Feasibility of an explainable AI-based therapeutic recommendation-tool utilizing tumor gene expression profiles in advanced and refractory solid tumors.

• DOI: 10.1200/jco.2022.40.16_suppl.1554
• Publication Date: 2022-06-06T15:57:14Z
• Authors: Ouissam Al Jarroudi; Coralie Williams; Rita Santos; Armelle Dufresne; Valéry Attignon; Anthony Ferrari; Sandrine Boyault; Laurie Tonon; Séverine Tabone-E...; Philippe Alexandr...; Nadège Corradini; Armelle Vinceneux; Aurélie Swalduz; Alain Viari; Sylvie Chabaud; David Pérol; Mohammad Afshar; Jean-Yves Blay; Olivier Tredan; Pierre Saintigny
• Keywords: 03 medical and health sciences; 0302 clinical medicine; 3. Good health

1554 Background: Precision oncology aims to guide patient (pts) treatment decisions by matching biological features with available drugs. Extensive genomic analysis allows to identify an actionable alteration in 40-60% of patients. In a recent study of 50 pts with advanced refractory diseases included in PROFILER (NCT01774409), whole exome and fusion transcripts had a limited value over a 90-tumor gene panel (TGP) to increase molecular-based treatment recommendations (MBTR). Herein, we evaluated the feasibility, in the same cohort of pts, of the AI-transcriptional-based therapeutic recommendation-tool OncoKEM to guide treatment recommendations. Methods: 77 fresh frozen (FF) and/or FFPE samples including paired specimens for 53 pts with available RNA-Seq gene expression profiles were included. For each pts, a tumor transcriptional profile (TTP) was generated by identifying differentially expressed genes between the pts tumor and a cohort of matched healthy tissue. A large database of drug transcriptional signatures (DTS) was queried in order to identify a “reversal relationship” between the TTP and a DTS. A total of 205 drugs were ranked, including a subset of 61 FDA and/or EMA approved targeted therapies (aTT). Results: Most common diagnoses were breast cancers (21% of which 63% were TNBC), followed by ovarian cancers (OC, 18%) and soft-tissue sarcomas (STS, 13%). The median number of previous treatment lines was 4 (range: 1 - 10). Among the 77 tumor samples analyzed, 54 (70%) specimens led to the generation of an OncoKEM report, with no differences between FF and FFPE samples (p = 0.85). The overlap between the top 10 proposed drugs between paired FF and FFPE samples was 56% on average. All patients had at least 2 propositions (range: 2-9) of aTT among the top 10 ranked drugs in the Onco KEM reports. Most frequently proposed drugs among the top 10 were palbociclib, talazoparib, infigratinib in TNBC; bosutinib, sapanisertib, SAR125844 in OC; ipilimumab, cabozantinib, sapanisertib in STS. Among the 30 pts (79%) without any MBTR based on TGP/WES/fusion transcript analysis, all had at least 2 proposed aTT in the Onco KEM report (median: 4, range: 2-9). Top ranked drugs were MET (18%), VEGFR (12%), Abl (12%), FGFR (11%), PI3K/AKT/mTOR (11%), PARP (10%) and CDK4/6 inhibitors (7%). Conclusions: AI-transcriptional-based therapeutic recommendation-tool OncoKEM is feasible and has the potential to expand personalized cancer treatment in pts with advanced & refractory diseases without tractable genomic alterations. The clinical relevance assessment is planned in an upcoming clinical trial.