Machine learning (ML) technologies have played a pivotal role in advancing precision immunotherapy by integrating multi-omics data to identify biomarkers, predict treatment responses, discover novel therapeutic targets (1, 2), characterize the tumor microenvironment, and optimize patient stratification. These predictive models have greatly enhanced clinical decision-making capabilities (3, 4). However, the application of ML in immunology has increasingly come under scrutiny. Traditional models often fail to capture the complexity of immune interactions (5), suffer from the “black-box” nature of deep learning (6), and lack standardized data preprocessing protocols (7).

Despite broad recognition that “correlation ≠ causation” is a fundamental statistical principle, this distinction is frequently overlooked in practice. A systematic review of 90 studies on immune checkpoint inhibitors (ICIs) revealed that while 72% employed traditional ML and 22% used deep learning, none incorporated causal inference. Consequently, these models were not included in phase III clinical trial designs or referenced in major clinical guidelines (8). This phenomenon is not isolated: a parallel analysis of 36 melanoma prediction models showed all studies were retrospective correlation-based analyses, with none applying causal inference. As a result, PROBAST evaluations rated them as having moderate to high bias, limiting their translational utility and clinical applicability (9, 10).

This disconnect between knowledge and practice highlights a broader issue in immunology research—an overreliance on digital correlations. Researchers may acknowledge the importance of causality but are deterred from applying causal frameworks due to the intrinsic complexity of immunological data. High-dimensional, noisy, and temporally dynamic immune responses, combined with treatment-induced nonlinear effects and substantial interindividual heterogeneity (across genotype, phenotype, and microenvironment), pose significant challenges to conventional causal inference methods (11–14).

Fortunately, recent methodological advances have made the integration of causal inference and ML increasingly feasible. For example, the Targeted-BEHRT model combines transformer architecture with doubly robust estimation to infer long-term treatment effects from longitudinal, high-dimensional data (15). Causal network models incorporating selection diagrams, missingness graphs, and structure discovery techniques outperform standard ML in risk evaluation and adverse event prediction for immunotherapies (16). CIMLA exhibits exceptional robustness to confounding in gene regulatory network analysis, offering insights into tumor immune regulation (17). CURE, leveraging large-scale pretraining, improves treatment effect estimation with gains of ~4% in AUC and ~7% in precision-recall performance over traditional methods (18). Causal-stonet handles multimodal and incomplete datasets effectively, crucial for big-data immunology research (19). LingAM-based causal discovery models have demonstrated high accuracy (84.84% with logistic regression; 84.83% with deep learning) and can directly identify causative factors, significantly improving reliability in immunological studies (20).

These innovations represent a confluence of causal reasoning and machine learning methodologies (21), which are now being increasingly applied in immunology research (22, 23). They help reveal true causal relationships, mitigate confounding (both observed and unobserved), enhance model interpretability and robustness (24), and integrate heterogeneous data types including genomics, proteomics, clinical phenotypes, and medical imaging (25, 26). Ultimately, they enable the construction of more realistic models with superior generalizability and predictive performance across diverse patient populations (27, 28).

This Perspective aims to systematically highlight the paradigm-shifting value of causal machine learning in immunological research. We focus on the following key questions ( Figure 1 ):

In current immunotherapeutic research, traditional machine learning (ML) models primarily rely on retrospective data mining of correlations (29), yet they often fail to explore the underlying causal mechanisms (30). For instance, in studies on the gut microbiome and immune checkpoint inhibitors (ICIs), although advanced algorithms such as Random Forests and SVMs were employed, only 4 out of 27 studies conducted cross-validation. Furthermore, key confounding factors such as antibiotic use and dietary differences were not adequately controlled, resulting in highly heterogeneous and unreliable conclusions regarding the efficacy of the same microbial strains (31). Similarly, in the analysis of immune-related adverse events (irAEs) and survival, traditional Cox regression yielded a hazard ratio (HR) of 0.37, implying a protective effect of irAEs. However, causal ML using target trial emulation (TTE) to correct for immortal time bias revealed a true HR of 1.02—completely overturning the conventional belief that irAEs improve prognosis (32). These findings underscore the urgent need for sound causal inference in immunological studies to avoid conclusions that contradict biological plausibility.

Moreover, the insufficient recognition of the importance of causal inference among researchers (33) has led to multiple problems. Notably, effective therapies may be erroneously rejected due to improper grouping strategies (34), while correlations that appear statistically significant (35) may be misinterpreted as causal relationships (36), leading to misleading clinical implications (37). For example, studies examining the impact of antibiotic exposure on ICI outcomes reported a statistically significant HR of approximately 1.3, yet the authors explicitly acknowledged the presence of residual unmeasured confounders. This raises the risk of inappropriate clinical decisions, such as the unjustified discontinuation of antibiotics due to a presumed class-wide harmful effect (38). Likewise, deep learning models based on CT radiomics for predicting ICI responses reported an AUC of ~0.71, but the signal captured largely reflected confounders such as tumor burden and treatment line rather than true drug sensitivity, casting doubt on the validity of the model’s conclusions (39).

