Demographic and clinical characteristics were significantly different between cohorts (Supplementary Data 2). Supplementary Data 2 offers detailed clinical and demographic breakdowns that reveal key differences in age, race, and survival across cohorts—context that anchors later stratification findings. For example, patients tended to be younger (avg. 42.7 years, p < 0.001), predominantly white (N = 260, 95.2%, p < 0.001) and had the highest rates of remaining with tumor (N = 139, 58.6%, p < 0.001) in the LGG cohort, more male patients (N = 111, 65.7%, p < 0.001) with the shortest overall survival (avg. 467.0 days, p < 0.001) and tumor-free survival (avg. 381.5 days, p < 0.001) in the GA cohort, and more diverse (non-white N = 53, 22.5%, p < 0.001) patients with the longest overall survival (avg. 1030.5 days, p < 0.001) and progression-free survival (avg. 929.1 days, p < 0.001) in the EC cohort.
MuTATE outperforms CART in simulations
In the evaluation of 18,400 simulations, MuTATE models consistently demonstrated superior performance over CART, yielding significantly improved error, as well as improved true and false discovery rates in multivariable analyses. (Fig. 2, Figs. S1–2, Supplementary Data 1). In each simulation, 100 synthetic multi-target GT trees were constructed, synthetic sets were divided into train and test sets (60/40 data split), and grid search assessed MuTATE trees and averaged single-target CART models for test error, true discovery rate (TDR), and false discovery rate (FDR) across model parameters in 18,400 synthetic datasets (Fig. 1b, c, Fig. S2).
a MuTATE enables explainable multi-endpoint ML by evaluating features across clinical endpoints45. Partitions are based on information gained (IG) using highest average multi-target IG (avgIG), highest IG in any target (maxIG), meaningful IG in the most targets (mostIG), lowest average p-value of statistically significant IG (avgPVal), lowest p-value, weighted by number of targets with significant IG (minPVal), significant IG in the most targets (mostPVal), or subtree lookahead (splitError). Trees predict endpoints and identify biomarkers and subtypes. b Synthetic multi-target data were generated using a positive definite covariance matrix of targets using a correlation structure (mean \(\mu\) = 1, SD \(\sigma\) = 1). Features were generated and sampled with replacement for ground truth (GT) definition, targets were divided into leaf quantiles and randomly assigned, resulting in multi-target tree-structured data with a known GT. Clinical cohorts with established expert trees were obtained from TCGA from the NCI Genomic Data Portal. 682 biopsies from three cohorts of 711 patients were included. c In simulations, synthetic data and GT are divided into train/test sets (60/40 data split), and grid search assesses model parameters for model test error, TDR, FDR in 18,400 synthetic datasets. Clinical cohorts were divided into train/test sets (60/40 data split), training sets underwent parameter tuning, model performance was captured. Tuned parameters used in trained models were applied to the full cohorts. Final trees were assessed for prognostic significance of partitions, biomarkers, and subtypes. See Figs. S1-6.
Figure 1 provides an overview of the MuTATE framework, including the model architecture (a), data preparation and cohort characteristics (b), and evaluation strategy across simulations and clinical datasets (c). Panel 1a outlines how multi-endpoint inputs are processed to generate interpretable decision trees. Panel 1b contrasts synthetic data generation with real TCGA cohort preprocessing. The left side shows how synthetic datasets were constructed to mimic multi-endpoint complexity, while the right side displays key statistics (e.g., sample size, endpoints, features) for each TCGA cohort. Panel 1c then illustrates the end-to-end evaluation pipeline—from data split and parameter tuning to model testing—clarifying the experimental setup across both synthetic and clinical datasets.
