, the analysis was independently performed in all samples available for that patient, treating each sample as a replicate. Accordingly, the analysis used triplicate samples for some patients, duplicate samples for others, and single samples for the remaining patients.

Bioinformatics analysis

Data availability statement

The human sequence data generated in this analysis are not publicly available due to patient privacy requirements but are available upon reasonable request to the corresponding author. Other data generated in this study are available within the article and its supplementary data files.

Patient population

Across the pooled studies, 97 patients contributed tumour samples with whole-transcriptome (RNA-seq) data and were therefore included. RNA-seqdata were available from three samples (in triplicate) for 34 patients, from two samples (in duplicate) for 42 patients, and as a single sample for 21 patients, yielding 207 tumour RNA-seq datasets for analysis. Patients were stratified based on oestrogen receptor (ER), progesterone receptor (PR), and HER2 status (Table I).

Molecular subtyping by gene expression analysis reveals canonical intrinsic breast cancer subtypes in Indian patients

We identified a median of 11,650 (range 100-33,100) transcripts across 207 samples from 97 patients. Transcript distribution across paired samples revealed a median of 11,300 (range 100–28,600) transcripts in at least three samples, 11,100 (range 100-33,100) transcripts in at least two samples, and 11,900 (range 1,100-21,400) transcripts in at least 1 sample. Transcript distribution and overlap have been shown in Supplementary Figure 1. The 97 patients were classified according to both AIMS and PAM50 intrinsic subtyping classifiers. AIMS identified five subtypes (Luminal A, n=27; Basal-like, n=25; HER2-enriched, n=23; Luminal B, n=20; Normal-like, n=2; Supplementary Table I), whereas PAM50 resulted in four subtypes (Luminal B, n=36; Basal-like, n=29; Luminal A, n=19; HER2-enriched, n=13; Fig. 1, Table II). A detailed comparison of these subtype distributions and their concordance is presented in Supplementary Table II. This comparative analysis (Supplementary Table II) reveals significant discordance in the classification of Luminal A and Luminal B subtypes between the two methodologies. In contrast, Basal-like and HER2-enriched tumours demonstrated high concordance across both classifiers. Patients classified as Luminal A by the PAM50 gene signature were found to be distributed among several distinct AIMS subtypes.

Close

Fig. 1.

PAM50 clustering and gene expression for 97 patients. The heatmap was generated using R software.

Export to PPT

Molecular architecture

Following stringent filtering, which mandated a minimum of five sequence reads for mutant allele support and excluded synonymous nucleotide changes, a total of 1,803 unique genetic variants were identified across the patient cohort. On a per-patient basis, a median of 302 such variants were detected. Of these 1,803 unique variants, 1,322 were absent from the dbSNP, COSMIC, and ICGC databases. These variants, potentially novel and specific to the Indian ethnic background (as they were not catalogued in the Indigenomes or TMC-SNPdb2.0 databases), underscore the genetic diversity within this population; they were subsequently excluded from further analysis.

The remaining 481 variants showed varied database representation: 179 were listed in dbSNP, COSMIC, and ICGC; 190 in dbSNP and COSMIC; 23 in dbSNP and ICGC; 40 in COSMIC and ICGC; 47 exclusively in COSMIC; and 2 solely in ICGC. A subsequent clinical significance assessment using ClinVar led to the exclusion of an additional 336 variants, comprising 239 with no ClinVar annotations and 97 classified as benign or likely benign.

From the 145 variants that proceeded past these filters, 60 were categorised as variants of unknown significance (VUS), 48 exhibited conflicting interpretations of pathogenicity, and 37 were determined to be pathogenic or likely pathogenic. Ultimately, 85 variants—those with conflicting significance or classified as pathogenic/likely pathogenic were selected for downstream analysis. As anticipated, mutations in TP53 (observed in 47% of the cohort) and PIK3CA (34%) were the most prevalent genetic alterations identified (Fig. 2, Supplementary Table III).

Close

Fig. 2.

Heatmap of genes mutated in at least 3% in cohort.

