Development and validation of AI-assisted transcriptomic signatures to personalize adjuvant chemotherapy in patients with pancreatic ductal adenocarcinoma.

ABSTRACT

BACKGROUND: After surgical resection of pancreatic ductal adenocarcinoma (PDAC), patients are predominantly treated with adjuvant chemotherapy, commonly consisting of gemcitabine (GEM)-based regimens or the modified FOLFIRINOX (mFFX) regimen. While mFFX regimen has been shown to be more effective than GEM-based regimens, it is also associated with higher toxicity. Current treatment decisions are based on patient performance status rather than on the molecular characteristics of the tumor. To address this gap, the goal of this study was to develop drug-specific transcriptomic signatures for personalized chemotherapy treatment. PATIENTS AND METHODS: We used PDAC datasets from preclinical models, encompassing chemotherapy response profiles for the mFFX regimen components. From them we identified specific gene transcripts associated with chemotherapy response. Three transcriptomic artificial intelligence signatures were obtained by combining independent component analysis and the least absolute shrinkage and selection operator-random forest approach. We integrated a previously developed GEM signature with three newly developed ones. The machine learning strategy employed to enhance these signatures incorporates transcriptomic features from the tumor microenvironment, leading to the development of the 'Pancreas-View' tool ultimately clinically validated in a cohort of 343 patients from the PRODIGE-24/CCTG PA6 trial. RESULTS: Patients who were predicted to be sensitive to the administered drugs (n = 164; 47.8%) had longer disease-free survival (DFS) than the other patients. The median DFS in the mFFX-sensitive group treated with mFFX was 50.0 months [stratified hazard ratio (HR) 0.31, 95% confidence interval (CI) 0.21-0.44, P < 0.001] and 33.7 months (stratified HR 0.40, 95% CI 0.17-0.59, P < 0.001) in the GEM-sensitive group when treated with GEM. Comparatively patients with signature predictions unmatched with the treatments (n = 86; 25.1%) or those resistant to all drugs (n = 93; 27.1%) had shorter DFS (10.6 and 10.8 months, respectively). CONCLUSIONS: This study presents a transcriptome-based tool that was developed using preclinical models and machine learning to accurately predict sensitivity to mFFX and GEM.