Accurate patient selection is essential for the successful implementation of advanced therapies such as CAR-T cell treatment. AI improves patient selection by integrating and analyzing clinical data, such as demographic variables, laboratory results, imaging biomarkers, to identify complex, non-linear correlations that often go unnoticed by conventional statistical methods. By training on large volumes of data, AI models can learn patterns that link specific patient characteristics with therapeutic response or toxicity risk. This enables the stratification of patients based on response or clinical risk, supporting more personalized and evidence-based decision-making (107, 108).

In the pivotal JULIET trial, patients with relapsed or refractory diffuse large B-cell lymphoma (DLBCL) who received a single infusion of Tisagenlecleucel demonstrated a 53% overall response rate (ORR) and a 39% complete response rate (CR). However, significant challenges persisted, evidenced by a median progression-free survival (PFS) of only 2.9 months, with 62% of patients ultimately experiencing disease progression or death. To address these limitations, secondary analyses utilized AI to enhance predictive outcomes and patient stratification. Integrating pre-infusion FDG PET/CT imaging with clinicopathological data, AI models, particularly attention-gated AG-CNNs, were applied to delineate disease foci and improve outcome predictions. The AI analysis revealed that 52.4% of patients exhibited a negative AI signature, and all patients in this group experienced poor outcomes. In contrast, 47.6% had a positive AI signature, of whom 55.6% experienced poor outcomes (109). These findings highlight AI’s ability to classify patients into distinct prognostic groups, particularly those at high risk of poor outcomes.

Similarly, a study on 39 adult B-cell lymphoma patients treated with CD19 CAR-T cells used transfer learning(a deep learning technique) to analyze 770 lymph node lesions. Pre-treatment diagnostic CT (dCT), low-dose CT (lCT) from PET/CT, and FDG-PET images were assessed to predict responses at both lesion and patient levels. Lesion response was defined by size or metabolic activity reduction. Using AlexNet, a pre-trained CNN-MLP, researchers achieved 82% accuracy in lesion-level predictions, outperforming the International Prognostic Index (IPI) in sensitivity, specificity, and accuracy (110).

Beyond imaging-based response prediction, researchers have also applied AI for response prediction using RNA sequencing data. A study of 39 mantle cell lymphoma patient samples (30 sensitive, 9 resistant to CAR-T treatment) employed single-cell RNA sequencing and performed differential gene expression analysis to identify genes that were expressed differently between sensitive and resistant samples. A total of 1,236 different genes were identify. The differentially expressed genes were used as inputs for a MLP. The number of neurons in the input layer corresponds to the number of differentially expressed genes, while the output layer consists of two neurons representing the probability of the patient’s categorical response labels (sensitive or resistant). The model achieved an average accuracy of 90.07% in classifying cells based on the genes expression as sensitive or resistant to therapy (111).

Collectively, these studies highlight AI’s role in improving patient selection and outcome prediction for CAR-T therapy compared to traditional methods and opens new avenues for personalized treatment strategies.

CAR-T cell production is complex and demands strict quality control to ensure safety and efficacy. As the demand for cell and gene therapies increases, automating manufacturing processes will become critical to achieve enhanced standardization. Innovations that eliminate intermediary steps, such as T cell separation using magnetic beads or nanomatrices coated with anti-CD3 and anti-CD28 antibodies (112), can substantially reduce production timelines, resource usage, and contamination risks inherent to manufacturing processes (113). Traditional flow cytometry (FCM) techniques also play an essential role in assessing cell populations; however, conventional methods involving fluorescent staining with antibodies and chemical reagents may compromise cell function, introduce toxicity, and reduce viability (114). Additionally, meeting Good Manufacturing Practice (GMP) standards necessitates specialized personnel, further increasing costs and operational variability (115, 116). These factors hinder the efficient scale-up of CAR-T cell production.

AI-driven sorting systems like ghost cytometry (GC), introduced by K. Sugimoto et al., offer a label-free alternative to antibody-based cell sorting. With this tool developed in 2020, cells are illuminated during imaging to produce distinct optical signatures and waveforms, reflecting various aspects of cell morphology and behavior. ML analyzes these aspects, enabling high-speed sorting when integrated with microfluidics. By training ML classifier with temporal and frequency-domain features extracted from the specific waveforms generated by each cell, allowing the capture of relevant morphological information, GC accurately distinguishes live from dead cells and identifies CD3-positive T cells in peripheral blood mononuclear cells (PBMC) populations (117). By 2024, label-free GC (LF-GC) further improved precision, differentiating live vs. apoptotic cells, T cells vs. non-T cells, and detecting particulate impurities. These advancements enhance real-time cell monitoring and quality control in cell therapy manufacturing without the need for surface markers or stains (118).