Therefore, neglecting causal inference not only compromises the reliability of study results (40), impedes clinical translation (41–43), and misguides clinical decision-making, but also wastes research resources and delays the development of effective therapies (44). A typical example is seen in COVID - 19 vaccine research, where including non-virus-related hospitalizations (“false-positive cases”) led to substantial underestimation of the protective effect of vaccines that primarily prevent severe post-infection complications rather than infection itself—ultimately resulting in misleading conclusions about vaccine efficacy (45).

Although the importance of causal inference has been increasingly recognized in clinical research, many studies still rely on conventional causal inference methods, which face significant challenges in practice. Randomized controlled trials (RCTs) are often infeasible due to high costs, ethical constraints, and heterogeneity among patients (46). Stratified designs in observational studies struggle with high-dimensional omics data, and multivariable regression fails to capture the nonlinear characteristics of the immune system (47). Propensity score methods (PSM), based on the unrealistic assumption that all confounders are measurable, have been misapplied in 72% of studies (8). Mendelian Randomization (MR) also faces methodological limitations, including susceptibility to false associations and estimation bias stemming from the quality of genetic instruments and core assumptions (48, 49). Specifically, MR applications in immunology face four major hurdles: violation of the instrumental variable assumption due to pleiotropy; weak instruments owing to low heritability of immune exposures; a mismatch between lifelong genetic effects and short-term therapeutic interventions; and systematic bias from population stratification (50–52). Collectively, these limitations have constrained the application and scalability of traditional causal inference approaches in immunology.

Table 1 presents representative cases where correlation-based analyses failed, while Table 2 summarizes the limitations of traditional causal inference methods.

To overcome the limitations of both traditional causal inference and conventional machine learning approaches, causal inference-based machine learning (causal ML) models have emerged ( Figure 2 ). Compared to classical causal methods such as propensity score matching (PSM), Cox regression, or linear models, causal ML lifts the constraints of strict parametric assumptions and rigid model forms, enabling more flexible modeling of the nonlinear dynamics and high-dimensional interactions inherent to immune systems (53–55). For instance, CV-TMLE, when applied in a small-scale study of only 168 ICU patients with COVID - 19, employed the Super Learner ensemble approach to effectively relax regularity conditions and increased the 95% confidence interval coverage by 10-20 percentage points compared to standard methods (53). Similarly, the ANN-DML estimator demonstrated a ~30% reduction in mean squared error (MSE) relative to conventional kernel smoothing methods when handling extremely high-dimensional scenarios where the number of immune biomarkers scales with sample size (p → 2n) (54).

Moreover, causal ML enables multi-modal modeling by integrating imaging, text, time-series, and genomic data. For example, Clinical Transformer can fuse clinical records, laboratory metrics, and sequencing data. By leveraging counterfactual perturbation strategies, it achieved an improvement of 0.05-0.10 in C-index across seven cancer types (56). MOFS effectively integrates MRI, pathology, and multi-omics data to identify glioma subtypes most responsive to anti-PD-1 therapy (57), while Bio-relevant AI combines imaging, pathology, and gene expression data to help 32% of stage II colorectal cancer patients avoid unnecessary chemotherapy (58). These unique strengths contribute to more accurate prediction of therapeutic outcomes (33), optimizing drug use and enhancing treatment efficacy (59).

In contrast to conventional machine learning methods such as random forests, LASSO, or deep learning—models that rely solely on correlational pattern discovery—causal ML shifts the focus from predicting associations to identifying causality. For instance, the Super Learner ITE framework estimates individual treatment effects (ITE) through model ensembling, achieving an AUC of 0.77 in external validation, with decision curve analysis showing a significantly higher net clinical benefit compared to treat-all or SAPS-II strategies (60). Similarly, in the MiCML platform study, Causal Forest utilized adaptive partitioning to estimate conditional average treatment effects (CATE), reducing prediction error for treatment–microbiome interaction effects by 25-40% compared to traditional LASSO regression (55).

Furthermore, causal ML effectively addresses key limitations of correlational models—namely spurious associations and confounding bias—by enabling robust control of unmeasured confounding (61). This facilitates the clarification of true causal relationships between immune cells and disease (36). For instance, COCA utilizes negative control outcome calibration to restrict estimation bias to less than 40% of that seen in conventional OLS models (62), and CV-TMLE improves 95% confidence interval coverage (53). Collectively, these advantages enhance model performance (63), clinical interpretability (43), and generalizability (64), providing robust scientific guidance for clinical decision-making (40).

In addition, mechanism-aware causal ML approaches embed biological prior knowledge into model structures, achieving a unification of data-driven and mechanism-driven strategies—a closed loop between computation and experimentation (65). This integration enables better capture of complex clinical phenotypes, deeper mechanistic insights (41), and enhanced feasibility and translational value of biomedical research (66). Consequently, causal ML provides promising avenues for early detection strategies (64) and novel drug development pipelines (34).

