Immune checkpoint inhibitors (ICIs) — drugs like pembrolizumab and nivolumab — have revolutionized treatment for several cancers including melanoma and lung cancer, dramatically extending survival for some patients. However, pancreatic cancer has proven stubbornly resistant to immunotherapy, with clinical trials showing little benefit in most patients.
The reason for this resistance lies in pancreatic cancer's unique tumor microenvironment: it is densely packed with fibrous tissue that physically blocks immune cells from entering, and it actively suppresses the immune cells that do manage to infiltrate. These 'cold' tumors lack the immune activity that makes immunotherapy effective, and predicting which rare patients might respond remains a major unsolved challenge.
Researchers are exploring multiple strategies to overcome pancreatic cancer's immune resistance. One approach is to convert 'cold' tumors to 'hot' ones — more immunologically active tumors — through combinations of radiotherapy, targeted therapy, or vaccines that stimulate immune recognition of the tumor.
Combination therapies are showing particular promise. Clinical trials are testing combinations such as PARP inhibitors with checkpoint inhibitors, kinase inhibitors with PD-1 blockers, and personalized mRNA neoantigen vaccines with immunotherapy. Some trials have shown improved progression-free survival compared to single agents, though overall survival benefits remain modest for most patients.
Radiomics — the extraction of quantitative features from medical images — can characterize the composition of the tumor microenvironment non-invasively from CT and MRI scans. Features reflecting tumor texture, heterogeneity, and shape have been correlated with immune cell infiltration patterns, offering a window into the immune status of a tumor without a biopsy.
Deep learning models can go further, analyzing entire imaging datasets to identify complex spatial patterns of immune activity. These models can potentially predict, before treatment starts, whether a patient's tumor is likely to respond to checkpoint inhibitor therapy, sparing non-responders from treatment toxicity while ensuring responders receive potentially life-extending therapy.
A major clinical challenge with immunotherapy is pseudoprogression — where tumors appear to enlarge on scans because they are being infiltrated by immune cells, even though the treatment is actually working. Standard response criteria like RECIST, developed for chemotherapy, misidentify pseudoprogression as treatment failure, leading to premature cessation of effective therapy.
AI models trained to distinguish true progression from pseudoprogression using serial imaging data could be transformative. By analyzing temporal changes in imaging features over multiple scans, deep learning can identify patterns associated with immune response that differ from the patterns of true disease progression, potentially guiding oncologists to continue effective therapy even when conventional criteria suggest otherwise.
The review highlights several promising AI-driven biomarker approaches for pancreatic immunotherapy. Tumor mutational burden (TMB) and microsatellite instability (MSI) status predict response to checkpoint inhibitors across cancer types, and AI models analyzing histology images can predict these genetic features non-invasively. The rare subset of pancreatic cancers with high TMB or MSI may respond to pembrolizumab.
PET imaging combined with AI analysis offers another promising avenue: novel tracers targeting immune cell activity or checkpoint proteins can visualize immune responses within the tumor in real time. AI-based analysis of these PET images could quantify the degree of immune activation, providing an earlier and more accurate signal of treatment response than anatomical size measurements alone.
The authors conclude that the integration of AI with immunotherapy represents one of the most promising frontiers in pancreatic cancer treatment. While no single AI tool is yet ready for routine clinical deployment, the combination of radiomics, deep learning, and multi-omic biomarker analysis is rapidly advancing toward that goal.
Future research should focus on multi-modal AI systems that combine imaging, genomic, and clinical data to create comprehensive immunotherapy response predictors. The authors also call for prospective clinical trials that incorporate AI-based patient selection as a prespecified stratification factor, which would generate the high-quality evidence needed to move AI tools from research into clinical practice.