While CART built separate models for each endpoint and was unable to explainably capture the GT, MuTATE accurately identified the GT with improved performance in one interpretable model, highlighting its superior ability to explainably and accurately represent complex data in a clear visual model (Fig. S1). CART had the highest test error (2.97, 95%CI: 2.93–3.01). MuTATE outperformed CART for model depths of two and above. As targets increased, MuTATE FDR dropped from 12.7% (2-target 95%CI: 8.7–16.7%) to 5.7% (5-target 95%CI: 2.7–8.7%), while CART maintained a constant FDR (11.0%, 95%CI: 9.0–12.0%). Sample size and inter-target correlation did not show a clear trend in performance of either method. As features increased, a decline in performance was observed across all methods, including CART. In multivariable logistic regression analysis of simulation performance, adjusting for simulated characteristics, all MuTATE partitioning options showed statistically significantly lower test error compared with CART (p < 0.001) (Fig. 2, Supplementary Data 1). MuTATE using splitError partitioning showed statistically significantly higher TDR (p < 0.001), and lower FDR compared with CART (p = 0.005), adjusting for simulated characteristics (Supplementary Data 1). MuTATE demonstrated superior explainability in constructing clear and interpretable models that accurately capture the underlying GT in complex data.
Multivariable analyses assessing method and model performance in 18,400 synthetic multi-target dataset simulations (adjusted for simulated sample size, number of targets, number of features, inter-target correlation, model depth, and run ID). Boxplots represent regression coefficient distributions across 18,400 simulations. Points indicate mean coefficient estimates with 95% CI. Positive coefficients reflect improved performance relative to CART (e.g., higher TDR), while negative coefficients reflect lower error or FDR. Whiskers represent interquartile range, and overlaid points display mean coefficients with 95% confidence intervals, derived from multivariable regression adjusting for simulation conditions.
Clinical cohorts: real-world validation of MuTATE
To assess real-world performance, we applied MuTATE to three independent clinical cohorts from TCGA (LGG, EC, and GA), each with corresponding expert-derived decision trees. These represent the only TCGA cohorts with such clinical models, allowing rigorous validation against established clinical subtypes. In cross-validated model selection across three cohorts, MuTATE significantly outperformed CART, highlighting its utility for clinical modeling (Supplementary Data 3). CART did not produce meaningful models on these clinical cohorts, as it failed to identify any partitions or key biomarkers in the datasets, resulting in 0 partitions. This outcome further demonstrates the robustness of MuTATE in stratifying disease heterogeneity and identifying relevant molecular features for precision medicine applications. Clinical endpoints for modeling included overall survival (OS, all), tumor-free survival (TFS, all), progression-free survival (PFS, LGG and GA), recurrence-free survival (RFS, GA and EC), vital status (all), neoplasm status (LGG and GA), new tumor (LGG and GA), recurrence (GA), progression (LGG and EC), and treatment response (GA) (Supplementary Data 2). Clinical cohorts were divided into train and test sets (60/40 data split), training data underwent k-fold cross-validated parameter tuning using grid search, model performance was captured, the best performing ML method was selected for tree construction, and final trees were applied to full cohorts (Fig. 1b, c). TDR and FDR were calculated based on manually constructed expert models as GT. MuTATE outperformed CART in cross-validated parameter tuning, highlighting improved ability to capture data complexity in clinical applications (Supplementary Data 3). Supplementary Data 3 outlines cross-validated performance metrics across all MuTATE partitioning strategies, enabling comparison of statistical heuristics in each cancer type. To identify the best-performing strategies across cohorts, we evaluated all MuTATE partitioning methods. Notably, avgIG performed best in LGG, splitError in GA, and mostPVal in EC—each outperforming CART on all performance metrics (Supplementary Data 3). MuTATE using avgIG (LGG), splitError (GA), and mostPVal (EC) partitioning achieved the highest TDR (LGG: 91%, 95%CI: 82–99%; GA: 42%, 95%CI: 24–59%; EC: 1.00, 95%CI: 1.00–1.00) and lowest test error (LGG: 0.07, 95%CI: 0.06–0.08, GA: 0.13, 95%CI: 0.06–0.20; EC: 0.15, 95%CI: 0.14–0.15) compared with CART, which had the lowest TDR and highest test error. Multi-target trees showed consistent improvement in test error, TDR, and model explainability, automating expert architectures across clinical cohorts.
