Development of prediction models for lymph node metastasis in endometrioid endometrial carcinoma

Abstract

Background: In endometrioid endometrial cancer (EEC), current clinical algorithms do not accurately predict patients with lymph node metastasis (LNM), leading to both under- and over-treatment. We aimed to develop models that integrate protein data with clinical information to identify patients requiring more aggressive surgery, including lymphadenectomy.

Methods: Protein expression profiles were generated for 399 patients using reverse-phase protein array. Three generalised linear models were built on proteins and clinical information (model 1), also with magnetic resonance imaging included (model 2), and on proteins only (model 3), using a training set, and tested in independent sets. Gene expression data from the tumours were used for confirmatory testing.

Results: LNM was predicted with area under the curve 0.72-0.89 and cyclin D1; fibronectin and grade were identified as important markers. High levels of fibronectin and cyclin D1 were associated with poor survival (p = 0.018), and with markers of tumour aggressiveness. Upregulation of both FN1 and CCND1 messenger RNA was related to cancer invasion and mesenchymal phenotype.

Conclusions: We demonstrate that data-driven prediction models, adding protein markers to clinical information, have potential to significantly improve preoperative identification of patients with LNM in EEC.

Fig. 1. Overview of training and test sets and corresponding models.

Lines with number of patients for each set connects the models to the training and test set used in each model. All models were trained using Bergen training (white colour) and tested in either the Bergen test set (dark grey) or in the MDACC test set (light grey).

Fig. 2. Prediction model (model 1) and RNA confirmation.

Receiver operating characteristic (ROC) curves for model 1 used as a continuous model (a, c) and as a categorical model (b, d) in the Bergen training and test cohorts. The number of patients is shown in the lower panel (grey scale). e Scatter plot of cyclin D1 and fibronectin RPPA protein levels vs. mRNA levels. Black round dots and square white dots illustrate fibronectin/FN1 and cyclin D1/CCND1 expression levels for each case, respectively. RNA validation for model 1 used as a continuous model (f), and as a categorical model (g). The number of patients is shown in the lower panel.

Fig. 3. Alternative lymph node metastasis prediction models using MRI and protein data (model 2) or protein data only (model 3).

Receiver operating characteristic (ROC) curves for model 2, using MRI and protein data in the Bergen training (a) and the Bergen test cohort (b), and model 3, using significantly altered proteins only in a subpopulation of presumed low-risk patients in the Norwegian training cohort (c) and the MDACC test cohort (d). The number of patients is shown in the lower panel (grey scale).

Fig. 4. Disease-specific survival in patients with high vs. low fibronectin and cyclin D1 levels.

Disease-specific survival (DSS) according to fibronectin and cyclin D1 high combined (i.e. 25% upper quartile of both proteins) vs. low/discordant expression (i.e. all other combinations: 75% lower quartile of at least one protein) (a). DSS was also calculated for the individual variables: fibronectin (b) and cyclin D1 (c) high (upper 25% quartile) and low (lower 75% quartile) expression. The number of events is given within parentheses.