CD33, also known as sialic acid-binding immunoglobulin-like lectin 3, is a differentiation-associated antigen predominantly expressed on myeloid progenitor cells and broadly distributed across various stages of myeloid lineage development (7). In AML, CD33 is highly expressed on leukemic progenitor cells, whereas its expression on normal hematopoietic stem and progenitor cells (HSPCs) is relatively low (8). This differential expression profile renders CD33 a compelling target for AML-directed immunotherapy.

Among CD33-targeted strategies, ADCs have achieved notable clinical progress. In 2017, the U.S. Food and Drug Administration approved gemtuzumab ozogamicin (GO) for the treatment of CD33⁺ AML, marking the first approved ADC for AML and representing a major milestone in precision medicine for hematologic malignancies (9). The clinical success of GO has verified the potential of CD33 as an effective therapeutic target and has also promoted the development of subsequent cell-based immunotherapies, including CAR-T cell therapy.

As CD33-directed CAR-T cells advance into clinical evaluation, a primary research objective is to enhance antileukemic efficacy while minimizing on-target, off-tumor toxicity. To improve therapeutic safety, researchers have developed CD33-targeted CAR-engineered natural killer (NK) cells, which have shown potent cytotoxicity against AML cells both in vitro and in vivo, with limited cytotoxic effects on normal HSPCs in phase I trials (10). In parallel, ‘controllable CAR systems’ have emerged as a promising safety-enhancing strategy. By integrating inducible safety switches such as inducible caspase-9 (iCasp9) into the CAR construct, these systems enable the selective elimination of CAR-T cells upon administration of small-molecule dimerizers such as AP1903, thereby mitigating prolonged myelosuppression and reducing the risk of severe adverse events (11).

CD33-targeted CAR-T therapy has now entered multiple phase I/II clinical trials. In the dose-escalation clinical trial, NCT03971799, conducted in adolescents and young adults with R/R AML, cytokine release syndrome (CRS), an acute systemic inflammatory toxicity caused by CAR-T-cell activation and cytokine release, occurred in 68% of patients, with grade ≥3 CRS reported in 21% (12). This highlights that, despite promising antileukemic activity, CD33-directed CAR-T therapy in AML may be associated with substantial treatment-related immune toxicity. However, cross-trial comparison of toxicity rates should be approached with caution due to differences in study population, dose level and trial design as these may substantially affect the observed safety profile. Despite these toxicities, promising efficacy was observed: In the highest dose level cohort, 40% of patients achieved complete remission (CR) with minimal residual disease (MRD) negativity (12) (Table I). To further improve the safety profile, the use of donor-derived CD33 CAR-T cells (such as VCAR33) in the post-hematopoietic stem cell transplantation (HSCT) setting has been explored. In a phase I/II trial, 41.7% (5/12) of patients achieved bone marrow CR, with four attaining MRD negativity. Additionally, patients with central nervous system (CNS) leukemia also showed MRD clearance following CAR-T infusion. The treatment was generally well tolerated, with most patients experiencing only grade 1–2 CRS and no cases of immune effector cell-associated neurotoxicity syndrome (ICANS). Some patients developed transient hepatic dysfunction and infections, which were manageable with supportive care. As of June 2024, six patients had either undergone a second HSCT or remained in remission following CD33 CAR-T therapy (13).

CLEC12A, also known as C-type lectin-like molecule-1 (CLL-1), is a type II transmembrane glycoprotein that is highly expressed on leukemia precursor cells and leukemia stem cells (LSCs) of AML, while its expression is extremely low on normal HSPCs, making it a potential target for AML treatment (14). This differential expression of CLL-1 confers high specificity for targeted therapy and reduces the risk of off-target effects. Moreover, the expression of CLL-1 remains relatively stable during disease progression, suggesting it has value in the application of MRD monitoring and has potential as a reliable prognostic biomarker (15). CLL-1 is still present in AML subtypes with low or absent expression of CD33/CD34 (16), which expands the treatment coverage and overcomes the limitations associated with traditional single antigen targets. Multicenter clinical trials have suggested that the CLL-1 targeted CAR-T cell therapy can achieve encouraging therapeutic effects. In a clinical trial of pediatric patients with R/R AML, the CR rate was >50%, and the safety profile appeared favorable. Specifically, all patients developed only grade 1–2 cytokine release syndrome (CRS) and no lethal events were reported (17).