Other approaches like COSMOS (Computational Sorting and Mapping of Single Cells), combines high-resolution imaging, DL, and microfluidics for real-time, label-free cell sorting. Using brightfield microscopy, it captures cellular images and analyzes size, shape, and structures. Using deep learning with CNNs and MLPs, it classifies cells by mapping their features into a representation space, enabling accurate and real-time identification (119). Other AI-enhanced systems like DeepCell integrate high-content imaging with neural networks to predict cell identity and function in real time. These platforms enable morphological profiling without staining and support cloud-connected analytics for remote QC, offering potential scalability for GMP-compliant CAR-T production (120).

Computer vision plays a fundamental role in enhancing these image-based sorting technologies by enabling the automated extraction of detailed morphological and structural features from high-resolution cellular images. This field within deep learning allows for accurate, real-time classification without the need for chemical staining, significantly improving throughput, consistency, and viability in CAR-T cell manufacturing. This integration accelerates decision-making and ensures greater standardization in cell quality assessment (121).

A key goal is engineering T-cell phenotypes that enhance cytotoxicity while maintaining a stem-like state for long-term persistence, which its traditionally made on trial-and-error methods (110). Neural networks can expedite CAR development by analyzing large datasets that integrate receptor structures with therapeutic outcomes.

One study constructed a library of costimulatory domains by recombining 13 signaling motifs, generating 2,379 unique combinations with distinct phenotypic traits. For instance, the 4-1BB domain enhances persistence, while CD28 boosts cytotoxicity but reduces longevity (122). ML models analyzed these motif interactions and generated diverse T-cell phenotypes, including separate cytotoxicity and stemness traits not typically observed with natural domains. This predictive approach significantly expands exploration of receptor combinations, potentially advancing the development of precise and potent cellular therapies (123).

Costimulatory domains also influence IS quality, directly affecting CAR-T activation and cytotoxicity. Unlike conventional T cells, CAR-T cells form disorganized IS structures, initiating rapid signaling via Lck/ZAP70 and exhibiting “serial killer” activity (17, 55, 56, 124–126). Traditional prediction methods, such as flow cytometry and in vivo models, are resource-intensive and inconsistent (127). In response to these challenges, researchers developed a DL tool that improves IS analysis using clinical trial data. They employed a segmentation neural network to accurately delineate kappa-CAR-T cells within the images, followed by a classification network that predicted therapeutic response based on the morphological analysis of the immunological synapse. This approach proved effective even in low-contrast images, accurately distinguishing between treatment responders and non-responders (128). Other approaches use the images obtained by optical diffraction tomography (ODT) combined with DL, which allows automated 3D IS tracking by segmentation and classification, further refining CAR-T assessments (129).

Beyond IS dynamics, tonic signaling is crucial for CAR-T function. This baseline signaling, independent of antigen engagement, must be carefully regulated—too little weakens persistence, while excessive signaling induces exhaustion (130–132). CAR antigen-binding domains with positive charge patches (PCPs) promote clustering, influencing tonic signaling (130). AI-driven tools like CAR-Toner leverage structural data from (Protein Data Bank)PDB and AlphaFold to optimize PCP scores, improving CAR-T expansion and reducing exhaustion. This method was validated in camelid single-domain nanobody (VHH)-based CARs targeting CLL1 in acute myeloid leukemia, identifying an optimal PCP score range (43–52, 255) that correlates with superior CAR-T function (133).

CAR-T therapies represent a breakthrough in cancer treatment but carry risks such as Cytokine Release Syndrome (CRS), Immune Effector Cell-Associated Neurotoxicity Syndrome (ICANS), and On-Target, Off-Tumor Toxicity (OTOT).These toxicities significantly impact clinical outcomes, requiring effective management strategies and early detection of this complications often rely on clinical observation, scoring and are often subjective and time-intensive (134, 135). CRS, the most common complication, affects 30%–100% of patients, with severe cases (grade ≥3) occurring in 10%–30% (136, 137). It results from excessive cytokine release and endothelial activation, leading to increased vascular permeability, hypoperfusion, coagulopathy, and potential multi-organ dysfunction (138–141).

