Immune checkpoint inhibitors have revolutionized cancer treatment by unshackling the immune system’s ability to recognize and obliterate malignant cells. These drugs target immune-suppressive proteins such as PD-1, PD-L1, and CTLA-4 that tumor cells use to cloak themselves from the body’s immune defenses. Since their FDA approval beginning in 2011, ICIs have extended survival for patients with various cancers that were once deemed incurable. Yet, their benefits are often limited to a minority subset, with response rates varying widely between 10 and 40 percent depending on cancer type. This heterogeneity in patient response remains poorly understood and limits the effective clinical application of these therapies.

Prior attempts to predict who will respond to ICIs have involved the identification of biomarkers and the use of machine learning models that analyze tumor microenvironment features, such as the presence of immune cells or so-called “immune deserts.” While these factors offer valuable clues, they fail to fully capture the biological complexity underlying patient responses, leading to unreliable predictions. Compounding this challenge is the heterogeneity not only of tumor genetics but also of the myriad ways the immune system can be activated or suppressed within the tumor milieu.

COMPASS addresses these challenges by leveraging a concept bottleneck transformer architecture, a sophisticated form of AI that prioritizes interpretability alongside predictive accuracy. Unlike traditional black-box models, COMPASS outputs transparent rationale based on the activity of nearly 16,000 genes implicated in immune cell function, tumor microenvironment interactions, and cellular signaling pathways. This architecture enables researchers and clinicians to understand the biological basis for each prediction, fostering trust and opening new avenues for scientific discovery.

The foundation of COMPASS’ training is a vast repository of genetic and molecular data derived from over 10,000 tumor samples encompassing 33 cancer types, sourced primarily from the Cancer Genome Atlas. Through this extensive dataset, the model “learned” how variations in gene expression correlate with successful responses to different ICIs. Subsequent fine-tuning involved a rigorous evaluation across 16 clinical cohorts, encompassing seven distinct cancer types and diverse ICI treatment regimes, where COMPASS demonstrated an impressive average prediction accuracy improvement of 8.5 percent over current state-of-the-art methods.

One of the most striking features of COMPASS is its capacity to elucidate atypical response patterns. For example, some nonresponding patients with tumors initially classified as “immune-inflamed” exhibited gene signatures linked to mechanisms that suppress immune attacks on cancer cells. Conversely, certain responders despite immune-desert tumor profiles showed gene expressions indicative of alternative immune pathways enabling therapeutic efficacy. These insights not only enhance the precision of patient stratification but also reveal novel biological processes at play in tumor-immune interactions.

The implications of this technology are monumental. COMPASS could soon transform clinical oncology by serving as a sophisticated decision-support tool, enabling oncologists to tailor immunotherapy choices with unprecedented precision. This would not only optimize patient outcomes but also minimize exposure to ineffective treatments and their associated toxicities. Furthermore, by enhancing patient selection in clinical trials, COMPASS could significantly accelerate the development pipeline for new immunotherapies, improving trial success rates and reducing costs.

Looking forward, the researchers plan to integrate additional layers of patient data into COMPASS, such as electronic health records detailing medical histories and previous treatment responses, as well as single-cell sequencing insights that unravel the heterogeneity within tumor and immune cell populations. Such integration holds the promise of refining the model’s predictive power even further, ushering in a new era of multi-modal precision oncology.

The design and development of COMPASS entailed close interdisciplinary collaboration, combining expertise in computational biology, oncology, and AI. The study’s first author, Wanxiang Shen, who completed this work as a research fellow in the Zitnik Lab before joining Zhejiang University, emphasizes the potential of interpretable AI methods to revolutionize cancer treatment paradigms. Their work stands as a testament to how cutting-edge technologies can tackle some of the most vexing questions in medicine.

Financial and institutional support for this study was extensive and multifaceted, involving grants and partnerships with organizations such as the National Science Foundation, pharmaceutical companies, and philanthropic foundations. This broad support underscores the high stakes and broad interest in overcoming the barriers to effective immunotherapy across cancer types.