The preliminary results of early clinical trials (NCT04219163 and ChiCTR1900027684) further corroborate the efficacy and safety of the CLL-1 targeted CAR-T therapy in pediatric and adult patients with AML, with the highest CR rate reaching 81.8%, and only 5% of patients experienced ≥3 grade CRS (17,18) (Table II). Nevertheless, these findings should be interpreted within the context of the individual study settings, as differences in patient population, disease status and trial design may limit direct comparison with other CAR-T studies.

CD7 is a transmembrane glycoprotein belonging to the immunoglobulin superfamily, primarily expressed on T cells and NK cells. In AML, 20–35% of patients display ectopic expression of CD7, particularly among those with cytogenetically high-risk subtypes (19), CD7⁺ AML is frequently associated with resistance to chemotherapy, increased disease aggressiveness and worse clinical outcomes (20). Therefore, despite its relatively limited expression prevalence, CD7 represents a compelling therapeutic target for selected AML subgroups.

CD7-targeted CAR-T cells have demonstrated potent, antigen-specific cytotoxicity against CD7⁺ AML cells in preclinical studies (21). However, two major challenges hinder their clinical translation: i) Fratricide: Since CAR-T cells inherently express CD7, they may attack each other during ex vivo expansion, compromising their viability and persistence; and ii) off-target toxicity: The ubiquitous expression of CD7 on normal T cells and NK cells increases the risk of profound immunosuppression and related complications (22).

To address these limitations, CRISPR/Cas9 gene-editing technology has been applied to eliminate the expression of CD7, T cell receptor and human leukocyte antigen (HLA) class II molecules, thereby generating ‘universal’ CD7 CAR-T cells. These engineered cells exhibit enhanced proliferative capacity, prolonged functional persistence and reduced immunotoxicity (23). In addition, a study has explored the transient use of tyrosine kinase inhibitors, such as ibrutinib and dasatinib, to reversibly suppress CAR signaling. This pharmacological approach mitigates fratricide during cell manufacturing while preserving the cytolytic function of CAR-T cells (22). In animal models, this strategy has yielded durable antileukemic responses, and early-phase clinical trials are currently underway (22).

Collectively, CD7-directed CAR-T cell therapy offers a promising treatment modality for patients with CD7⁺ AML. Continued advancements in gene editing and pharmacologic modulation are expected to further enhance the safety and therapeutic efficacy of this approach.

CD123, the α subunit of the IL-3 receptor α, is highly expressed on leukemic blasts and LSCs in AML, while its expression is either absent or minimal in normal hematopoietic stem cells (HSCs). Several therapeutic approaches targeting CD123, such as bispecific antibodies and ADCs, are currently in clinical development (24). As early as 2002, Testa et al (25) conducted a systematic analysis of CD123 expression in AML and reported that ~45% of cases exhibited high levels of CD123. This upregulation was associated with increased blast proliferation, elevated white blood cell counts and hypersensitivity to IL-3 signaling, all of which were associated with worse prognosis. These observations have since been consistently validated in subsequent studies (26,27).

Clinically, CD123 has proven useful for risk stratification and prognostic assessment in AML. In pediatric cohorts, CD123 expression was stratified into quartiles, and patients in the highest-expression quartile (Q4) were notably enriched for adverse genetic alterations, including FMS-like tyrosine kinase 3 (FLT3)-internal tandem duplications (ITD) and lysine methyltransferase 2A rearrangements (26). By contrast, favorable genetic markers such as t(8;21), inv (16) and CCAAT enhancer binding protein α mutations were more frequently observed in the low-expression group. Patients with high CD123 expression demonstrated markedly shorter OS and event-free survival, establishing CD123 as an independent biomarker of worse prognosis (26). To address the on-target/off-tumor toxicity associated with CD123-targeted CAR-T therapy, several research groups have incorporated safety switches into CAR constructs. These include inducible systems such as iCasp9 and CD20, which can be externally activated in response to severe treatment-related toxicity (28,29). Activation of these switches enables rapid and selective elimination of CAR-T cells, thereby minimizing non-specific damage to normal tissues and improving the overall safety profile of the therapy.

