Abstract
Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal malignancy with limited effective therapeutic options. Gemcitabine plus nab-Paclitaxel (GnP) and FOLFIRINOX (FFX) are commonly used first-line chemotherapies for advanced PDAC. However, disease heterogeneity and late-stage diagnosis remain to result in a poor prognosis. There is a need for predictive biomarkers for selecting patients likely to respond to either regimen. Here, we analyzed data from the COMPASS trial to identify predictive mutational and transcriptional features that can guide treatment selection between GnP and FFX in advanced PDAC patients.
We utilized NetraAI, an alternative artificial intelligence/machine learning (AI/ML) technology to analyze 87 PDAC patients. NetraAI identifies explainable patient personas characterized by a set of 2-4 variables that favor a specific treatment outcome and is capable of learning from smaller data. By focusing on uncovering causal high effect size subpopulations, this approach avoids overfitting and the reinforcement of biases about PDAC. The patient population was categorized into two groups: GnP responders or FFX non-responders (GnPR/FFXNR; n =45) and GnP non-responders or FFX responders (GnPNR/FFXR; n =42). Tumor responses were assessed according to standard criteria (stable disease, partial response, and complete response).
NetraAI identified specific variables that characterize preferential response to either GnP or FFX. A key FFX preferential responder persona (n =26; 16 FFXR/GnPNR, 10 GnPR/FFXNR) was characterized by low expression levels of C3AR1 (<3), MIEF1 (<8), and HOXB6 (<20), demonstrating a significant effect size (Cohen’s D=1.086, p =0.00799). A key GnP preferential responder persona (n =28; 17 FFXR/GnPNR, 11 GnPR/FFXNR), was characterized by TRIM25 expression between 22-25, low expression levels of RAB40B (<10), and NUTF2 (<35), with a significant effect size (Cohen’s D=1.2, p =0.00294). Additionally, lower levels of Major Vault Protein (MVP) were found to be associated with response to chemotherapy in both regimens, particularly favoring GnP response when combined with other variables.
Our findings suggest that specific transcriptional and mutational biomarkers can predict preferential response to GnP or FFX in advanced PDAC patients. The use of NetraAI enables the identification of patient subpopulations characterized by a small set of variables, paving the way towards personalized treatment selection. These results highlight the potential of AI/ML modeling approaches in enhancing precision medicine for PDAC and could inform clinical decision-making to improve patient outcomes.