To enhance early detection, ML models like M2-CRS have been developed. This meta-analysis-informed ML approach integrates statistical data from clinical studies with ten predictive cytokine biomarkers, including IL-2, IL-6, IFN-γ, and GM-CSF. By leveraging a robust Knowledge Base (KB) of cytokine interactions, M2-CRS addresses data scarcity while maintaining high interpretability. The model, based on Support Vector Machines, consistently achieves accuracy and precision above 90%, making it a promising tool for CRS risk stratification and timely intervention. To further enhance interpretability, SHapley Additive exPlanations(SHAP) values were applied, allowing the identification of the most influential features within the model. This analysis revealed IFN-γ and IL-10 as the most relevant proteins contributing to CRS risk prediction in acute myeloid leukemia and B cell Lymphomas (142).

ICANS, another significant CAR-T toxicity, affects 30%–50% of patients, with severe cases in 12%–30% (143). It arises from cerebral endothelial activation, blood-brain barrier disruption, and cytokine-induced astrocyte and pericyte dysfunction, leading to cerebral edema, thrombosis, hemolysis, and disseminated intravascular coagulation (DIC). Logistic regression model can analyze longitudinal patient data faster, predicting ICANS onset and severity with accuracy of 77%. By identifying key risk factors such as age, fever, IL-6 levels, and procalcitonin, ML can forecast ICANS progression up to three days in advance. This predictive capability enhances clinical decision-making, optimizing resource allocation, hospitalization planning, and early therapeutic interventions (144).

CAR-T cell and nanoparticle production are interdependent biotechnological processes, requiring precise control for consistency and efficacy. Nanoparticles are essential for mRNA-based CAR expression, facilitating gene delivery to T cells, but large-scale manufacturing faces challenges in maintaining size, charge, and composition. Batch variability affects cellular uptake and gene transfer, making precise formulation critical (145, 146). Similarly, CAR-T cell production demands careful monitoring of viability and expansion in bioreactors.

Integrating real-time monitoring with ML can optimize synthesis parameters, reducing variability and enhancing scalability in both nanoparticle formulation and CAR-T manufacturing (147, 148). Lipid nanoparticles (LNPs) are currently the most widely utilized carriers for mRNA delivery, comprising ionizable lipids, phospholipids, cholesterol, and PEG-lipids (149–151). Ionizable lipids facilitate protein expression, while nanoparticle size and charge critically influence cellular uptake, immune responses, and delivery efficiency; therefore is essential to ensure consistent and effective large-scale production (152–155). XGBoost is a gradient boosting algorithm based on decision trees, known for its high predictive accuracy and computational efficiency in structured data. It builds models iteratively, adding new trees to correct previous errors while balancing accuracy and complexity through regularization. In the context of nanoparticle production, XGBoost has demonstrated significant effectiveness in optimizing nanoparticle production. Sato et al. used XGBoost to predict key parameters in the manufacturing of lipid LNPs, such as particle size, polydispersity index (encapsulation efficiency (EE Independent regression models were built using process variables such as total flow rate, pH, and concentrations of lipids, ethanol, mRNA, and buffer. The models demonstrated high accuracy, with correlation coefficients of 0.998 for particle size, 0.990 for PdI, and 0.977 for EE (156).

Other notable advancement is presented by Amirreza Mottafegh et al. Introducing an autonomous platform for rapidly synthesizing drug-loaded nanoparticles, such as liposomes and polymeric nanoparticles. Their system features an Automated Nanoparticle Synthesizer (ANS) using nanoprecipitation to precisely control flow rates and reactant concentrations. It also incorporates a ML-driven multi-objective Bayesian optimization (MOBO) algorithm, which dynamically adjusts parameters based on real-time feedback. This innovation reduces production time from hours or days to just 20 minutes while ensuring high reproducibility and quality (157). AI-driven optimization in nanoparticle manufacturing can enhance CAR-T therapy by boosting initial production and improving efficiency.

