Integrating genetic polymorphisms and clinical data to develop predictive models for skin toxicity in breast cancer radiation therapy

Abstract

Background: We aim to develop and validate predictive models for acute and late skin toxicity in breast cancer (BC) patients undergoing radiation therapy (RT). Models incorporate a genetic profile-comprising candidate single nucleotide polymorphisms (SNPs) in non-coding RNAs and previously reported toxicity-associated variants-combined with clinical variables.

Methods: The study involved 1979 BC patients monitored for two to eight years post-RT in a multi-centre study. We assessed acute (oedema/erythema) and late (atrophy/fibrosis) toxicity using logistic regression and Cox proportional hazards models. The cohort was divided into training and validation datasets.

Results: Six SNPs demonstrated to be predictors of acute (rs13116075, rs12565978, rs72550778 and rs7284767) and late toxicity (rs16837908 and rs61764370) either in the training or validation cohort. However, none of these SNPs were consistently associated with toxicity across both stages of analysis. The rs13116075, rs12565978 and rs16837908 were previously reported to be associated with RT toxicity. In the validation phase, SNP-based models showed limited predictive ability, with AUC values of 0.49 and c-index of 0.54 for acute and late toxicity, respectively. Models incorporating either clinical variables alone or in combination with SNPs achieved similar AUC and c-index values of ∼0.60 for acute and late toxicity, respectively. However, the combined model exhibited the highest predictive accuracy for acute and late toxicity, both in the training and the validation cohorts.

Conclusions: Our findings highlight the importance of combining clinical data with genetic markers to enhance the accuracy of models predicting RT toxicity in BC.

Keywords: Breast cancer; Literature candidate polymorphisms' validation; Non-coding RNA polymorphisms; Predictive models; Radiation therapy-induced side-effects.

Fig. 1

Study flow diagram depicting numbers of analysed patients included in training and validation cohorts. RT: radiation therapy. 12m: 12 months.

Fig. 2

Sankey diagram illustrating the distribution of acute and late toxicity among the training and validation cohorts. Patients without acute toxicity are depicted in blue, while those experiencing acute toxicity are represented in red. Patients without late toxicity are shown in green, and those with late toxicity are displayed in violet. Excluded patients are represented in grey.

Fig. 3

Forest plots showing the association of clinical (orange), SNPs (blue) and SNPs + Clinical factors (violet) with radiation therapy-induced according to dichotomized high and low risk toxicity groups, in the training and validation cohorts from the acute toxicity study (A) and late toxicity study (B). CI: Confidence interval.

Fig. 4

Comparative performance metrics of the three predictive models (Clinical in orange, SNPs in blue, and SNPs + Clinical in violet) in training and validation cohort for acute toxicity. A) Receiver operating characteristic (ROC) curves in the training cohort. B) ROC curves in the validation cohort. C) Accuracy comparison in the training and validation cohorts. D) Sensitivity evaluation in the training and validation cohorts. E) Specificity assessment in the training and validation cohorts.

Fig. 5

Cumulative incidence curves reporting the late toxicity probability in the high risk and low risk strata. Analysis of the SNP model: training cohort of 520 high versus 219 low risk (A, left); validation cohort of 588 high versus 233 low risk (B, left). Analysis of the clinical data model: training cohort of 252 high versus 487 low risk (A, middle); validation cohort of 438 high versus 383 low risk (B, middle). Analysis of the combined model (clinical + SNP data): training cohort of 221 high versus 518 low risk (A, right); validation cohort of 371 high versus 450 low risk (B, right). P values by Cox model are reported.

Fig. 6

Comparative performance metrics of the three predictive Models (Clinical in orange, SNPs in blue, and SNPs + Clinical in violet) in training and validation cohorts for late toxicity. A) c-index comparison in the training, internal validation and validation cohorts. B) Accuracy comparison in the training and validation cohorts. C) Sensitivity evaluation in the training and validation cohorts. D) Specificity assessment in the training and validation cohorts.