145P - Identification of biomarkers of survival across multiple cancer types using eXplainable artificial intelligence

Background

In this study, we use KEM® (Knowledge Extraction and Management) explainable Artificial Intelligence (xAI) platform to systematically explore association rules in a heterogeneous patient database accounting for above 30 cancer types and thus identify biomarkers characterizing patients with higher chances of survival. The list of candidates’ biomarkers included drug scores generated by OncoKEM®, an AI-transcriptional-based therapeutic recommendation-tool that computes scores for up to 205 drugs based on the drug's transcriptional signatures and the tumor transcriptional profile. The goal was to demonstrate the biological relevance of OncoKEM® by confronting its results with the findings obtained through a standard pathway analysis.

Methods

Data was retrieved from the PROFILER study (NCT01774409), a molecular screening program, and aggregated into a consolidated database totaling 247 patients and 215,670 variables that included survival, baseline descriptors, gene expression, derived REACTOME pathway dysregulations, and OncoKEM® scores. KEM® xAI platform extracted 55,335 relations associating candidate biomarkers and survival. These results were then filtered based on Support (number of examples), Lift (relative probability) and statistical significance. The remaining relations were finally split into 2 sets to study the associations between survival and respectively pathways dysregulations or OncoKEM® scores.

Results

Our analysis first identified 4 pathway dysregulations that were associated with the overall survival (hazard ratio ranged from 2.36 to 2.80). 3 of these pathways were related to tubulin. Consistently, 4 OncoKEM® scores were then identified as associated with survival (hazard ratio ranged from 2.20 to 2.52) and all 4 corresponded to microtubule inhibitor drugs: ixabepilone, cabazitaxel, vinflunine and brentuximab vedotin.

Conclusions

Our analysis enabled the identification of biomarkers of survival across cancer types. The consistency of the findings both demonstrated the biological relevance of OncoKEM® for microtubule inhibitor drugs and paved the way for the use of this tool as a prognostic marker for refractory cancers.

Clinical trial identification

NCT01774409.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Ariana Pharma.

Disclosure

All authors have declared no conflicts of interest.