As promising as these results are, the true test for COMPASS will come with prospective clinical trials designed to validate its predictions in real-world oncology settings. Should these trials confirm the model’s performance, the medical community may soon wield a powerful tool that not only predicts outcomes but also deepens our biological understanding of cancer-immune dynamics. Ultimately, COMPASS exemplifies the convergence of AI and biomedical research, heralding a future where cancer treatment is increasingly personalized, effective, and reasoned.

Subject of Research: Predictive modeling of patient response to immune checkpoint inhibitor cancer immunotherapy using tumor gene expression data.

Article Title: Generalizable AI predicts immunotherapy outcomes across cancers and treatments

News Publication Date: 3-Jul-2026

Web References: https://www.nature.com/articles/s41591-026-04502-7

Keywords: Cancer immunotherapy, immune checkpoint inhibitors, AI-driven prediction, gene expression analysis, tumor microenvironment, precision oncology, interpretable artificial intelligence, immune response biomarkers, machine learning, clinical trial optimization, personalized medicine

At the core of tumor biology, the TME represents a complex, multicellular ecosystem comprising not only malignant cells but also diverse immune cells, stromal components, blood vessels, and a surprising influence of neural elements. These components are not static entities; instead, they co-evolve and interact in a highly coordinated manner that influences tumor initiation, progression, immune evasion, and, critically, therapeutic outcomes. Harnessing the power of single-cell sequencing and spatial omics, researchers have transcended traditional bulk analyses, enabling a high-dimensional, panoramic view that captures cellular diversity and spatial relationships at an unprecedented resolution.

Among the immune effectors, lymphocytes stand out as the frontline warriors in anti-tumor immunity, with CD8+ cytotoxic T lymphocytes (CTLs) playing a quintessential role by recognizing tumor-specific antigens presented via major histocompatibility complex class I (MHC-I) molecules and mediating tumor cell lysis through cytotoxic molecules such as perforin and granzymes. However, the suppressive nature of the TME frequently drives these CTLs into an exhausted functional state, marked by reduced cytotoxicity and proliferative capacity. Intriguingly, a subset of CD8+ T cells expressing the chemokine CXCL13 has been identified as pre-exhausted but functionally significant, correlating with favorable responses to immune checkpoint blockade (ICB), signaling a nuanced balance within T cell states that could be exploited for therapeutic benefit.

Beyond classical T cells, B cells and natural killer (NK) cells constitute essential, though often underappreciated, components of the tumor immune milieu. Tumor-associated B cells, characterized by high expression of FCRL4 and MHC-II molecules, demonstrate a potent antigen-presenting capacity that is linked to enhanced patient prognosis and improved ICB responses. Conversely, NK cells within the TME frequently adopt a dysfunctional phenotype marked by downregulated cytotoxic pathways, as observed by DNAJB1 expression, contributing to poor clinical outcomes and resistance to PD-1-directed therapies. These observations underscore the complexity of immune cell states within solid tumors and their critical role in shaping therapeutic responses.

The myeloid compartment within the tumor also presents a diverse cellular repertoire, with macrophages, dendritic cells (DCs), neutrophils, and mast cells exhibiting distinct polarization states and functional repertoires. The traditionally simplistic M1/M2 macrophage paradigm is being supplanted by more sophisticated models, such as one centered on mutually exclusive CXCL9 and SPP1 expression. Notably, SPP1+ tumor-associated macrophages have emerged as key pro-tumorigenic players, fostering tumor angiogenesis, extracellular matrix remodeling, and hypoxic adaptations, all hallmarks of aggressive disease and poor prognosis. Likewise, LAMP3+ dendritic cells, particularly subsets derived from conventional type 1 DCs (cDC1) producing CXCL9 and interleukin-15, are instrumental in recruiting and sustaining CD8+ T cell effector responses and mediating responsiveness to immunotherapies.

The stromal compartment adds another layer of complexity; cancer-associated fibroblasts (CAFs), especially those expressing LRRC15, exemplify terminal differentiation states associated with immune exclusion and resistance mediated through transforming growth factor-beta (TGF-β) signaling pathways. Endothelial tip cells marked by CXCR4 expression catalyze aberrant angiogenesis, frequently correlating with adverse outcomes. On the other hand, tumor-associated high endothelial venules and ACKR1+ endothelial cells facilitate immune infiltration, highlighting a dualistic role of vasculature in tumor immunity. More recently, the intersection of neural biology and oncology has revealed TGFBI+ Schwann cells within tumors, which are induced by TGF-β and potentiate tumor cell migration, underscoring a complex neuro-immune-tumor crosstalk that was previously unappreciated.

