br We compared the gene expression
We compared the gene DETA NONOate detected in our cohort with data in the Human Protein Atlas [38,39] comprising mRNA-seq data of 1063 female breast cancer patients from TCGA. The 17-marker panel consists of 16 markers that have an adverse eﬀect on survival when over-expressed. Seven of these genes also showed statistically significant adverse OS outcome in the Human Protein Atlas dataset (BUB1B, DCAF13, DIAPH3, MARS, MTERF3, SPDL1, ZDHHC23). The nine re-maining genes showed no statistical significance when stratified in the Human Protein Atlas dataset. Overexpression of PIEZO2 is connected to unfavourable survival outcome in renal cancer but showed no statistical significance in the breast cancer dataset of the Human Protein Atlas. A possible explanation for these discrepancies can be technical diﬀerences (mRNA-seq and gene expression microarray) and diﬀerences in popu-lation cohorts (United States and Sweden). In the microarray-based KM-plotter datasets (n = 3955) , overexpression of ten genes was sig-nificantly associated with adverse recurrence-free survival (RFS) out-come (BUB1B, CDCA3, DEPDC1B, EED, EPCAM, KIF14, MARS, MTERF3,
NUP153, SPDL1), while overexpression of five genes was significantly associated with favourable RFS outcome (DIAPH3, FAM131B, PIEZO2, TAF5L, ZDHHC23). However, univariable stratification of cohorts has only limited impact in this setting as we aimed to identify networks of genes connected to genomic instability. Furthermore, the reciprocal eﬀects of overexpression of some genes highlight that the same path-ways can be deregulated in diﬀerent ways that go beyond the expres-sion of single genes.
The 17-marker panel comprises genes selected based on their vari-able importance in the segregation between genomically stable and unstable tumours. The stability of the genome classified by the G2I could be translated into a 17-marker gene expression signature. Hence, the 17-marker panel has the potential to be easily implemented into clinic routine, e.g. by using quantitative real-time reverse-transcriptase polymerase chain reaction (qRT-PCR). Despite the relatively low nu-merical diﬀerences in HRs between the 17-marker panel and competing signatures, the 17-marker panel could prove a high and stable pre-dictive accuracy through the entire follow-up period that exceeded the competitor's accuracy. The universal applicability of the 17-marker panel was validated in three external validation cohorts. The novelty and clinical significance of the 17-marker panel was demonstrated by the increased predictive power compared with the 12-gene signature and the Oncotype Dx-based signature.
Combining the 17-marker panel with the molecular subtype re-sulted in slightly increased C-indices but less informative HRs and thus had a marginal eﬀect on stratification of patient outcome. However, a strong increase in HRs was detected when the 17-marker panel was applied to only the Luminal B-subcohort (n = 93) indicating that the 17-marker panel shows increased applicability for Luminal B tumours. This eﬀect can be partly explained by the cohort selection, which consisted to a large part of Luminal B tumours posing a selection bias and a major limitation of the study. A further limitation is that the comparison with the 12-gene signature only included 11/12 genes due to the use of diﬀerent microarray platforms. Furthermore, the Oncotype Dx-based signature applied array-based gene expression data instead of qRT-PCR data. As the gene expression microarray data was normalized with normal breast tissue, we did not include the five reference genes for Oncotype Dx and instead built a multivariable model on the 16 cancer-related genes. Additionally, in the validation cohorts dis-crepancies between (i) the probes of the two microarray platforms (Illumina Human HT-12 Whole-Genome Expression BeadChip and Aﬀymetrix Human Genome U133 Set) and (ii) the diﬀerent experi-mental approaches (gene expression microarray and mRNA-seq) need to be taken into consideration. The microarray-based cohorts GSE1456 and GSE4922 consist of samples from the Swedish population, while the mRNA-seq-based TCGA cohort presents a diﬀerent type of experiment and comprises samples primarily from the United States. These tech-nical and diﬀerences in population cohorts could account for the lower HR in the TCGA cohort. Nonetheless, true biological eﬀects should be detectable using diverse validation datasets. Another limitation is that the cut-oﬀ between genomically stable (G2I-1, G2I-2) and unstable tumours (G2I-3) as suggested by the G2I algorithm might need to be optimised as some tumours of the intermediately stable group G2I-2 appear among the clusters of genomically unstable tumours. However, discretisation always presents a loss of information; therefore, it might be more sensible to define genome stability based on a continuous variable.