Our proposed AI-revolution could also extend to cell manufacturing with smart bioreactors equipped with nanosensors enabling precise, real-time monitoring and adaptive control of culture conditions. These advanced systems integrate thin-film nanosensors—such as iridium oxide for pH, platinum-based for oxygen, and graphene-coated for glucose—capable of detecting even subtle environmental changes (158–161). For example, a Pd/Pt-functionalized glucose sensor on graphene ensures continuous nutrient monitoring, preventing both starvation and overfeeding. When paired with wireless electronics, these sensors provide continuous feedback, allowing automated adjustments to optimize culture conditions, reduce costs, and enhance scalability without compromising quality (160). AI analytics further enhance this approach by enabling the interpretation of complex, high-dimensional nanosensor data. In the work by Herpin et al., MLP network was employed to disentangle overlapping infrared spectral signals from proteins, lipids, nucleotides, and peptides, enabling real-time monitoring of dynamic molecular interactions with high accuracy (162, 163). Similarly, Leong et al. demonstrated the use of CNNs to analyze multiplexed Surface-Enhanced Raman Scattering (SERS) spectra, achieving over 86.8% classification accuracy in distinguishing small-molecule metabolites such as ATP, glucose, and lactate in complex biological environments. These deep learning models transform intricate biochemical signatures into interpretable and actionable insights, paving the way for real-time diagnostics and precise metabolite monitoring at the point of need (164).

In bioreactors, AI models analyze cell growth, nutrient levels, and environmental fluctuations, dynamically adjusting conditions to optimize expansion. For example, in the work by Grzesik et al. regression models (including Elastic Net and Random Forest) were trained on donor-specific data to predict T cell viability and expansion with high accuracy (R² > 0.92, RMSE < 1.5). These models enabled the in-silico design of optimized culture media, which were experimentally validated and outperformed traditional regression approaches, illustrating how machine learning can streamline media optimization and enhance bioprocess robustness (165, 166).

Beyond manufacturing, it’s important to look at the molecular level to integrate innovative approaches aimed at enhancing CAR-T therapy. Initially, it is key to optimize mRNA delivery for CAR-T therapy success. Within the acidic environment of endosomes, lipid cores become charged, facilitating endosomal escape—a crucial step for successful nucleic acid delivery (167, 168). However, not all nanoparticle formulations perform equally well. Some formulations may accumulate within sub-endosomal membranes, impeding efficient release and consequently reducing transfection efficiency (169). While nanoparticles show great potential, optimizing their performance through conventional trial-and-error methods remains challenging (170).

DL is a powerful tool for overcoming these bottlenecks, enabling the analysis of molecular patterns, deciphering chemical structures, and predicting the behavior of novel substances. Computational models can simulate cationic polymer behavior within endosomes, facilitating the design of more efficient gene delivery systems (171). This significantly reduces the time and cost involved in developing new platforms. AI-driven models also allow detailed study of ion channel physiology and genomics, supporting the design of systems that promote enhanced endosomal and lysosomal escape (96, 172–174). AI can also evaluate critical factors like cargo fusion rates after internalization, optimizing vectors for better performance (26, 175). In this context, multimodal neural networks (which integrate heterogeneous data sources such as chemical structure, physicochemical descriptors, biological response profiles, and imaging data) are emerging as advanced tools for optimizing nanoparticle design. By jointly analyzing multiple modalities, these models offer a more holistic understanding of how structural and functional features interact, allowing for more precise prediction of mRNA transfection outcomes and nanoparticle behavior across complex biological environments (176).

A prominent example of AI’s potential in gene delivery is the AGILE platform, developed by Yue Xu and colleagues. AGILE uses DL, combining combinatorial chemistry with high-throughput screening (HTS), to rapidly discover ionizable lipids optimized for mRNA delivery. At the core of AGILE is a multimodal deep learning model that integrates two complementary inputs: molecular graphs and calculated descriptors. The graph encoder, based on graph neural networks (GNNs), encodes the structural topology of lipid molecules, while a molecular descriptor encoder processes physicochemical features derived from each compound. Initially, the GNN is pre-trained in a self-supervised manner on a vast dataset of small molecules to capture generalizable structural patterns. It is then fine-tuned with experimental data from 1.200 synthesized lipids formulated into lipid nanoparticles (LNPs) and assessed for mRNA transfection potency (mTP) in specific cell lines. This refined, multimodal model predicts the mTP of new lipid candidates, enabling the prioritization and selection of high-performing structures for experimental validation. AGILE identified H9 and R6, ionizable lipids optimized for mRNA delivery to muscle cells and macrophages, respectively. H9 outperformed industry-standard lipids like ALC-0315 (used in Pfizer’s COVID-19 vaccine) in muscle tissue, minimizing liver off-target effects, while R6 exhibited superior transfection efficiency in macrophages (175). By rapidly screening extensive lipid libraries, AI-driven platforms like AGILE can optimize ionizable lipids for T-cell mRNA transfection, improving CAR expression, T-cell expansion, and addressing critical factors for broader clinical adoption.

