AI-designed CAR binders push cancer stem cell immunotherapy toward safer targets

Cancer stem cells have long frustrated immunotherapy researchers. These tumor cells, defined functionally by their ability to self-renew, differentiate and regenerate tumors, can shift between molecular states, survive therapeutic pressure and seed disease recurrence. A new review published in Clinical Cancer Bulletin examines how artificial intelligence could help cellular immunotherapies, particularly chimeric antigen receptor (CAR)-based approaches, overcome the resistance and safety problems that cancer stem cells create. The authors, Visula Abeysuriya, Sachintha N. Alwis and Vasitha Abeysuriya, argue that AI offers a framework for integrating high-dimensional molecular, cellular, spatial and clinical data, but they are equally careful to state that most current applications remain computational or preclinical, and that clinical utility has not yet been established prospectively.

The review identifies a cluster of interconnected obstacles. Cancer stem cell-targeted cellular therapy is constrained by heterogeneous antigen expression, phenotypic switching, tumor microenvironmental suppression and the risk of on-target, off-tumor toxicity. Candidate targets such as CD133, CD44, EpCAM and CD123 may be heterogeneously expressed and can overlap with normal progenitor or stem-cell populations, meaning that attacking them could damage healthy tissue. These challenges are especially pronounced in solid malignancies, where limited immune-cell trafficking, hypoxia, metabolic competition and stromal barriers complicate treatment compared with hematological cancers, which generally provide more accessible and better-defined cellular targets and have already seen substantial clinical benefit from CAR-T therapy.

A central theme of the review is that cancer stem cells should not be treated as a permanently defined population identified by a single surface marker. Single-cell studies increasingly support a model in which stemness represents a dynamic cellular state that can emerge or disappear in response to developmental programs, environmental signals and therapeutic pressure. This plasticity matters therapeutically because phenotypic switching can alter the expression of therapeutic antigens during treatment, allowing tumor cells to move between target-positive and target-low states rather than remaining within a fixed compartment. Eliminating an antigen-positive population does not necessarily eliminate all cells capable of regenerating the tumor, and antigen escape can amplify the problem: CAR-T therapy has demonstrated that loss or reduction of the targeted antigen can produce relapse, prompting the development of dual-target approaches to reduce the probability of single-antigen escape.

Resistance is not produced by a single mechanism, the authors emphasize. Cancer stem cell plasticity, antigen heterogeneity, therapeutic-cell dysfunction and tumor microenvironment-mediated suppression can interact dynamically to generate treatment failure. Persistent antigen exposure, inhibitory checkpoint signaling and the tumor microenvironment can promote T cell dysfunction and exhaustion. Cancer stem cells can participate in this process through interactions with macrophages, regulatory immune cells and stromal populations, and through expression of immunosuppressive molecules such as PD-L1. Hypoxia, metabolic competition and lactate accumulation can further impair immune-cell activity within solid tumors, while cancer-associated fibroblasts can provide additional protective signals and have been implicated in resistance to CAR-T cell therapy. The review frames these mechanisms as an interconnected resistance network in which plasticity, antigen loss, checkpoint activation, metabolic reprogramming and microenvironmental suppression reinforce one another rather than operating in isolation.

Against this backdrop, the review describes how machine learning could treat cancer stem cell identity as a multidimensional and dynamic phenotype. Conventional approaches frequently rely on one or a small number of markers, whereas single-cell sequencing combined with machine learning can integrate thousands of molecular features simultaneously. Random forests, gradient-boosting methods and neural-network models can analyze high-dimensional single-cell and multi-omics datasets to identify stemness-associated states, predict treatment response or resistance, and prioritize therapeutic targets. The authors sketch an illustrative workflow: single-cell RNA-sequencing, chromatin-accessibility or multi-omics data from untreated and treatment-exposed tumors would be annotated for malignant and normal compartments; classifiers would rank candidate targets such as CD133, CD44, EpCAM and CD123 using marker abundance, co-expression, stemness scores, treatment state and normal-tissue expression; and outputs would include cross-validated discrimination, calibration, feature importance and a tumor-versus-normal expression margin. Crucially, the authors stress this is a proposed, testable workflow, not a claim that a validated AI pipeline has already established any of these four markers as safe, cancer stem cell-selective CAR targets.

The strongest direct evidence that AI can influence cellular therapy comes from computational protein design rather than biomarker discovery. Xia and colleagues developed computationally designed binders against EGFR and CD276 and incorporated them into CAR-T cells targeting glioblastoma. The engineered CAR-T cells showed improved proliferation, cytokine production, resistance to exhaustion and antitumor activity compared with CARs using antibody-derived binding domains. More recently, Chow and colleagues evaluated 1,758 AI-designed protein binders against BCMA, CD19 and CD22, combining computational design with biochemical and cellular CAR assays. That study demonstrated that AI-assisted CAR design can identify sequence and structural determinants associated with tonic signaling, epitope occlusion and off-target activity, providing experimentally testable routes for receptor optimization. The review is explicit, however, that neither study should be presented as proof that AI has solved cancer stem cell-state-specific targeting of CD133, CD44, EpCAM or CD123; they demonstrate computational design in tumor models, not a complete AI pipeline validated for cancer stem cell antigens.

