Chimeric antigen receptor (CAR)-based therapies represent a transformative class of cancer immunotherapies in which patient immune cells, primarily T cells and natural killer (NK) cells, are genetically engineered to express synthetic receptors that recognize and kill tumor cells with high specificity. CAR T-cell therapy has achieved regulatory approval for B-cell lymphomas, multiple myeloma, and acute lymphoblastic leukemia, and is under active investigation for non-small-cell lung cancer, glioblastoma, and HIV. Despite these gains, the broader field is constrained by manufacturing complexity, tumor microenvironment (TME) immune suppression, antigen escape, and unpredictable toxicity profiles. AI is being positioned as the tool to address each of these constraints systematically.
The immune suppression problem: The TME in lymphoma and other cancers contains PD-L1-expressing cells that engage the PD-1 receptor on CAR T cells, dramatically reducing their cytotoxic activity. PD-1 (also called CD279) is an inhibitory receptor upregulated on activated T cells, and its binding by PD-L1 or PD-L2 triggers apoptosis of effector T cells and promotes regulatory T-cell activity. Approximately 20-50% of patients treated with FDA-approved immune checkpoint inhibitors (pembrolizumab, avelumab, ipilimumab) have a meaningful clinical response, meaning the majority derive limited benefit from current approaches. Combining PD-1 blockade with CAR T therapy has shown promise but carries systemic autoimmune risks.
Scope of this review: Published in 2024, this comprehensive review by Shahzadi, Rafique, Waheed, and colleagues surveys the current applications of AI across the full CAR therapy pipeline, from CAR construct design and antigen targeting to manufacturing optimization, toxicity prediction, and outcome forecasting. The authors synthesize evidence from deep learning (DL), natural language processing (NLP), computer vision, and reinforcement learning (RL) applications, and frame the outstanding challenges in data standardization, model validation, and regulatory translation.
The review explicitly frames AI not as a replacement for clinical expertise but as a complementary computational layer that can process and integrate multidimensional biological data at scales that exceed human capacity. Given the enormous complexity of CAR therapy design space, including antigen selection, co-stimulatory domain configuration, linker design, and infusion scheduling, AI offers a principled way to navigate this space more efficiently than empirical trial-and-error.
Immune checkpoint inhibitors (ICIs) are monoclonal antibodies targeting inhibitory receptors on CD4+ T cells and antigen-presenting cells. The three major approved classes target CTLA-4 (ipilimumab), PD-1 (nivolumab, pembrolizumab), and PD-L1 (avelumab, atezolizumab). These agents have reshaped treatment for melanoma, non-small-cell lung cancer (NSCLC), renal cell carcinoma, and head and neck squamous cell carcinoma (HNSCC). In NSCLC specifically, PD-1 inhibitors versus chemotherapy comparisons have shown 5-year overall survival rates of 13-25% in second-line treatment and up to 32% in first-line settings, a major improvement over historical baselines.
Response rate gaps: Despite these gains, approximately 60% of ICI-treated patients do not achieve a clinically significant response. The mechanisms of resistance include macrophage-mediated removal of anti-PD-1 antibodies from T-cell surfaces (documented by Arlauckas et al.), impaired T-cell memory formation, and tumor-driven changes in the TME that sustain angiogenesis and suppress immune infiltration. Combination ICI regimens, such as nivolumab plus ipilimumab, amplify efficacy but also increase immune-related adverse events (irAEs): Wolchok et al. found that this combination caused significant irAEs, including cardiac and neurological inflammation, in up to 60% of patients, with 42% of melanoma patients requiring treatment discontinuation.
irAE complexity: irAEs vary substantially in timing and organ involvement. Ipilimumab-related irAEs typically emerge around week four of treatment, while nivolumab toxicities peak around week ten. Importantly, ICI toxicity is not dose-proportional, so dose reductions do not reliably prevent adverse events. Some irAEs (dermatitis, pneumonitis) are reversible due to organ regenerative capacity, while others (adrenal insufficiency, insulin-dependent diabetes) cause permanent endocrine damage. The clinical challenge of distinguishing true progression from pseudoprogression (PP) and hyperprogression (HP) on imaging also remains unsolved without AI-assisted analysis.
These well-characterized limitations make ICIs a primary target for AI augmentation: AI can potentially identify the 40% of patients likely to respond before treatment begins, predict which irAE profiles a given patient is likely to experience based on baseline clinical and genomic data, and distinguish PP from true progression on serial imaging by detecting quantitative signal patterns invisible to the human eye.