Optimizing nanoparticle formulations is essential for effective mRNA delivery; however, identifying optimal targets for immune recognition is equally crucial for successful CAR-T therapies. Neoantigens—tumor-specific antigens generated by cancer cell mutations—are fundamental to CAR-T therapy design. Integrating AI into this process can boost neoantigen discovery, significantly improving therapeutic precision and safety. Traditional neoantigen prediction tools primarily assess peptide-MHC binding strengths to identify potential epitopes. While mouse models utilize in vivo and ex vivo methods to assess immunogenicity, equivalent approaches for humans remain unavailable (177, 178). Additionally, predicted neoantigens often exceed experimentally validated immunogenic peptides, highlighting the need for advanced computational tools capable of accurately predicting neoantigen immunogenicity (179, 180). Here, bioinformatics plays a key role by enabling the integration and interpretation of multi-omic data (such as transcriptomic, proteomic, and genomic profiles) to uncover biologically relevant patterns. When enhanced with AI, especially deep learning, these bioinformatic pipelines become powerful tools capable of modeling complex immune interactions, improving the prediction of truly immunogenic neoantigens and accelerating the design of personalized immunotherapies (181).

A study made by Fabiana Perna et al. demonstrated computational discovery of potential antigens by identifying acute myeloid leukemia (AML) targets—ADGRE2, CCR1, CD70, and LILRB2—via surface proteomics and transcriptomics. These antigens showed high AML cell expression with minimal presence in normal tissues (182). AI-based methods could enhance such processes further by rapidly analyzing data and detecting complex protein expression patterns, significantly improving prediction accuracy and speed (183, 184). ML algorithms like Random Forest and XGBoost effectively predict immunogenic peptide responses. In one study, researchers trained models on a dataset of immunogenic and non-immunogenic peptides, analyzing characteristics like hydrophobicity, peptide size, and amino acid preferences. This allowed the models to predict peptide immunogenicity efficiently, accelerating antigen discovery for immunotherapy (184).

Similarly, the DL-based tool MHCnuggets employs Long Short-Term Memory (LSTM) neural networks to predict peptide binding to MHC class I and II molecules. Its adaptable architecture handles variable-length peptides and integrates both binding affinity measurements and mass spectrometry-derived data, substantially improving prediction performance. The model supports both regression, estimating continuous IC50 binding affinity values, and classification, identifying binders versus non-binders from immunopeptidomic (HLAp) datasets. MHCnuggets rapidly analyzes millions of peptide-MHC interactions, identifying immunogenic mutations across various cancer types. Although CAR-T therapies do not directly rely on MHC molecules, adapting similar approaches could help identify highly specific surface antigens, enriching overall immune responses and enhancing therapy effectiveness (185). Likewise, nanotechnology addresses challenges in neoantigen stability and immunogenicity, improving antigen spreading and delivery. Zhao et al. Showed myeloperoxidase nanovaccines capable of inducing immunogenic cell death (ICD), triggering an immune response cascade, enhancing neoantigen delivery to lymph nodes, and promoting a pro-inflammatory tumor microenvironment. Integrating AI tools like MHCnuggets could optimize this process further by rapidly prioritizing immunogenic neoantigens for personalized cancer immunotherapy (186–188).

It’simportant to mention that, besides the discovery of neoantigens, the quest for highly specific and effective targeting mechanisms leads us to explore alternative binders: Nbs present a promising alternative to traditional scFvs in CAR-T therapy. Due to their smaller size, enhanced stability, reduced immunogenicity, and ability to target unique epitopes, nanobody-based CAR-T cells hold significant potential to improve cancer treatments (76). CAR-Toner has demonstrated that nanobodies-based CARs exhibit optimal positive charge patch (PCP) scores, correlating with improved CAR-T cell functionality (133). Accelerating the discovery of high-affinity nanobodies enables more precise and effective therapeutic interventions. Artificial intelligence further enhances this process, notably through the ML-driven platform NbAffinity, which predicts nanobody-target binding affinity by analyzing critical molecular interactions such as hydrogen bonding, aromatic stacking, and ionic interactions. NbAffinity integrates advanced algorithms—including Random Forest (RF), Rotation Forest (RotF), and Support Vector Machines (SVMs)—to efficiently analyze extensive structural data. Trained on a comprehensive dataset of 991 nanobody-ligand complexes supplemented by 444 protein-protein interaction pairs, NbAffinity achieves robust, generalizable predictions.On the test set, SVM reached an accuracy of 75.87%, reflecting a strong ability to classify affinitive versus non-affinitive nanobodies based on structural interaction features (189).