FLT3 is a glycosylated protein belonging to the class III receptor tyrosine kinase family. It is predominantly expressed on HSCs and myeloid progenitors, where it plays a key role in regulating cell survival, proliferation and differentiation. FLT3 mutations are detected in >50% of patients with AML (30). These mutations are primarily categorized into two major types: ITDs within the juxtamembrane domain, accounting for ~25% of cases, and point mutations in the tyrosine kinase domain, which occurs in 6–8% of patients (31); the current standard of care for FLT3-mutated AML involves induction chemotherapy in combination with midostaurin, followed by allo-HSCT (31). A recent clinical study demonstrated that gilteritinib, when administered in conjunction with allo-HSCT, can notably extend relapse-free survival in patients with FLT3-ITD⁺ AML (32). Additionally, quizartinib has shown promise as a post-transplant maintenance therapy, sustaining FLT3 inhibition to reduce MRD and prevent relapse (33). While the majority of research to date has focused on FLT3-ITD⁺ AML, the therapeutic potential of FLT3 targeting in patients that are FLT3-ITD⁻ remains largely unexplored. Further prospective studies are needed to elucidate the role of FLT3 in non-canonical signaling pathways and to assess its broader applicability as an immunotherapeutic target in diverse AML subtypes.

The clinical outcomes summarized in Tables I and II are derived from heterogeneous studies and are not directly comparable. Differences in trial phase, sample size, patient age, disease status, target antigen selection, dose-escalation design and response criteria may substantially influence the reported CR rates and toxicity profiles. Therefore, these tables are intended to provide a contextual overview of the current clinical landscape rather than to indicate the superiority of one CAR-T target over another.

Dual-targeted CAR-T cell therapies have emerged as promising strategies to overcome the limitations of single-antigen targeting in AML. For instance, Ma et al (47) developed Loop33 × 123 CAR-T cells, a bispecific CAR-T design incorporating tandem antigen-recognition domains that enables simultaneous targeting of the two myeloid antigens CD33 and CD123. This dual targeting effectively eliminates both AML blasts and LSCs. In vitro assays demonstrated that Loop33 × 123 CAR-T cells exhibited superior cytotoxicity compared with single-target CAR-T cells, while an in vivo study showed notable prolongation of survival in murine models without evident toxicity (48).

Similarly, dual-target CAR-T cells targeting CD123 and CLL-1, designed either via a bicistronic vector or tandem scFv architecture, have been developed to achieve precise recognition and elimination of diverse AML subtypes. A study reported that CD123/CLL-1 dual-targeted CAR-T therapy exerted robust antileukemic effects in animal models and reduced the likelihood of antigen escape (49). Furthermore, comparative analyses of two tandem CAR constructs (CD123/CLL-1 vs. CLL-1/CD123) revealed that dual-targeted CAR-T cells possessed enhanced tumor-killing efficacy and improved resistance to antigen escape relative to single-target counterparts (50). Beyond these, other dual-target CAR designs such as CD33/CLL-1 are also under investigation, further broadening the immunotherapeutic target landscape in AML.

While these findings highlight the potential of dual-target strategies, their translational relevance in AML remains uncertain because the available evidence is still largely preclinical. Expanding antigen coverage may help reduce immune escape, but it may also increase the risk of hematopoietic toxicity when target expression overlaps with normal cells. In addition, greater structural complexity may introduce further challenges for signaling coordination, persistence and manufacturing consistency (5). Whether these theoretical advantages can be translated into a clinically meaningful improvement in efficacy without compromising safety will require validation in future clinical studies.

Beyond target selection, Fig. 2 highlights the importance of engineering strategies to improve T cell fitness, controllability and resistance to immunosuppressive cues in AML CAR-T therapy.