Predicting which patients will benefit from immunotherapy, whether ICI or CAR T, is one of the most consequential clinical problems AI can address. The standard pipeline involves acquiring multiscale medical data from the training cohort, filtering and segmenting it, extracting and selecting features, and then training an AI model. Data sources include genomics, proteomics, pathology tissue images, CT and MR imaging-omics (radiomics), and clinical variables. The ultimate goal is a model that can determine, before treatment begins, whether immunotherapy will benefit a specific patient.
Histopathological features: AI applied to histopathology slides can predict immunotherapy response by analyzing tumor-infiltrating lymphocyte patterns, PD-L1 expression distribution, and spatial relationships between immune and tumor cells. Traditional histopathology depends on specialists manually extracting information from complex images, an approach that is inherently limited in throughput and reproducibility. Machine learning classifiers trained on histopathology images provide fresh approaches to predicting tumor immunotherapy response, with studies showing that AI achieves a recognition accuracy of 91.66% for MHC (major histocompatibility complex) patterns linked to immunotherapy response.
Genomic and multi-omics approaches: Next-generation sequencing (NGS) and whole-genome sequencing (WGS) generate data on tumor mutational burden, microsatellite instability (MSI), copy number alterations, and gene expression profiles that are strongly associated with immunotherapy outcomes. AI models, particularly DL techniques, are needed to extract actionable predictions from these high-dimensional genomic datasets. PD-L1 expression, TME composition, TP53 mutations, ALK rearrangements, gut microbiota composition, Fc-gamma receptor polymorphisms, serum complement levels, and microRNA anomalies have all been identified as contributors to immunotherapy response. AI can integrate all of these variables simultaneously, which manual clinical scoring cannot achieve.
Immunotherapy prediction scores: AI has enabled the development of immunotherapy-associated composite scores, including the Immunoscore and Immunophenoscore, which quantify the immune contexture of tumors and predict the probability of responding to ICB treatment. Both scores are designed to assist clinicians in determining recovery probability from ICB, and they represent a concrete example of AI-derived tools entering clinical decision-making. AI can also standardize evaluations across institutions, reducing dependence on inter-rater-variable physician interpretation, which is a recognized source of inconsistency in immune scoring systems.
AI methods create systems that improve over time at specific tasks, including differentiating immunohistochemistry scores, classifying cancer subtypes, and identifying prognostic biomarkers. The convergence of AI and precision medicine is addressing several categories of complexity simultaneously: genomic variability between tumors, environmental influences on treatment response, and clinical comorbidity profiles that modulate treatment tolerance. A seminal demonstration from Capper et al. showed that AI-based whole-genome methylation analysis could classify 82 different brain cancer types with 93% accuracy using the Illumina HumanMethylation450 or Methylation EPIC arrays, outperforming pathologist accuracy and classifying over 70% of tumors that were unclassifiable by traditional methods.
Genomic considerations: Genome-informed prescription using machine learning represents one of the earliest and most mature applications of AI in precision medicine. ML algorithms trained to determine which patients are most likely to benefit from specific treatments based on genetic data can generate real-time treatment suggestions. Radiogenomics, a field that establishes connections between cancer imaging qualities and gene expression to predict individual toxicity risk from radiation treatment, has emerged directly from AI's success in image recognition tasks applied to cancer imaging.
Clinical prediction models: Researchers have analyzed 30 comorbidities using ML classifiers to determine which critical care patients would benefit from long-term tracheostomy and mechanical ventilation, demonstrating the breadth of AI's clinical applicability. For cancer-specific applications, AI models integrating clinical data and follow-up CT scans build predictive models that differentiate immunotherapy responders from non-responders based on patient age, sex, medical history, conventional lab tests, and imaging trajectories. Liquid biopsy approaches measuring circulating tumor cell DNA, serum complement levels (C1q and LDH), and cytokines enable real-time monitoring: a favorable correlation between declining circulating tumor DNA and improved overall survival has been established, and immunotherapy success has been linked to baseline serum C1q and LDH levels.
Organoid models: Tumor organoids, three-dimensional in vitro structures that recapitulate the TME and immune-tumor interactions, are emerging as powerful platforms for AI-assisted immunotherapy testing. AI in organoid analysis is expected to address the safety and personalization concerns of traditional prediction tools by enabling tumor culture, growth analysis, medication screening, and tissue collection in a controlled setting that closely mirrors in vivo biology.