Building on the precision offered by AI-driven nanobody discovery, it becomes clear that understanding and enhancing dynamic interactions within the tumor microenvironment is equally crucial for advancing CAR-T therapies. Predicting real-time interactions between CAR-T cells, nanoparticles, and tumor cells remains a complex challenge. Emerging technologies like digital twins and virtual patient models are poised to improve cancer treatment by leveraging computational methodologies, including quantitative systems pharmacology (QSP) and physiologically based pharmacokinetics (PBPK), to simulate complex biological processes (190–194). Digital twins provide real-time, personalized tumor replicas, capturing distinct biological and physical attributes. In contrast, virtual patient models incorporate population-level variability, offering valuable insights into diverse subgroup responses (195). These models integrate multi-omics data, predicting tumor dynamics, immune interactions, and therapeutic outcomes, thus enhancing targeting accuracy and minimizing off-target effects (196–198). Such approaches significantly improve simulation of interactions within the CAR-T cell-nanoparticle-tumor axis, optimizing tumor targeting and delivery efficiency (199–202).

Integrating AI further amplifies these capabilities by analyzing tumor morphology, metabolism, and pharmacokinetic profiles (200). This data-driven approach refines tumor microenvironment simulations, optimizing CAR-T cell and nanoparticle design and administration. Platforms such as TumorScopePredict illustrate practical applications by generating dynamic 3D tumor models and accurately predicting therapeutic responses from pre-treatment imaging data processed via CNNs, simulating tumor behavior, drug sensitivity, and resistance (203, 204). The adaptive nature of AI ensures these models remain at the forefront of cancer research, continuously evolving with new data (205, 206).

One significant challenge for CAR-T therapy in solid tumors is overcoming physical and immunosuppressive barriers within the tumor microenvironment (TME). Unlike hematologic malignancies, solid tumors present dense extracellular matrices (ECM) and aberrant vasculature, impeding T-cell migration. Additionally, immunosuppressive cells such as tumor-associated macrophages (TAMs), regulatory T cells (Tregs), and myeloid-derived suppressor cells (MDSCs) create a hostile environment, limiting T-cell effectiveness through cytokine secretion and checkpoint inhibition (207–213).

Nanotechnology addresses these challenges by remodeling the TME, degrading ECM, and reprogramming immunosuppressive cells like TAMs. It also facilitates targeted delivery of supportive cytokines (e.g., IL-12, IL-15) and immune checkpoint inhibitors (e.g., anti-PD-1) to enhance CAR-T cell performance. Additionally, nanoparticles address antigen escape via multi-antigen targeting and epitope spreading (17, 64, 84, 84, 214, 215). However, clinical translation remains limited due to poor delivery efficiency (DE), with only ~0.67% of nanoparticles reaching the tumor, while the remainder accumulates off-target, primarily in the liver, spleen, and lungs (216–221). This challenge highlights the necessity for improved nanoparticle design to enhance tumor targeting while minimizing systemic distribution.

AI-driven approaches can significantly enhance nanoparticle design, optimizing physicochemical properties to improve tumor targeting and reduce systemic distribution. A DNN utilizing the Nano-Tumor Database (534 tumor datasets, 1972 tissue datasets) accurately predicted nanoparticle DE in tumors and key organs based on nanoparticle properties and therapy strategies. This AI-driven approach identified core materials as critical determinants of DE and provided a web-based tool, Nano-AI-QSAR, facilitating nanoparticle optimization (222). Other organ-specific studies have used AI to optimized brain-targeted nanoparticle delivery, identifying influential factors such as release rate and molecular weight, enhancing nanoparticle effectiveness (223). Additionally, AI-driven optimization of mRNA nanovaccines demonstrated significant improvements in transfection efficiency and lymphatic delivery, successfully activating the stimulator of interferon genes (STING) pathway and boosting antitumor immunity in vivo (224). Given the high dimensionality, heterogeneity, and non-linear relationships within biological and physicochemical data, deep learning models are increasingly favored for their capacity to capture complex patterns and interactions that traditional models may overlook (225).

