Combining different genomic signatures to improve the predictive power for LRC after PORT-C in HNSCC

Combining different genomic signatures to improve the predictive power for LRC after PORT-C in HNSCC

Schmidt, S.; Linge, A.; Zwanenburg, A.; Leger, S.; Großer, M.; Lohaus, F.; Gudziol, V.; Nowak, A.; Tinhofer, I.; Budach, V.; Stuschke, M.; Balermpas, P.; Rödel, C.; Grosu, A.-L.; Abdollahi, A.; Debus, J.; Belka, C.; Combs, S. E.; Mönnich, D.; Zips, D.; Baretton, G. B.; Buchholz, F.; Baumann, M.; Krause, M.; Löck, S.

Purpose: To improve compare and improve the performance of a hypothesis-driven 7-gene signature by with a signature based on whole transcriptome analysis for the prognosis of loco-regional tumour control (LRC) in patients with HPV-negative locally advanced head and neck squamous cell carcinoma (HNSCC) after postoperative radiochemotherapy (PORT-C).

Material and methods: Gene expression analyses were performed on a multicentre retrospective cohort of 125 patients with HPV16 DNA negative HNSCC using the GeneChip® Human Transcriptome Array 2.0 (Affymetrix) for whole transcriptome analysis. To identify a gene signature prognostic for LRC from the whole transcriptome data, 3085 genes were considered, which previously have been related to radioresistance or response to radiotherapy [1-4]. The final gene signature was based on the comparison of different signature sizes, feature selection algorithms and prognostic models. The performance of the whole transcriptome-based signature was compared to a previously identified 7-gene signature based on nanoString analysis of a hypothesis-driven gene set containing 171 genes, using the concordance index (ci). The signatures were applied independently and combined to stratify patients into groups of low (LR) and high (HR) risk of recurrence.

Results: The identified gene signatures based on whole transcriptome data showed improved performance (ci 0.79-0.87) compared to the signatures based on the hypothesis-driven gene set (0.72-0.78). The model with the best performing gene signature contained genes related to tumourigenesis, invasion, cell cycle regulation and immune response. Patient stratification into low and high risk groups was performed for both signatures, see figures (A) and (B). The difference in LRC between both groups was highly significant (p<0.001). Compared to the 7-gene nanoString signature, the LR group showed a slightly improved LRC for the Affymetrix signature, similar to that of HPV positive tumours. Finally, a combined high risk group was defined, including patients who were classified as high risk patients by both gene signatures. This patient group showed a poor LRC of only about 45% compared to the individual signatures, see figure (C).

Conclusion: We determined a gene signature predicting LRC in a cohort of 125 HPV16 DNA negative HNSCC patients after PORT-C based on whole transcriptome analysis.
This signature showed improved performance compared to the 7-gene signature identified on a limited hypothesis-driven gene set, indicating that the inclusion of additional genes during feature selection may lead to a better performing signature.s may further enhance this signature.
The combination of both models allowed for the identification of a patient group with HPV-negative HNSCC who are on a particularly high risk of developing a recurrence, and may be considered for future dose-escalation trials.

Keywords: Head and neck squamous cell carcinoma; Genomics; Machine Learning; HNSCC; Cancer; Radiotherapy

  • Lecture (Conference)
    ESTRO 37, 20.-24.04.2018, Barcelona, Spanien
  • Open Access Logo Abstract in refereed journal
    Radiotherapy and Oncology 127(2018), S140-S141
    DOI: 10.1016/S0167-8140(18)30586-3

Publ.-Id: 26207