AI applications in the design of CAR constructs and the discovery of novel anticancer drugs represent some of the most technically specific and rapidly evolving areas covered by this review. CAR construct design involves multiple interacting parameters: the extracellular antigen-binding domain (typically a single-chain variable fragment, scFv), the transmembrane region, the intracellular co-stimulatory domain (CD28 or 4-1BB), and the signaling domain (CD3-zeta). Each of these components influences CAR T-cell persistence, exhaustion resistance, cytokine secretion profile, and on-target off-tumor toxicity. Navigating this multi-dimensional design space computationally is far more tractable than iterative experimental screening.
CAR-Toner and AI-driven CAR design: Qiu et al. developed CAR-Toner, an AI-powered positively charged patch (PCP) computation tool that evaluates and scores CAR construct designs for antigen binding optimization, and also provides suggestions for improving the PCP score. While the precise effects of recommended modifications on CAR specificity and binding affinity still require experimental validation, CAR-Toner represents a concrete demonstration of AI advancing CAR T-cell engineering beyond what iterative bench experimentation can achieve at comparable speed.
TCR selection platforms: Bujak et al. proposed an AI-based platform for selecting specific T-cell receptors (TCRs) by building a database of pHLA (peptide-human leukocyte antigen) and TCR sequences from a multicenter observational trial enrolling 100 patients with stage II, III, or IV colon cancer adenocarcinoma from eight hospitals. DNA and RNA extraction was conducted on 86 samples, and 57 underwent RNA sequencing for gene expression profiling plus whole-exome sequencing for somatic mutation detection. The resulting pHLA:TCR database is intended to power an AI platform for TCR selection in precision cancer therapy, with the potential to significantly impact colorectal cancer management.
scFv screening by AI molecular dynamics: Martarelli et al. demonstrated that AI-assisted molecular dynamics and molecular docking analysis of anti-CD30 monoclonal antibody clones could identify the optimal scFv binding configuration before any CAR T-cell engineering was performed in the laboratory. Virtual computational scFv screening, surface plasmon resonance validation, and functional CAR T-cell tests showed concordant results for tumor binding and killing. The authors concluded that this in silico analysis approach could substantially cut expenses, time, and the requirement for laboratory animals in CAR construct development.
Anticancer drug susceptibility: Wang et al. developed an elastic net regression ML method to model drug susceptibility across multiple cancer types. In patients with gastric cancer (5-FU), ovarian cancer, and endometrial cancer (paclitaxel), ML algorithms predicted treatment susceptibility for patients with dismal prognoses. Deep learning-based screening predicted which patients might benefit from PARP inhibitor therapies by recognizing cancer cells with homologous recombination (HR) abnormalities at a 74% detection rate. These results underscore AI's potential to match patients to effective treatments at the molecular level rather than relying on histological type or stage alone.
AI algorithms trained on large clinical datasets have demonstrated improved diagnostic accuracy compared to medical professionals across a range of cancer imaging tasks. The FDA has approved several AI systems for cancer-related applications, including detection of suspicious lesions and interpretation of MRI and CT scans. AI algorithms now exist for cancer detection, tumor classification, treatment trend analysis, and dataset assessment across multiple cancer types, including breast cancer (identifying mammographic abnormalities) and lung cancer (determining pulmonary nodule malignancy risk).
CAR therapy diagnostics: Computer vision applied to CAR T-cell products analyzes the morphology and phenotype of CAR cells before infusion, enabling quality assessment that correlates product characteristics with clinical outcomes. Since the manufacturing variability of autologous CAR T-cell products is a recognized contributor to outcome heterogeneity, automated morphological and phenotypic profiling of cell products before infusion could identify batches with suboptimal effector characteristics before they are administered to patients.
Deep learning for tumor imaging in immunotherapy: Wang and Xu developed a semi-supervised DL approach to extract data from CT scans for predicting high-grade serous ovarian cancer recurrence. Effland et al. used a DL-based variation network for joint reconstruction of images and segmentation to characterize how immune cells communicate with melanoma cells at the tissue level, linking cellular spatial organization to immunotherapy response. A multiscale CNN approach was developed for volumetric segmentation of lung tumors, and deep residual learning was applied to predict mismatch repair (MMR) status from H&E-stained histology slides, a finding with direct implications for ICI patient selection given MMR-deficiency's established predictive value.
Watson for Oncology and clinical decision support: Watson for Oncology, a specific AI-assisted decision system, has demonstrated satisfactory concordance with choices made by multidisciplinary oncology teams, and could facilitate faster, less resource-intensive decision-making at the patient level, particularly in settings without access to full multidisciplinary tumor boards. Smartphone app-based AI tools (such as the "Diagnosis" app referenced by the authors) for annotating medical images and videos are also entering clinical use, reflecting the trend toward accessible, deployable AI tools beyond academic hospital settings.