Based on the foundation of AI-optimized nanoparticle design for improved delivery efficiency (DE), an exciting frontier lies in magnetically guided nanoparticles, a strategy ripe for further refinement through AI-driven models. Certain nanoparticles possess magnetic properties that facilitate their targeted redirection to tumor sites—a strategy that can be refined through AI-driven models. Yasmeen Akhtar et al. developed computational simulations integrating factors like magnetic field strength, blood flow, and chemical dynamics to optimize delivery efficiency (DE) and minimize off-target accumulation. Pulsatile flow experiments revealed that stronger magnetic fields enhance nanoparticle retention, while pulsatile conditions improve tumor penetration (62, 226).

Such AI-driven magnetic guidance methodologies hold considerable promise for CAR-T therapy, especially since magnetic control has already demonstrated potential for directing T-cell migration. This suggests substantial opportunities to further apply AI techniques to optimize these processes. Innovative strategies, such as CAR-T-cell-based microrobots (M-CAR Ts), which incorporate immunomagnetic beads onto CAR-T cells, facilitate precise targeting and deep tumor infiltration through sequential magnetic-acoustic actuation, markedly increasing T-cell accumulation and therapeutic efficacy (227). Similarly, CAR-T cells functionalized with superparamagnetic iron oxide nanoparticles (SPIONs) can be magnetically guided to tumor sites, enhancing localized cytolytic activity while simultaneously reducing systemic cytokine release and associated toxicities (228). Integrating AI-driven predictive modeling with advanced nanoparticle engineering can actualize the therapeutic potential of nanotechnology for solid tumors. AI can overcome critical barriers in CAR-T therapy by optimizing nanoparticle properties for superior delivery efficiency, precision targeting, and incorporating complementary approaches such as magnetic guidance to achieve robust and targeted therapeutic outcomes.

Now that the key components for precise and efficient targeted delivery are in place—including AI-optimized nanoparticles and advanced TME-modulating strategies—the prospect of in vivo CAR-T cell generation could be within reach. Current CAR T-cell production involves leukapheresis to isolate patient T cells, activation with cytokines and co-stimulatory molecules, genetic modification via lentiviral or γ-retroviral vectors to introduce CAR genes, and subsequent cell expansion into therapeutic doses. This complex and costly procedure requires specialized facilities and trained personnel (14, 229). In contrast, the proposed in vivo CAR T-cell production, involving direct in-patient delivery of CAR genes and activation signals, simplifies manufacturing, reduces costs, and enhances accessibility. Preclinical studies in mouse models have shown comparable antitumor efficacy between in vivo-produced and ex vivo-engineered CAR-T cells (53, 54, 230).

Successful in vivo CAR-T cell therapy must meet several critical criteria: high gene-editing efficiency, precise T cell targeting, the ability to overcome solid tumor barriers, sustained functionality and persistence of CAR-T cells, and minimal toxicity (51, 231). CAR T cells face major barriers in the tumor microenvironment (TME), including physical, chemical, and immunological obstacles. To overcome these, nanoparticles are being developed not just for drug delivery but also to modulate the TME and create a “theranostic window” where tumors are both treatable and detectable. These particles (10–200 nm) accumulate in tumors via the EPR effect and can be functionalized with ligands targeting tumor or stromal cells. They can also carry enzymes like collagenase to degrade the ECM, enhancing T cell infiltration and nanoparticle penetration. Additionally, they can deliver siRNA, miRNA, or inhibitors (e.g., TGF-β, IDO, checkpoint blockers) to reprogram the immunosuppressive TME. The Smart CAR-T Nanosymbionts interface exemplifies this advancement, integrating functionalities for precise targeting, enhanced persistence, and minimized toxicity by leveraging the advanced technologies discussed in this paper. Furthermore, AI-driven platforms such as AGILE are being employed to optimize the design of ionizable lipids, significantly enhancing the transfection efficiency of gene delivery systems. This ensures robust CAR expression in T cells, further supporting the development of effective in vivo CAR-T therapies (175).

