Digestive system cancers, including hepatobiliary and gastrointestinal malignancies, remain a major global oncological burden. Although immune checkpoint inhibitors have changed the treatment landscape for some patients, their overall efficacy remains limited, with only about 15% to 30% of patients achieving durable clinical benefit. One major reason for this challenge is the highly complex and heterogeneous tumor microenvironment, in which tumor-associated macrophages play a central role in regulating immune responses and therapeutic outcomes. These macrophages are highly plastic and can dynamically shift between a pro-inflammatory, antitumor M1-like state and an immunosuppressive, tumor-promoting M2-like state; this M1/M2 balance strongly influences how patients respond to immunotherapy. However, current clinical decisions still rely mainly on static biomarkers such as PD-L1 and tumor mutational burden, which cannot adequately capture the spatial, temporal, and treatment-induced evolution of the immune microenvironment. "With the development of single-cell sequencing, multiplex immunohistochemistry, and spatial omics technologies, researchers can now characterize tumor immune ecosystems in much greater detail." said the author Lei Zhou, a researcher at Central South University, "Yet translating these high-dimensional, multimodal data into interpretable, generalizable computational models that can guide individualized therapy optimization remains a key challenge for precision cancer immunotherapy."
This study developed a deep generative immune "digital twin" framework called GI-ImmunoWorld to simulate tumor-associated macrophage immune dynamics in hepatobiliary–intestinal tumors and optimize individualized immunotherapy strategies. The researchers integrated multimodal data from 2,847 patients across 5 digestive cancer types, including hepatocellular carcinoma, cholangiocarcinoma, colorectal cancer, gastric cancer, and pancreatic ductal adenocarcinoma. These data included approximately 1.2 million single-cell transcriptomes, multiplex immunohistochemistry-based spatial immunophenotyping, and longitudinal clinical treatment and outcome records. The model first used a hierarchical variational encoder to integrate different data sources: a graph attention network captured single-cell interactions and macrophage polarization states, a convolutional neural network–Transformer module extracted spatial immune architecture features, and a multilayer perceptron encoded clinical variables. Cross-modal attention was then used to generate a unified latent representation of each patient's immune state. Next, the team introduced a Transformer-based temporal dynamics module to simulate how the immune microenvironment evolves over time under different therapeutic interventions. Finally, using model-predictive control and cross-entropy optimization, the framework searched through a clinically validated treatment action space to identify patient-specific therapeutic strategies that maximize predicted response, macrophage repolarization potential, and survival benefit while controlling potential toxicity risk.
The results showed that GI-ImmunoWorld could effectively characterize the heterogeneity of macrophage immune states in digestive system cancers and improve immunotherapy response prediction. Single-cell and spatial immune analyses revealed clear M1/M2 polarization differences among tumor types and spatial regions. A higher M2/M1 ratio was associated with lower treatment response rates and shorter survival, whereas stronger spatial interaction between CD8+ T cells and macrophages was linked to better therapeutic response. For immunotherapy response prediction, GI-ImmunoWorld achieved an AUROC of 0.847 in the test set, outperforming conventional clinical biomarker models and several deep learning baselines. In survival prediction, it also achieved a concordance index of 0.786, exceeding comparator models. Latent space analysis further showed that several learned dimensions corresponded to the M1/M2 polarization axis and CD8+ T-cell infiltration, suggesting that the model representations were biologically interpretable. The temporal dynamics module also showed good trajectory prediction performance in longitudinal biopsy data, with predicted immune-state trajectories strongly correlated with observed trajectories, and it helped identify potential macrophage repolarization windows for therapy. In an exploratory, nonrandomized model-guided treatment analysis, among patients whose model recommendations differed from physicians' original plans, those receiving model-guided regimens showed higher objective response rates and longer median progression-free survival, along with increased M1/M2 ratios, greater CD8+ T-cell infiltration, and reduced immunosuppressive signatures. However, the authors emphasized that these findings may still be affected by selection bias and require prospective randomized trials for further validation.
The significance of this work lies in moving immunotherapy optimization for digestive system cancers beyond static biomarker-based judgment toward computational simulation of the dynamic evolution of the tumor immune microenvironment. By integrating single-cell transcriptomic data, spatial immunophenotyping, and clinical trajectories, GI-ImmunoWorld builds a deep generative model centered on tumor-associated macrophage dynamics. It can not only predict immunotherapy response and survival risk, but also simulate immune-state changes under different therapeutic interventions and provide reference for individualized treatment timing and regimen selection. In particular, the model's representation of M1/M2 macrophage polarization, CD8+ T-cell infiltration, and immunosuppressive states gives its outputs a degree of biological interpretability, offering a new computational perspective on when to treat and how to combine therapies. However, this study remains at an early validation stage. The model-guided treatment results were derived from a nonrandomized exploratory analysis and may still be affected by selection bias and unmeasured confounding, so they cannot be interpreted as direct causal clinical evidence. "Future work should validate its clinical utility through larger multicenter prospective randomized trials, while also reducing the barriers of multimodal profiling and computational deployment, developing simplified models based on routine clinical data, and incorporating liquid biopsy markers, patient quality-of-life measures, broader tumor types, and stronger explainability methods to help bring immune digital twin frameworks into precision oncology practice." said Lei Zhou.
Authors of the paper include Lei Zhou, Yingjie Tan, Weigang Lv, Kening Lin, Feifeng Li, Wen Ouyang, and Di Zhang.
The project is funded by the National Natural Science Foundation of China (No. 81971028), the Natural Science Foundation of Hunan Province (No. 2023JJ30858), and the "co-PI" project from The Third Xiangya Hospital of Central South University (No. 202431).
The paper, "Deep Generative Model of Macrophage Immune Response for Hepato-intestinal Tumor Therapy Optimization" was published in the journal Cyborg and Bionic Systems on Jun 16, 2026, at https://doi.org/10.34133/cbsystems.0559.
