Background: Cancer-associated fibroblasts (CAFs) are major orchestrators of immunosuppression, therapeutic resistance, and extracellular matrix (ECM) remodelling in pancreatic ductal adenocarcinoma (PDAC). Building on our ongoing development of prognostic CAF-derived transcriptional signatures, we aim to identify tractable stromal vulnerabilities that can reprogram CAF states toward tumour-restraining phenotypes while preserving cellular fitness and enhancing T-cell-supportive cytokine profiles. We are currently constructing a multi-component prognostic CAF signature by integrating bulk RNA-seq (TCGA-PAAD), curated stromal gene sets, and pathway-level Cox proportional hazards modelling. This evolving signature forms the mechanistic benchmark against which perturbation-induced state transitions will be quantified. To functionally validate and expand these computational predictions, we are implementing a targeted CRISPR interference (CRISPRi) perturbation screen in primary CAFs. Approximately 50–200 ligand, receptor, and stromal-regulatory genes prioritised via NicheNet ligand-receptor inference, spatial transcriptomics, and PDAC-specific stromal networks will be perturbed. For each perturbation, bulk RNA-seq and viability readouts will be integrated into a composite TargetScore: a multi-objective metric combining ΔGSVA/ΔNES shifts toward the favourable CAF signature, an immune-support proxy (e.g., CXCL9/10 induction, MHC-II upregulation), and a viability-penalising toxicity term. These data will serve as the foundation for Tier-2 machine-learning models capable of predicting transcriptomic outcomes of unseen perturbations, enabling iterative computational–experimental refinement.