Additionally, AI-guided nanoparticle design improves tumor-specific targeting, minimizing off-target accumulation, facilitating extracellular matrix degradation for improved tumor infiltration, delivering targeted immunomodulatory agents (e.g., IL-12, anti-PD-1), and enabling efficient mRNA delivery into T cells (222). Advanced targeting strategies like magnetically guided M-CAR Ts and SPION-functionalized CAR-T cells enhance tumor localization and deep tissue penetration as necessary (228). Maintaining CAR-T cell functionality involves localized cytokine delivery facilitated by nanotechnology, complemented by AI-driven optimization tools such as CAR-Toner, which refine tonic signaling, and computational approaches optimizing co-stimulatory domains (CD28, 4-1BB) to balance T-cell activation and memory formation directly within the patient (55–60). Finally, addressing potential toxicity necessitates a comprehensive strategy incorporating precise targeting, tumor-specific neoantigen prediction, and controlled-release systems for cytokines and checkpoint inhibitors, collectively ensuring both safety and therapeutic efficacy (232–234). Figure 4 represents the congregation of these technologies in the “In vivo Smart CAR-T Nanosymbionts manufacturing”.

AI-driven image analysis has the potential to revolutionize cancer imaging by shifting from qualitative interpretation to objective, quantitative evaluations. This transformation facilitates earlier detection, precise lesion characterization, and improved monitoring of disease progression and treatment responses (235). In CAR-T therapy specifically, multimodal network models integrating imaging and clinical data have successfully predicted treatment outcomes (109, 110).

Iron and gold nanoparticles have emerged as powerful contrast agents for magnetic resonance (MR), computed tomography (CT), and positron emission tomography (PET) imaging, significantly enhancing the resolution and visibility of both solid tumors and hematological malignancies. These advancements not only improve diagnostic accuracy but also reduce imaging duration and resource utilization (236–239). When combined with AI-driven innovations in biodistribution and targeted delivery, nanoparticles further amplify tumor contrast, enabling precise lesion delineation and superior assessment of therapeutic responses. CNNs trained on nanoparticle-enhanced imaging data refine tumor characterization and support real-time clinical decision-making, fostering automated and precise cancer monitoring (240). Additionally, in vivo tracking of CAR-T cell distribution with contrast agents enhances persistence and tumor targeting. This enables real-time response assessment and better lesion characterization (228). The previously mentioned CNN-based approach could further improve non-invasive CAR-T tracking methods, such as ferumoxytol-enhanced MRI for iron oxide- or gold-labeled CAR-T cells, optimizing the evaluation of therapy distribution and effectiveness (20, 21, 88).

Besides imaging monitoring, tracking CAR-T cells in peripheral blood is critical for assessing therapy efficacy and ensuring patient safety, particularly in blood malignancies like acute leukemia (241). Traditional identification methods are limited due to the similarities between CAR-T cells and other immune cells. However, the RCMNet model, which integrates CNN-MLP trained on Peripheral Blood Cells (PBC) datasets, achieves a top-1 accuracy of 99.63%, revolutionizing CAR-T cell identification in blood samples. This approach alleviates manual evaluation burdens, provides real-time automated analysis, and markedly improves diagnostic accuracy (242). Transformers, originally developed for natural language processing, rely on self-attention mechanisms that allow the model to weigh the importance of different input features relative to one another (243). In this context, they enhance the model’s ability to focus on subtle, context-dependent cellular traits that differentiate CAR-T cells from other immune populations.

While monitoring therapeutic responses is crucial, it is equally vital to detect adverse effects such as cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS) at an early stage. Nanoelectronics biosensors, such as silicon nanowire field-effect transistors (SiNW FETs), provide highly sensitive, real-time detection capabilities that surpass traditional methods like ELISA (244, 245). These nanosensors enable precise tracking of cytokines and toxicity markers, facilitating rapid clinical intervention—all of which can be further optimized and enhanced through AI-driven models (246). The detailed data generated by these biosensors can strengthen AI-driven predictive models, such as M2-CRS, facilitating early toxicity predictions and personalized treatment management. Additionally, machine learning algorithms, particularly Random Forest (RF), can integrate nanoparticle physicochemical properties—such as size, shape, surface coating, and zeta potential—along with exposure conditions like dose, duration, and tissue type to accurately predict nanoparticle-induced toxicity. Such predictive models improve safety assessments, enabling proactive risk management and tailored therapeutic adjustments (247).