The authors evaluated whether the 11 Integrative Cluster (IntClust) genomic subtypes of breast cancer, originally defined using integrated copy-number and gene-expression profiling, can be accurately inferred from routine clinical DNA targeted gene-panel sequencing and whether these inferred subtypes meaningfully stratify overall survival (OS).
A machine-learning classifier was developed and trained on public METABRIC data (N=1,980), externally validated using TCGA-BRCA (N=1,066), and then applied to a contemporary cohort of clinical targeted-panel sequencing results from 5,368 patients in the MSK-CHORD dataset derived from the MSK-IMPACT programme. The analysis focused on OS and tested whether inferred IntClust assignments resolve prognostic heterogeneity not captured by standard clinical markers such as ER or HER2.
The classifier was trained on METABRIC and externally validated on TCGA-BRCA before deployment on the MSK-CHORD clinical cohort. The MSK-CHORD cohort used for application comprised 5,368 patients with targeted sequencing data from the MSK-IMPACT clinical sequencing programme. The authors provide per-patient predicted IntClust labels, the feature matrix used for prediction, and processed matrices as supplementary files and archived data.
All primary data analysed in the study are publicly available via cBioPortal under the identifiers cited for MSK-CHORD, METABRIC, and TCGA-BRCA. The MSK-CHORD predicted labels and input features are archived at Zenodo as reported by the authors.
Inferred IntClust labels significantly stratified overall survival in both localized and metastatic breast cancer cohorts (log-rank P<0.0001). The stratification was robust across disease settings and showed clinically meaningful separation of outcomes that was not captured solely by ER or HER2 status.
Within ER-positive metastatic cases (N=2,689), median OS varied substantially across IntClust groups, with reported median survival ranging from 46 months for IC10 to 116 months for IC3. The authors used a pre-specified categorization that grouped ER+ subtypes into a worse-prognosis set (IC1, IC2, IC6, IC9) and a better-prognosis set (IC3, IC4ER+, IC7, IC8). This pre-specified separation was highly significant (P<0.0001) and showed the same pattern in localized disease.
In metastatic triple-negative breast cancer, the IntClust groups IC10 and IC4ER- exhibited marked survival separation (approximately 28 vs 47 months; HR 1.58, P<0.0001).
Within HER2-positive metastatic disease, IC5 trended toward longer OS (HR 0.69, P=0.11). In triple-positive disease comparisons, IC5 demonstrated improved survival compared with IC4ER+ (HR 0.59, P=0.027). These findings indicate that IntClust assignments can refine prognostic expectations within established clinical subtypes.
The inferred IntClust framework identified subgroups with distinct prognoses within standard receptor-defined categories. For ER-positive metastatic disease, the marked spread in median OS across IntClust groups highlights substantial heterogeneity within ER+ tumors that is resolved by genomic subtyping inferred from targeted DNA data.
For HER2-positive and triple-negative disease, particular IntClust groups (for example, IC5 in HER2+ and IC10/IC4ER- in triple-negative) showed differential outcomes, suggesting that IntClust captures biology relevant to survival beyond receptor-defined treatment groups.
The study reports that ESR1 mutations were strongly enriched in metastatic biopsies (odds ratio 6.73, FDR<0.0001) with heterogeneous enrichment across IntClust groups (P=0.0017). The strongest enrichment of ESR1 mutations was observed in ER-positive subtypes IC3 and IC4ER+.
The authors performed a systematic test of gene-by-IntClust survival associations. Of 134 testable gene-by-IntClust combinations, 26 reached an FDR threshold <0.10. TP53 mutation was associated with shortened survival across most IntClust groups in metastatic disease (reported HR range 1.55–1.92), except in IC10, where TP53 mutation is near-universal. By contrast, PIK3CA mutations were associated with worsened survival in IC10 (HR 2.39) but were neutral in the ER+ better-prognosis group.
These results underscore that the prognostic impact of recurrent driver mutations can be context-dependent and differ by inferred IntClust subtype.
All primary datasets analysed are publicly available on cBioPortal under the specified study identifiers for MSK-CHORD, METABRIC, and TCGA-BRCA. The authors state that per-patient predicted IntClust labels for MSK-CHORD, the 125-feature input matrix, and processed feature matrices are provided as supplementary files and archived at Zenodo.
The work is a secondary analysis of de-identified, publicly available data. The MSK-IMPACT programme (source of MSK-CHORD) obtained IRB approval and patient consent for research use at the originating institution. METABRIC and TCGA-BRCA data were used under their public data-use agreements. The authors report no new patient contact, specimen collection, or generation of identifiable data.
Competing interests declared include that one author co-led the METABRIC study and another author reports honoraria and support from several industry sources; other authors declared no competing interests. The manuscript is a preprint and has not been peer-reviewed; the authors explicitly note that the report should not be used to guide clinical practice until formally evaluated in peer review.
Limitations explicitly stated in the source are limited; specific model performance metrics, feature importances, and additional methodological details beyond the training and validation cohorts are not reported in the abstract. For full technical and statistical details, the reader should consult the full preprint and supplementary materials archived by the authors.