Crucially, these individual cellular players do not exist in isolation but form spatially organized, functionally integrated multicellular networks within the TME. The identification of ‘immunity hubs’—cellular modules comprised of LAMP3+ dendritic cells, TCF7+ T cells, and CCL19+ fibroblasts—illustrates how coordinated cellular consortia establish niches critical for effective immune surveillance and response. The integrity and spatial arrangement of these hubs strongly predict immunotherapy outcomes. However, tumor progression drives the degradation of healthy multicellular networks and the emergence of aberrant, conserved oncogenic modules, providing insights into shared TME remodeling trajectories that transcend tumor types and offer targets for broad-spectrum therapies.

Looking toward the future, the review highlights a visionary framework termed the “AI virtual tumor”—a computational ecosystem that integrates cellular composition, spatial tissue architecture, intercellular communication, and response to perturbations to model tumor-scale dynamics in silico. This AI-driven paradigm could revolutionize patient stratification, enable in silico hypothesis testing, optimize combination therapy design, and predict treatment efficacy with unprecedented accuracy. Such digital twin models combine high-dimensional biological data with advanced computational algorithms, driving precision oncology toward a new horizon.

In the domain of immunotherapy, the review delineates three promising frontiers. Immune checkpoint blockade (ICB) therapies benefit from biomarkers such as CXCL13+ T cells that predict favorable clinical responses, whereas cell types like CCR8+ regulatory T cells, SPP1+ macrophages, and LRRC15+ CAFs are associated with resistance mechanisms. Remarkably, novel dual checkpoint inhibitors, such as the combination of LAG-3 and PD-1 blockade, have demonstrated encouraging clinical success. Meanwhile, adoptive cell therapies progress with CAR-T cells revolutionizing hematological malignancy treatment and emerging CAR-macrophage (CAR-M) therapies showing potential in solid tumors due to superior tumor infiltration, currently undergoing early-phase clinical trials.

Further, personalized cancer vaccines are gaining traction, with cDC1-targeted vaccines offering strategies to circumvent ICB resistance, exemplified in pancreatic cancer models. mRNA neoantigen vaccines evaluated in high-risk renal cell carcinoma patients have demonstrated safety and immunogenicity, heralding a new era of patient-specific immunotherapy that synergizes with insights from spatial and single-cell analyses. Collectively, these advances exemplify an integrated pathway from fundamental tumor biology investigation to innovative, AI-supported immunotherapy modalities.

The synthesis provided by this review offers an indispensable roadmap linking cell biology, spatial organization, and computational modeling with clinical applications in cancer immunotherapy. By illuminating specialized cellular subtypes and their coordinated networks within the TME, this research advances our understanding of tumor heterogeneity and therapeutic resistance. Moreover, the AI virtual tumor concept promises to catalyze a paradigm shift, enabling in silico experimentation and rational design of next-generation, mechanism-based precision immunotherapies that could significantly improve patient outcomes.

As the realm of cancer treatment moves toward increasingly personalized approaches, the interweaving of single-cell genomics, spatial biology, and computational intelligence foretells a future where detailed biological knowledge is harnessed alongside artificial intelligence to confront the multifaceted challenges posed by tumors. This work by Zhu, Zhang, and colleagues exemplifies how multidisciplinary integration can transform cancer research, inspiring new strategies that transcend existing therapeutic limitations and usher in a new era of immuno-oncology.

Subject of Research: Not applicable
Article Title: The cellular actors of the tumor microenvironment: a single‑cell atlas perspective on specialized subtypes, coordinated networks, and immunotherapy
News Publication Date: 5-Jun-2026
References: DOI 10.1007/s44466-026-00043-3
Image Credits: Professor Zemin Zhang and Dr. Linnan Zhu from Peking University, China, and Chongqing Medical University, China