Refined molecular subtypes enhance prognostic stratification
MuTATE molecular models for LGG, GA, and EC have revealed novel molecular signatures when validated against manually-constructed established models, augmenting the existing clinical understanding of these diseases (Figs. 3–5, Figs. S4–6). Figure 3 provides a visual comparison of expert-derived vs. MuTATE-derived LGG subtypes, helping readers understand how MuTATE refines patient stratification. The middle panel illustrates how mutation status (e.g., CIC, NOTCH1) further stratifies risk within groups defined by traditional markers like IDH and 1p19q. The color gradient reflects increasing clinical severity across endpoints. In the LGG cohort, MuTATE identified CIC and NOTCH1 (HGNC:7881), ATRX, and NF1 (HGNC:7765) as key markers of heightened LGG severity, augmenting risk stratification and potentially reshaping clinical strategies, while also successfully recognizing the previously-established manual expert partitions on IDH1 variant and 1p19q codeletion (Fig. 3, Fig. S4). Surprisingly, patients initially classified as “low-risk” in established manual models due to IDH variant and 1p19q codeletion LGG, may face a higher risk of severe disease (23.2% mortality, 45.5% progression, 36.4% new tumor) if they harbor concurrent CIC and NOTCH1 variants (inactivating molecular alterations associated with 1p19q codeletion), signifying the need for intensified monitoring and personalized therapeutic strategies. NF1 variant (an inactivating molecular alteration associated with IDH1 wild-type) emerged as an indicator of aggressive disease (45.5% mortality, 54.5% progression, 36.4% new tumor) in patients with IDH wild-type LGG, underscoring the imperative for tailored therapeutic approaches and vigilant surveillance. Lastly, ATRX variant (an inactivating molecular alteration associated with IDH variant) correlates with increased disease progression (30.4%) in IDH variant LGG, reinforcing MuTATE’s role in refining risk paradigms and personalized treatment in LGG.
The MuTATE-generated multi-endpoint decision tree stratifies LGG patients first by IDH1 and 1p19q status (top section, aligned with expert subtypes), and then by CIC, NOTCH1, NF1, and ATRX mutation status (middle section, representing MuTATE-defined final subtypes). To interpret this figure, follow each branching path from the expert subtype (top) to the MuTATE-defined subgroups (middle). Each node summarizes how the presence or absence of specific mutations alters risk across endpoints. Icons and color gradients reflect increasing clinical severity and inform potential clinical actions (bottom). Each node includes the number of patients (N, %) and subtype-specific percentages for key clinical endpoints: mortality, progression, new tumor events, and neoplasm status. Color shading reflects estimated clinical severity, with darker shades indicating higher-risk subtypes (defined by rates of death, progression, new tumor events, and neoplasm status). Summary statistics (% of patients with each clinical outcome) are provided for each subtype to illustrate clinical heterogeneity. Statistical associations between MuTATE subtypes and clinical outcomes were evaluated using logistic and Cox regression models and are reported in Fig. S4 and Supplementary Data 4–6. These results demonstrate MuTATE’s ability to replicate expert-defined classifications (e.g., IDH1-1p19q) while revealing more granular, higher-risk subgroups that were not captured by existing clinical models. The bottom section of the figure summarizes potential implications for clinical decision-making based on MuTATE-defined subtypes, including options for tailored monitoring, therapeutic escalation, or intensified surveillance. Notably, CIC and NOTCH1 variants stratified a higher-risk group within the traditionally low-risk IDH-mutant population, while NF1 and ATRX variants flagged more aggressive disease courses—highlighting MuTATE’s potential to inform post-resection therapeutic decisions and targeted surveillance strategies.
