473P - A novel machine learning (ML) model integrating clinical and molecular data to predict response to second-line treatment in recurrent IDHwt-glioblastoma (rGBM)

Background

Nitrosoureas (lomustine or fotemustine) and antiangiogenic (bevacizumab or regorafenib) are second-line treatment options for patients with rGBM, but to date predictors or efficacy are lacking. We aimed to develop a multi-classification ML algorithm to predict the response to different second-line therapies.

Methods

We conducted a retrospective study to assess the combined predictive value of molecular profiles (tested by FoundationOne®CDx on tissue) and clinical data (clinical-pathological and treatment characteristics) in a cohort of patients with first rGBM treated with Regorafenib, Lomustine/Fotemustine or Bevacizumab from Oct 2019 to Jan 2023. WHO2021 classification was used for pathological diagnosis, and RANO criteria for neuroradiologic response evaluation. ML was applied to identify the best responders, using a gradient boosting framework basedf on tree-based algorithms LightGBM to integrate clinical and genomic data. Matthews correlation coefficients (MCC) were used as a metric. Best responders were labeled for having a median progression free survival (mPFS) above the third quartile for each treatment.

Results

153 patients were enrolled, treated with Regorafenib (n=95), Bevacizumab (n=19), Nitrosoureas (n=39). The mPFS in each treatment cohort was, respectively, 2.1 m (95% CI 1.9-3.3), 3.6 m (95% CI 2.1-4.8) and 3.0 m (95% CI 2.4-5.7). The mOS was, respectively, 12 m (95% CI 9.1-14), 7.0 m (95% CI 6.0-11), and 8.0 m (95% CI 5.0-13). By performing the multi-classification model we obtained a MCC of 0.32. The SHAP analysis showed age at diagnosis, gender, MGMT and CDKN2B, CDKN2A, PTEN, EGFR, TP53, MTAP mutational status as the most important predictors. These variables were ranked in a different order for each second-line treatment, thus demonstrating that the multi-classification ML approach could discriminate which patients are likely to benefit from each second-line treatment.

Conclusions

The multi-classification ML model developed in this study was able to identify clinical and molecular signatures of rGBM responding to Bevacizumab,Regorafenib or Nitrosuree. This model could be useful to choose the more effective second-line therapy in rGBM.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

G. Lombardi: Financial Interests, Personal, Advisory Board: Janssen, Braun, Helath4U; Financial Interests, Personal, Invited Speaker: Bayer, Orbus, Novartis, Servier; Financial Interests, Institutional, Coordinating PI: Bayer SpA. All other authors have declared no conflicts of interest.