Export to PPT

Actionable alterations

The 481 variants identified in at least one of the dbSNP, COSMIC, or ICGC databases underwent further assessment for clinical actionability. This analysis was performed using the ESMO Scale for Clinical Actionability of molecular Targets (ESCAT) framework, facilitated by a local academic installation of the Cancer Genome Interpreter (CGI).^32,33 This evaluation revealed that a significant portion of the cohort, 51 out of 97 patients (52%), possessed at least one actionable mutation. The evidence supporting these actionable findings was categorised as follows: 20 mutations were supported by Level A evidence, 10 by Level B evidence, and 15 by Level C evidence. Among the 51 patients with actionable mutations, alterations in PIK3CA and TP53 were prominent. Specifically, 24 of these 51 patients had a PIK3CA mutation, and 17 had a TP53 mutation. Furthermore, 9 patients within this group harboured concurrent mutations in both PIK3CA and TP53. These identified mutations were predominantly well-characterised, previously documented hotspot mutations (Supplementary Fig. 2).

Validation of mutations

An attempt was made to validate the mutations initially identified through RNA-sequencing using DNA extracted from formalin-fixed paraffin-embedded (FFPE) tumour tissues. DNA was successfully extracted from 10 available tissue samples. However, the inherent degradation caused by the paraffinization process resulted in poor-quality DNA for most samples, rendering them unsuitable for subsequent assays. Of the evaluated samples, only two yielded DNA of acceptable quality, defined by a DIN greater than 2.5 and a concentration of at least 100 ng/µl. From these two samples, one mutation (specifically, PIK3CA H1047R in sample 14T) was successfully validated. The attempted validation of a second mutation (PIK3CA E545K in sample 69T) was unsuccessful. It is likely that this could be due to the limit of detection of the qPCR assay used (1% for this particular mutation). The tumour allele fraction in the FFPE sample might have been reduced due to the presence of normal tissue or fat infiltration, falling below the qPCR detection threshold, though it was identifiable by the more sensitive high-depth sequencing.

Network analysis in each intrinsic subtype

Our analysis identified 227 unique genes harbouring mutations that met our defined criteria. Twenty-one (9.3% of the 227) genes were found to be altered across all molecular subtypes. Further analysis revealed distinct sets of uniquely mutated genes in each subtype. Specifically, 30 genes (13.2%) were mutated exclusively in the Luminal A subtype, 32 (14.1%) were unique to Luminal B, 28 (12.3%) were specific to the HER2-enriched subtype, and 33 (14.5%) were uniquely mutated in the Basal-like subtype (Supplementary Fig. 3). Pathway analyses were performed on these subtype-specific gene pools. This revealed that pathways related to immune response, DDD (Disease, Drug, and Development), haematopoiesis, and growth/development were significantly impaired in both Luminal A (Supplementary Fig. 4A) and Luminal B subtypes (Supplementary Fig. 4B). In contrast, the HER2-enriched subtype showed enrichment in pathways associated with Epidermal Growth Factor Receptor Family (ERBB) signalling, PI3K signalling, and T-cell regulation (Supplementary Fig. 4C). The Basal-like subtype was characterised by an enrichment of pathways involved in cell cycle regulation, including those governing microtubule and cytoplasmic regulation, nuclear excision repair, and developmental growth processes (Supplementary Fig. 4D).

Fusion transcript analysis

Across 207 patient samples (derived from 97 patients), a total of 225 potential fusion events were initially identified. Predicting fusion transcripts from RNA-sequencing data is challenging due to misalignment errors, library preparation artefacts, and the absence of germline control data. To address these limitations, the analysis focused only on fusion transcripts detected in multiple tumour samples from the same patient, as true fusions are more likely to recur across independent experiments.

Applying this stringent criterion, 91 of the initial 225 fusion events were classified as recurrent, i.e., present in more than one sample from the patient. Notably, this filtered set included novel fusion transcripts involving partner genes previously implicated in fusions, such as ERBB2, CBX1, MED1, CDK12, ELF2, and KANSL1. Despite these findings, the analysis suggests that commonly known fusion transcripts reported in other breast cancer populations do not appear to be prevalent in this Indian breast cancer cohort.