The top section illustrates the expert-defined GA subtypes based on EBV, MSI, CIN, and GS classifications, including corresponding patient counts and risk categories. The middle panel shows the MuTATE-derived multi-endpoint decision tree, which first stratifies patients by RHOA and EBV status, then further divides subgroups based on ARID1A mutation status—revealing clinically distinct phenotypes overlooked by expert-defined GA subtypes. To interpret this figure, follow each branching path from the expert subtype (top) to the MuTATE-defined subgroups (middle). Each node summarizes how the presence or absence of specific mutations alters risk across endpoints. Icons and color gradients reflect increasing clinical severity and inform potential clinical actions (bottom). Each node displays the number of patients (N, %), and subtype-specific percentages for key clinical endpoints: mortality, recurrence, new tumor events, neoplasm status, and treatment response. Color shading reflects clinical severity, with darker shades indicating higher-risk subtypes (defined by rates of death, recurrence, new tumor events, neoplasm status, and treatment response). MuTATE identified subgroups with markedly worse outcomes—such as RHOA-mutant GA with high rates of non-remission and recurrence, and ARID1A wild-type tumors within the traditionally “genomically stable” group—highlighting hidden risk missed by expert models. The bottom section outlines potential treatment implications, such as intensified post-operative therapy and closer surveillance. These findings underscore MuTATE’s ability to uncover granular, prognostically meaningful subtypes that could support more personalized and equitable treatment decisions in GA. Statistical associations were evaluated using logistic and Cox regression (Fig. S5, Supplementary Data 4–6). EBV Epstein Barr-Virus, MSI microsatellite instability, CIN chromosomal instability.
The top section presents expert-defined EC subtypes based on POLE, MSI, MLH1, and copy number (CN) classifications, including corresponding patient counts and risk categories. The middle panel shows the MuTATE-derived multi-endpoint decision tree, which first stratifies patients by MLH1 hypermethylation status, followed by copy number alterations (for hypermethylated patients), and POLE and CN status (for non-hypermethylated patients). To interpret this figure, follow each branching path from the expert subtype (top) to the MuTATE-defined subgroups (middle). Each node summarizes how the presence or absence of specific mutations alters risk across endpoints. Icons and color gradients reflect increasing clinical severity and inform potential clinical actions (bottom). Each node reports the number of patients (N, %) and subtype-specific percentages for key clinical endpoints: mortality, progression, and disease status. Color shading reflects estimated clinical severity, with darker red tones indicating higher-risk subtypes based on endpoint distributions. MuTATE uncovered high-risk subgroups that were not visible in expert-defined models. Notably, patients with MLH1 hypermethylation and high copy number alterations—traditionally considered “intermediate-risk”—exhibited markedly worse outcomes, including the highest observed rates of death, persistent disease, and progression. Similarly, patients with high CN and wild-type POLE status among the MLH1 non-hypermethylated group were reclassified as high-risk, despite not being flagged by expert models. These findings underscore MuTATE’s capacity to reveal granular, multi-endpoint phenotypes that may better inform post-operative risk stratification, treatment intensification, and surveillance strategies. The bottom panel summarizes potential clinical implications for each MuTATE-defined subtype, ranging from routine monitoring to close progression and new tumor surveillance. Statistical associations between MuTATE subtypes and clinical outcomes were evaluated using logistic and Cox regression and are reported in Fig. S6 and Supplementary Data 4–6. CN copy number, MSI microsatellite instability, LTF lost-to-follow-up.
Figure 4 highlights MuTATE’s ability to uncover clinically relevant subtypes in GA that were overlooked by expert models. For example, MuTATE distinguishes a high-risk ARID1A wild-type subgroup within the genomically stable category, shown in darker red to reflect increased recurrence and disease progression. In the GA cohort, MuTATE expanded upon the expert model by identifying EBV, RHOA, and ARID1A as key indicators of GA severity, enriching clinical phenotype and treatment response classification and reshaping clinical decision-making (Fig. 4, Fig. S5). MuTATE identified EBV (the primary GA expert partition), RHOA (frequently mutated in the “genomically stable” expert subtype) and ARID1A (a GA tumor suppressor46), revealing novel clinical stratifications that were not captured by the expert model. Remarkably, individuals with RHOA variant GA (including patients from all expert subtypes) exhibited a strikingly high progression rate (44.4% non-remission) during treatment, revealing a critical clinical paradox of robust survival alongside elevated disease aggressiveness, thus highlighting a pressing need for tailored therapeutic interventions. This progression was under-recognized in the expert model, demonstrating MuTATE’s potential to highlight key clinical differences. EBV signaled severe disease (35.3% mortality, 58.3% non-remission), aligning with the EBV expert subtype, warranting targeted therapies and vigilant surveillance among those with RHOA wild-type and EBV-positive GA. Crucially, MuTATE uncovered novel insights regarding ARID1A status in GA. Individuals harboring ARID1A wild-type GA exhibited elevated rates of mortality (31.4%), recurrence (24.5%), new tumors (23.7%), and progression (36% non-remission) during treatment, suggesting that the ARID1A wild-type group, previously overlooked by the expert model, represents a higher-risk category that should be monitored for timely interventions. Notably, most patients in the GS expert subtype, were reclassified into the higher-risk ARID1A wild-type group, indicating that the expert model underestimated the clinical severity of these patients. This reclassification suggests that ARID1A wild-type status may define a previously overlooked high-risk subgroup in GA.