Structural and generative AI adds another layer at the receptor-design stage. AlphaFold demonstrated the potential for highly accurate protein-structure prediction, while newer generative systems such as BindCraft extend computational design toward automated generation and screening of candidate protein binders. The review notes that AI-assisted structural design has progressed beyond structure prediction toward generation and optimization of functional antigen-binding proteins, creating a direct connection between computational modeling and CAR engineering. These advances may reduce the experimental search space involved in receptor development, but computational predictions cannot replace biochemical, cellular and in vivo validation, because native membrane geometry, antigen density, epitope accessibility, receptor clustering and intracellular signaling may produce outcomes not completely captured by structural models.

Safety receives parallel attention. Because several cancer stem cell-associated antigens are not completely tumor-specific, AI-assisted target selection should evaluate differential tumor-versus-normal expression rather than simply identifying antigens enriched in cancer stem cells, since bulk tumor measurements can conceal biologically relevant normal-cell expression. Computational target prioritization should incorporate tumor enrichment, normal-tissue expression, cellular specificity, antigen stability and co-expression with complementary targets rather than relying on absolute antigen abundance alone. AI may also contribute to systemic toxicity prediction: a machine-learning study comparing XGBoost, random forest and logistic-regression approaches for cytokine release syndrome risk stratification reported that XGBoost provided the strongest predictive performance in its evaluated cohort, although the review cautions that model performance from individual cohorts requires independent validation before clinical implementation. Future toxicity models could incorporate antigen density, tumor burden, CAR expansion, inflammatory markers, immune-cell phenotype and tumor-specific target expression, addressing both systemic toxicities and target-specific toxicities caused by recognition of normal cells.

The review organizes the current evidence into four maturity levels: computational prediction, preclinical validation, retrospective clinical modeling and prospective clinical validation. This stratification matters because computational accuracy, experimental efficacy and clinical effectiveness represent different levels of therapeutic evidence. The 2026 study of AI-designed CAR binders provides strong molecular and preclinical evidence but does not yet establish clinical benefit in patients, highlighting the gap between AI-enabled therapeutic design and prospective clinical validation. Longitudinal AI-guided treatment adaptation, in which serial single-cell measurements, circulating tumor DNA, imaging and immune-cell phenotyping could identify emerging antigen loss, cancer stem cell enrichment or therapeutic-cell exhaustion before overt clinical progression, remains an emerging concept rather than a clinically validated strategy. Digital twins for CAR-T therapy, patient-updated computational representations combining tumor composition, antigen density, spatial microenvironment features and toxicity measurements, are likewise described as an emerging modeling concept, not an established clinical decision tool.

Looking forward, the authors propose dynamic cancer stem cell-state modeling using longitudinal single-cell and spatial datasets, graph neural networks that represent cancer stem cell states, antigens, signaling pathways and immune populations as nodes with experimentally supported interactions as edges, and multimodal-fusion models combining transcriptomic, chromatin, imaging and clinical data. They also stress that inferred network edges are not automatically causal, and that models must be compared with simpler baselines, tested in independent cohorts and experimentally perturbed. Responsible clinical translation, they argue, requires external validation, model calibration, uncertainty quantification, domain-shift assessment and reproducibility across institutions, together with explainability, fairness, privacy and regulatory assessment. The most realistic translational pathway, the review concludes, is integration of AI prediction with molecular engineering, experimental validation and prospective clinical evaluation rather than replacement of experimental immunotherapy by AI, potentially moving cancer stem cell-targeted cellular immunotherapy toward more adaptive, precise and safer therapeutic strategies.

Subject of Research: Artificial intelligence approaches to overcoming resistance and safety challenges in cancer stem cell-targeted cellular immunotherapy

Article Title: Artificial intelligence for overcoming resistance and safety challenges in cancer stem cell-targeted cellular immunotherapy

Article References: Abeysuriya, V., Alwis, S. N., & Abeysuriya, V. (2026). Artificial intelligence for overcoming resistance and safety challenges in cancer stem cell-targeted cellular immunotherapy. Clinical Cancer Bulletin, 5(1), Article 25. https://doi.org/10.1007/s44272-026-00078-y

Image Credits: AI Generated

DOI: 10.1007/s44272-026-00078-y

Keywords: artificial intelligence, cancer stem cells, CAR-T cell therapy, antigen escape, computational protein design, tumor microenvironment, single-cell multi-omics, on-target off-tumor toxicity, resistance prediction, glioblastoma, clinical translation, machine learning