Table 3 summarizes the unique advantages of causal ML methods, while Table 4 outlines their applications in multi-dimensional data integration.

At the data acquisition level, the presence of inaccurate or incomplete data significantly hinders the implementation of causal inference models. In particular, measurement errors can amplify causal bias, thereby undermining the reliability of results (67). Moreover, when missing data violate identifiability assumptions, no estimator can recover the true causal effect, rendering any derived causal inference invalid (68).

At the clinical application level, causal machine learning (Causal ML) models often exhibit a “black-box” nature, which severely limits clinician acceptance (69). When internal parameters and computational processes become overly complex, it becomes difficult for clinicians to understand how conclusions are derived, ultimately impeding clinical translation (70, 71).

At the research methodology level, both methodological selection difficulties and interdisciplinary collaboration barriers constrain the advancement of Causal ML in immunological research. Causal relationships vary in structure and often require tailored methods, yet the abundance of available approaches—each with unique limitations—makes optimal selection challenging, especially for researchers with limited formal training in causal modeling (33). Furthermore, interdisciplinary efforts are frequently impeded by cultural and conceptual gaps between domains. For instance, biomedical scientists tend to focus on clinical applicability, statisticians emphasize methodological validity, and computer scientists prioritize algorithmic performance. These differing priorities can lead to communication breakdowns and ultimately slow scientific progress (72, 73).

Table 5 summarizes the three major challenges faced by Causal ML.

Over the next five years, addressing the two core challenges—data quality and model interpretability—will require the development of innovative technical solutions. In terms of data quality, the integration of multiple imputation with the G-formula has significantly reduced bias caused by missingness in cystic fibrosis studies (74). Likewise, the MI-BART method has demonstrated superior robustness in multi-treatment comparisons (75), offering promising prospects for enhanced data control and quality improvement over the next 5-10 years.

Regarding interpretability, studies have shown that Causal-XAI hybrid frameworks can generate causal attribution heatmaps, enabling physicians to better understand image-based decisions (76). In addition, CLARUS, an interactive counterfactual reasoning platform, allows clinical experts to directly manipulate and verify model reasoning chains (77). This effectively addresses the “black-box” issue by clarifying causal pathways underlying model outputs (78), thereby improving both clinical decision-making and regulatory trust, ultimately facilitating clinical translation (79). In the future, the integration of Bayesian nonparametric models and natural language processing (NLP) is expected to further enhance model performance by extracting authentic causal structures from large-scale biomedical data (80), revealing deep causal relationships and identifying novel therapeutic targets (81, 82).

In the next 5-10 years, methodological integration will become a central theme. The emerging “triangulation framework” will be more widely adopted. This framework enhances the robustness of causal inference by integrating and cross-validating multiple approaches such as instrumental variables (IVs), regression discontinuity (RD), and propensity scores (83). In parallel, strengthening interdisciplinary collaboration and talent development will become essential. Multidisciplinary teams can develop shared terminologies and workflows, promoting effective integration across epidemiology, economics, and clinical medicine (84) and enabling each field to contribute its strengths to solve complex problems (42, 73). Cultivating versatile professionals capable of navigating the intricacies of immune-related biological systems (85) will help dismantle disciplinary silos and address challenges in resource allocation and coordination (86). This integrated approach will enable more comprehensive solutions (87) to meet the rapidly evolving demands of immune drug research (88). Furthermore, academic institutions should establish dedicated programs and curricula to train cross-disciplinary talent in causal inference and immunotherapy (42, 73), fostering the convergence of modern science and specialized education (89), promoting skills development (90), and facilitating global collaboration in immunology research (91), injecting new vitality and opportunity into the field.

In the next 5-10 years, causal inference models are expected to be widely implemented in clinical immunology. One notable development is the “Perturbation Cell Atlas” proposed by Rood et al., which represents a conceptual turning point. Future research will likely build on this by leveraging large-scale CRISPR-scRNA-seq perturbation datasets to train and deploy foundational causal models for practical guidance (92). Technologically, tools such as Velorama, which has shown great promise in immune differentiation studies, will play a pivotal role. By integrating RNA velocity to express cellular developmental trajectories as directed acyclic graphs (DAGs), these tools enable causal network inference at single-cell resolution, a capability expected to be expanded in future research (93).

With the continued advancement of artificial intelligence, AI-assisted vaccine design is poised to become a prevailing trend. This will necessitate the use of target trial emulation, causal NLP, and federated causal estimation frameworks to identify causally relevant endpoints and accelerate critical discoveries (94, 95). Moreover, as federated learning frameworks mature across institutions (76), interpretable causal tools such as CIMLA will likely become standardized (96), enabling a full transition of causal inference from theoretical development to routine clinical decision support. This process will be further facilitated by improvements in data quality and model robustness through rigorous control of covariates and confounding variables (97–99), which are essential for enhancing the credibility, transparency, and real-world applicability of causal models in clinical settings.

Table 6 presents strategies to address the three major challenges, while Table 7 outlines the projected applications of causal ML in immunology over the next 5-10 years.