Figure 5 shows how MuTATE reclassifies EC patients by incorporating copy number and MLH1 methylation status. Patients previously considered intermediate-risk based on expert models are reclassified into a highest-risk group, reflected in the red-shaded node under the “MLH1 hypermethylated + high CN” branch. In the EC cohort, MuTATE identified copy number (CN), POLE, and MLH1 hypermethylation as key markers of severity, enriching clinical phenotype classification and potentially reshaping clinical decision-making (Fig. 5, Fig. S6). Strikingly, among EC patients typically regarded as “intermediate-risk” in established manual models due to MLH1 hypermethylation, those with concurrent high CN exhibited not only the lowest survival (15.2% mortality) rates but also the highest incidence of persistent disease (15.9%) and a higher progression rate (10.2%), casting a spotlight on a previously underestimated high-risk EC subgroup for intensified monitoring and specialized treatment strategies. POLE variant signaled less severe disease in MLH1 not hypermethylated EC, aligning with the POLE expert subtype and reinforcing the robustness of our classifications. High CN signals severe disease (10.6% mortality, 26.5% progression, 14.3% persistent disease) in those with POLE wild-type and MLH1 not hypermethylated EC, aligning with the CN-high expert subtype, warranting tailored therapeutic approaches and vigilant surveillance. Lastly, low CN signals the least severe disease in POLE wild-type and MLH1 not hypermethylated EC, suggesting less severity than the expert subtype and underscoring MuTATE’s ability to capture multiple endpoints for clinical decision-making.
MuTATE subtypes are prognostically informative
MuTATE partitions and the biomarkers they identify are independent prognostic markers of disease, defining subtypes associated with multiple endpoints, revealing novel phenotype heterogeneity in all cohorts.
In LGG, the partition on ATRX variant predicted new tumor events (aOR=2.89, 95%CI: 1.00–10.54, p = 0.071) compared to those without ATRX variant among patients with IDH1 variant (Fig. S4). In multivariable biomarker analyses, IDH1 variant protected against death (aOR = 0.33, 95%CI: 0.14–0.76, p = 0.010; aHR = 0.39; 95%CI: 0.20–0.78, p = 0.008), progression (aOR=0.33, 95%CI: 0.14–0.74, p = 0.007; aHR = 0.36, 95%CI: 0.20–0.65, p = 0.001), and new tumor events (aHR=0.37, 95%CI: 0.20–0.71, p = 0.003) (Fig. S7b, Supplementary Data 6–7). 1p19q codeletion protected against death (aHR=0.19, 95%CI: 0.05–0.66, p = 0.009) and progression (aHR = 0.35, 95%CI: 0.15–0.85, p = 0.020), while NF1 variant predicted death (aHR = 2.50, 95%CI: 1.08–5.79, p = 0.032). In multivariable LGG subtype analyses, the IDH1 wild-type and NF1 variant subtype predicted death (aOR = 7.08, 95%CI: 1.49–37.09, p = 0.015; aHR=15.08, 95%CI: 3.87–58.78, p < 0.001), progression (aOR = 10.12, 95%CI: 1.92–79.21, p = 0.011; aHR = 8.44, 95%CI: 2.82–25.23, p < 0.001), and new tumor events (aHR=6.69, 95%CI: 1.88–23.74, p = 0.003) compared to the 1p19q codeletion and CIC wild-type subtype (Fig. S7a, Supplementary Data 4–5). The IDH1 wild-type and NF1 wild-type subgroup showed significant lower risk of these outcomes and were more likely to remain “with tumor” (aOR = 3.00, 95%CI: 1.18–7.91, p = 0.023), suggesting NF1 variant marks increased risk of severe disease in the already high-risk IDH1 wild-type LGG subtype.