Retrospective study designs and lack of prospective validation: Most AI research in cancer immunotherapy to date has been retrospective, and randomized controlled studies comparing treatment outcomes with and without an AI-based clinical decision support system are very rare. Few prospective investigations have been conducted despite significant advances in internal validation methodology, and this gap has eroded trust in the validity and applicability of published AI findings. Benchmark datasets representing diverse patient cohorts are not yet in widespread use, limiting algorithm generalizability across different demographics and healthcare systems.
Collider bias and geographic distribution: AI-based models have a biased geographic distribution, which limits their applicability to populations with different demographics and healthcare systems. Collider bias, which occurs when training and validation cohorts share a common selection variable not representative of the broader population, and inaccurate association production between features and outcomes both arise from characteristic distribution mismatches between training and validation datasets. This is a particularly acute concern for CAR therapy AI models given that most clinical trial data comes from major academic centers in North America, Europe, and East Asia, with limited representation of patients from diverse genetic backgrounds or healthcare infrastructure.
Data distribution shift and calibration drift: ML algorithms are contextual and time-sensitive: their performance changes as the patient population, clinical protocols, imaging hardware, or data collection practices at a deploying institution differ from those in the training dataset. Special-cause variations, defined as unexpected declines in model performance attributable to a change in the relationship between input features and target variables, require continuous monitoring of deployed AI systems for calibration drift. Ongoing AI monitoring and model updating using new data are prerequisites for sustained clinical utility, but few organizations have the infrastructure to do this continuously.
Data volume and annotation requirements: Convolutional neural networks and other DL architectures require massive annotated datasets that most institutions cannot assemble independently. High-quality standardized data across imaging, genomics, and clinical variables are in short supply for CAR therapy specifically. The Cancer Imaging Archive is one example of an open-access image database being used to address this problem, but it covers a narrow slice of the data types needed for comprehensive CAR therapy outcome modeling. An unsupervised DL method for chest CT segmentation demonstrated up to 98% accuracy, and a self-supervised lung disease classification approach on the NIH chest X-ray dataset showed strong performance, pointing toward unsupervised and self-supervised methods as pathways to reduce annotation dependence.
Reinforcement learning for CAR therapy dosing: One of the most forward-looking applications reviewed is the use of reinforcement learning (RL) to optimize the dose and schedule of CAR T-cell infusion. RL agents learn optimal decision policies through iterative reward-based feedback on outcomes, making them well-suited to the sequential, time-dependent nature of CAR T-cell therapy scheduling where infusion timing, lymphodepletion conditioning, and dose level all interact to determine cytokine release syndrome (CRS) severity and antitumor efficacy. This remains conceptual for CAR therapy but is under active development for drug dosing optimization in other contexts.
NLP for clinical data extraction: Natural language processing applied to clinical reports and literature offers the ability to extract relevant treatment decision information from unstructured text at scale. NLP can mine published case reports, trial results, and institutional clinical records to build the large, structured datasets needed for CAR therapy AI model training. Given that most CAR therapy outcome data resides in heterogeneous unstructured clinical records rather than structured databases, NLP is a critical enabling technology for the field.
AI governance frameworks: The review addresses AI governance directly, defining it as a system of rules, practices, processes, and technological tools ensuring that AI use aligns with organizational strategies, legal requirements, and ethical AI principles. As AI enters clinical practice in oncology and immunotherapy, governance frameworks must address algorithm transparency, accountability for AI-assisted decisions, bias auditing, and patient privacy. The current AI ecosystem in medicine is fragmented, with inadequate verification restricting clinical translation. Establishing shared governance standards comparable to Good Clinical Practice (GCP) for trials or TRIPOD reporting guidelines for prediction models is identified as a priority.
Multidisciplinary collaboration as a prerequisite: The authors are explicit that translating AI into routine CAR therapy practice requires multidisciplinary collaboration among researchers, clinicians, regulators, and patients. Radiologists, oncologists, engineers, and programmers must work together in bidirectional learning processes. Image and data harmonization, curation and annotation, preprocessing standardization, and legal/ethical frameworks must all be addressed collectively. The CHAIMELEON project is cited as an example of a methodology that uses continuous learning to update AI models incrementally with new descriptions and training data, allowing algorithm performance to improve progressively over the project timeline. This type of continuously learning, multi-institutional infrastructure is the architecture needed to bring CAR therapy AI from research setting to clinical deployment at scale.