Digestive tract cancers, including hepatobiliary and gastrointestinal malignancies, remain a major oncological burden globally. Immunotherapy efficacy rates are low, with only 15% to 30% of patients experiencing responses following treatment. Tumor-associated macrophages, which change phenotype between a pro-inflammatory and an immunosuppressive state, play a key role in determining the response to therapy, and current static biomarkers are inadequate for capturing the spatial–temporal changes associated with the immune response. We developed a bioinspired digital twin platform integrating variational representation learning with causal sequence modeling. The platform incorporates heterogeneous biological data (1.2 million single-cell transcriptomes, spatial immunophenotyping, and clinical trajectories) from 2,847 individuals across 5 digestive cancer types. Graph-based attention mechanisms encode intercellular interactions, while transformer-based temporal modules simulate immunological state transitions. A model-predictive optimization layer identifies patient-specific interventions maximizing repolarization potential. The biomimetic model predicted the outcome of therapy response better than conventional biomarker models did (area under the receiver operating characteristic curve: 0.847 compared to 0.692 with a statistically significant difference at P below 0.001). In an exploratory, nonrandomized analysis of discordant cases (n = 156) where model and physician recommendations differed, model-guided treatment was associated with higher response rates (47.4% versus 28.2%) and longer median progression-free survival (9.8 months versus 6.0 months; P = 0.003); however, selection bias cannot be excluded. This study provides preliminary evidence for the feasibility of a computational framework for immunotherapy optimization; prospective randomized trials are required to establish clinical utility.