166P - Metabolomic prediction of breast cancer treatment toxicities

Background

Long term treatment related toxicity is a major issue for breast cancer patients in the adjuvant setting. Predicting toxicities may allow us to adapt the treatment strategy. We assessed whether the metabolomic profile of patients may predict long-term toxicities.

Methods

High-resolution untargeted metabolomics was performed at baseline for 857 ER-positive, HER2- breast cancer patients from the CANTO prospective cohort. Four metabolomic profiles per patient were produced: (i) shared and annotated metabolites (n=224), (ii) annotated but not always common metabolites (n=456), (iii) annotated but not always shared metabolites (n=766) and (iv) all metabolites (n=1693, FullMet). Samples were split into a discovery and validation set. We benchmarked algorithms adapted for high dimensional analysis (LASSO, Adaptive LASSO, machine learning, and deep learning) in order to select best models for prediction.

Results

30.0% of patients were >65 years old, 24.4% <50 years old, 20.4% had BMI>30, 12.7% had previous history of neurological disorders, 6.1% had diabetes. 69.6% presented with pT1, 25.7% with pT2 and 3.4% with pT3; 11.1% had lymph node involvement. Among all benchmarked, adaptive LASSO was the most interesting statistical method with limited optimism bias. It also allows the selection of a subset of metabolites of particular interest. The addition of rare metabolites as well as non-annotated metabolites significantly increase the predictive power of models. Metabolic toxicity prediction mainly relied on endogenous metabolites while neurological toxicities were partly predicted using exogenous/environmental metabolites. In the validation set, compared to clinical data alone (AUC 0.50-0.54), addition of metabolomics data shows moderate (AUC = 0.55-0.60) but significant (p<0.05 adjusted for multiple comparison) predictive ability for neurological and metabolic toxicities.

Conclusions

Breast cancer patient metabolomic profile at baseline improves toxicity prediction after adjuvant chemotherapy, similar to what is reported for genomic fingerprints. Untargeted metabolomics allows the achievement of higher performance by taking into account environmental exposure, metabolites linked to microbiota as well as rare and uncommon metabolites.

Clinical trial identification

NCT01993498.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

CANTO consortium.

Disclosure

All authors have declared no conflicts of interest.