Conventional CAR-T therapy relies on autologous T cells harvested from patients with hematologic malignancies, which entails challenges such as prolonged manufacturing times, high costs and considerable product heterogeneity (51). UCAR-T cells, derived from healthy donors and genetically engineered via CRISPR/Cas9 or TALEN technologies to disrupt T cell receptor and HLA genes, represent ‘off-the-shelf’ universal CAR-T products designed to minimize the risk of GVHD. Recent preclinical studies have demonstrated that optimized UCAR-T cells exhibit a defined memory phenotype and dose-dependent antitumor efficacy, along with favorable safety profiles in xenograft models (52).

UCAR-T is attractive because it may simplify production and reduce delays associated with autologous manufacturing. However, its broader clinical application is still constrained by several unresolved issues, including alloreactivity, host immune rejection, limited persistence after infusion and the need for robust large-scale manufacturing. Gene editing may help address some of these barriers, but it also raises additional concerns regarding genomic stability and regulatory oversight (53,54). Therefore, the clinical value of UCAR-T will depend on whether these products can achieve reproducible safety and durable efficacy in patients.

Logic-gated CAR-T cells regulate T cell activation and cytotoxicity by integrating multiple cellular signals, enabling precise discrimination between malignant and normal cells. Key design strategies include: i) OR gate CARs, which recognize multiple antigens simultaneously and are thus referred to as multi-antigen CARs (55); ii) gate CARs, which require co-expression of two target antigens (such as CD19 and CD22) to become fully activated (56); and iii) NOT gate CARs, which detect markers expressed on healthy cells (such as CD34) and trigger inhibitory signaling pathways to protect normal tissues and reduce off-tumor toxicity (55).

These designs are conceptually attractive as they offer a more selective way to distinguish leukemic cells from normal tissues. Their clinical application, however, may be more complicated than their design principle suggests. Because logic-gated systems depend on coordinated multi-signal responses, their activity may vary with antigen density, dynamic target expression and the surrounding leukemic microenvironment. Greater circuit complexity may also complicate vector construction, manufacturing workflows and quality control (57,58). Whether this added precision is sufficient to justify the increased complexity remains an important translational question.

The clinical translation of CAR-T therapy is advancing rapidly in parallel with technological innovations that integrate manufacturing processes with AI. Leading pharmaceutical companies including Gilead, Novartis, Johnson & Johnson and Bristol-Myers Squibb have successfully reduced CAR-T cell manufacturing timelines from >30 days to ~14 days, with ongoing efforts aimed at achieving a 7-day production cycle. This acceleration is primarily driven by the adoption of automated, good manufacturing practice-compliant platforms and streamlined, rapid quality control procedures (51). For instance, fully automated, closed-loop manufacturing systems such as CliniMACS Prodigy, coupled with expedited 24-h production workflows, have demonstrated feasibility in preclinical and early clinical settings while preserving key naïve and memory T cell subsets essential for therapeutic efficacy (59). Concurrently, AI has emerged as a transformative tool in CAR-T development across multiple domains: i) CAR design and target recognition optimization, where machine learning algorithms predict antigen affinity, structural stability and intracellular signaling activation to inform the engineering of enhanced CAR constructs; ii) intelligent manufacturing process control, leveraging digital twin models and reinforcement learning to enable real-time monitoring and regulation of bioreactor parameters, thereby improving process stability and product yield; and iii) quantitative prediction of therapeutic efficacy and toxicity through the integration of multi-omics, imaging and clinical datasets, providing robust support for dose optimization and personalized treatment strategies.

Collectively, these innovations herald the emergence of a ‘digital pharmaceutical’ paradigm in CAR-T therapy, offering promising avenues to enhance safety, manufacturing efficiency and patient accessibility. From an engineering perspective, these optimization strategies have substantially expanded the design space of AML CAR-T therapy, however, technical sophistication does not necessarily guarantee clinical feasibility. For dual-target CARs, universal CAR-T platforms and logic-gated systems alike, the central translational challenge is whether theoretical advantages can be converted into reproducible clinical benefit without introducing excessive toxicity, manufacturing complexity or regulatory barriers. Addressing these issues will be essential for translating next-generation CAR-T strategies into meaningful and durable clinical benefit.