Consensus model based on 2 models

In a first step, two models based on nCounter® or GeneChip® measurements will be combined following the (i) high risk approach. For nCounter® measurements, patients will be identified as combined high risk if

• the patient specific prognostic index assigned by the LRC 7-gene signature model with clinical parameters is above 0.37 (Schmidt et al., 2018), and

• they are stratified as hypoxic by the 15-gene hypoxia signature (Linge, Löck, Krenn et al., 2016; Linge, Löck, Gudziol et al., 2016).

Based on GeneChip® measurements, the classification for the patients are assigned to the combined high risk group if:

• the patient specific prognostic index assigned by the LRC 5-gene signature model with clinical parameters is above 0.51, and

• the patients are stratified as hypoxic by the 15-gene hypoxia signature based on GeneChip® gene expression data.

Identification of gene expression signatures as prognostic biomarkers

Patients with HPV16 DNA positive tumours are considered as low risk by the LRC models but are shown as a separate group. Furthermore, only those patients are considered who have a complete GeneChip® and nCounter® data set. This may lead to differing patient numbers in figures presented in this section even though the same parameters and stratifications are shown.

Combining the 15-gene hypoxia-associated signature & LRC-signature based models The Kaplan-Meier estimators of more and less hypoxic tumours are shown in Figure 4.16 (B and E). Figure 4.16 (A and D) show the stratification based on the LRC specific signatures, extended by clinical parameters assessed for nCounter® (A) and GeneChip® measurements (D). The less hypoxic tumours and the low risk groups of each measurement method have a similar behaviour, i.e. they are nearly equal or just slightly below the Kaplan- Meier estimator of the HPV16 DNA positive patients. Less than 10% of the patients have a loco-regional recurrence. The groups of more hypoxic tumours and the high risk patients of the gene signatures are very similar, too. All 4 groups show a significant difference (p < 0.01) to their corresponding less hypoxic or low risk group. Still, only 25-35% patients of these high risk groups have a loco-regional failure.

In Figure 4.16 (C and F) the combined model, according to the high risk approach is presented, i.e. only patients who have a more hypoxic tumour and who are considered as high risk by the LRC-specific models are classified as high risk. The resulting low and high risk groups are still significantly different (p < 0.001). Furthermore, the combined high risk group shows a poorer outcome, than the high risk or the more hypoxic group of the individual models. The expected decline of the combined low risk group did only occur for the GeneChip® data. The combined low risk group based on the nCounter® measurements was similar to the individual model’s low risk and less hypoxic groups. This may be explained by the worse stratification of the 15-gene hypoxia-associated signature based on the GeneChip® expression data, see section 5.2.3.

The model combination based on the nCounter® measurements was also applied to the validation cohort. All patients assigned to the high risk group by the 7-gene signature with clinical features were also assigned into the more hypoxic group by the 15-gene hypoxia signature, see Supplementary Figure 2 A-C. After the independent normalization, the

Combined models for PORT-C

Figure 4.16. Kaplan-Meier estimators of loco-regional tumour control for the patients of

the PORT-C cohort stratified according to different models. Patients with HPV16 positive tumours are represented by the brightest colour in each Kaplan-Meier estimator. The used models are trained on nCounter® gene data in the first row and on GeneChip® data in the second row. The HPV16 DNA negative patients are stratified based on the LRC specific models (A, D), their hypoxia status (B, E) or a combination of both models (C, F). The combined high risk (comb. HR) is composed of patients who have a more hypoxic tumour and who are in the high risk group of the LRC specific model. The p-value is based on the comparison of the HPV16 DNA negative patient risk groups.

Identification of gene expression signatures as prognostic biomarkers

Combining LRC signatures

In the following subsection, the high risk approach will be used based on the stratification of both LRC specific gene signatures: 7-gene and 5-gene signature. Both measurement methods will be combined, i.e. the low and high risk status is assessed by each model on its specific gene expression data, respectively. The obtained signature includes different biological mechanisms. A combination may show an improved performance for patient stratification.

In Figure 4.17 (A and B) the same stratifications as in Figure 4.16 (B and E) are presented for better visualisation. The combination of both models is presented in Figure 4.17 C. The combined low and high risk groups are significantly different (p < 0.001). In contrast to the expectation, the combined low risk group did not show a poorer outcome compared to the individual models, even though some former high risk patients are now included in the combined low risk group. As expected, the combined high risk group (40 patients) is smaller than the individual high risk groups (7 genes: 57 patients, 5 genes: 65 patients) and shows a poorer outcome compared to the individual models.

Figure 4.17. Kaplan-Meier estimators of loco-regional tumour control for the patients of

the PORT-C cohort stratified according to different LRC-specific models. Patients with HPV16 DNA positive tumours are represented by the grey lines. The used models are (A) the 7-gene signature trained on nCounter® gene data and (B) the 5-gene signature trained on GeneChip® data. Both models are combined in (C). Patients who are classified as high risk in (A) and (B) are stratified into the combined high risk group (comb. HR). The other

Combined models for PORT-C