In GA, the partition on EBV+ predicted poor treatment response (aOR = 3.12, 95%CI: 0.92–11.27, p = 0.069) and remaining with tumor (aOR = 4.16, 95%CI: 1.39–13.33, p = 0.012) compared to those who were EBV- among patients with RHOA wild-type (Fig. S5). The EBV biomarker was also significantly associated with neoplasm status (aOR = 4.63, 95%CI: 1.51–15.22, p = 0.008) and poor response to treatment (aOR = 4.31, 95%CI: 1.17–17.62, p = 0.031) (Fig. S7b, Supplementary Data 6–7). ARID1A variant protected against recurrence (aOR=0.29, 95%CI: 0.07–0.91, p = 0.056), poor treatment response (aOR = 0.34, 95%CI: 0.09–1.00, p = 0.069), and remaining with tumor (aOR=0.32, 95%CI: 0.09–0.93, p = 0.053) compared to those with ARID1A wild-type among patients with RHOA wild-type and EBV-. Multivariable subtype analyses showed significantly increased odds of remaining with tumor in the RHOA wild-type and EBV+ subtype (aOR = 3.28, 95%CI: 1.08–10.67, p = 0.039) compared to the RHOA wild-type, EBV-, and ARID1A wild-type subtype (Fig. S7a, Supplementary Data 4–5). Results suggest RHOA wild-type, EBV- and ARID1A variant GA has better prognosis, while the RHOA wild-type EBV+ subtype has poorer prognosis.
In EC, the partition on high CN cluster predicted progression (aOR=8.65, 95%CI: 1.71–157.88, p = 0.038) compared to those with lower CN clusters among patients without MLH1 hypermethylation and with POLE wild-type (Fig. S6). In multivariable biomarker analyses, high CN cluster was associated with increased risk of death (aOR = 4.02, 95%CI: 1.10–26.01, p = 0.070; aHR = 3.59, 95%CI: 0.83–15.51, p = 0.087) and predicted progression (aOR = 4.07, 95%CI: 1.52–14.19, p = 0.011; aHR = 2.97, 95%CI: 1.04–8.45, p = 0.042) (Fig. S7b, Supplementary Data 6–7). While POLE variant had a protective univariable association with progression (aOR = 0.27, 95%CI: 0.04–0.96, p = 0.085), this was not observed in multivariable analyses, suggesting CN cluster and MLH1 hypermethylation status explain some of this association. Multivariable EC subtype analyses also showed significant associations across clinical endpoints (Fig. S7a, Supplementary Data 4–5). The MLH1 non-hypermethylated, POLE wild-type, and low CN subtype protected against progression (aOR = 0.12, 95%CI: 0.01–0.58, p = 0.038; aHR = 0.17, 95%CI: 0.02–1.23, p = 0.079) compared to the MLH1 non-hypermethylated, POLE wild-type, and CN high subtype, the most prevalent EC subtype. The MLH1 non-hypermethylated and POLE variant subtype protected against progression (aOR = 0.15, 95%CI: 0.01–0.80, p = 0.075).
MuTATE models consistently exhibited superior performance compared to CART, resulting in significantly enhanced accuracy, and improved true and false discovery rates, as corroborated by cross-validation analyses of biopsies from three clinical cohorts. MuTATE surpassed CART, highlighting its substantial potential to advance our comprehension of clinical scenarios. Furthermore, the MuTATE molecular models developed for LGG, GA, and EC unveiled innovative molecular signatures, enriching the established clinical models. MuTATE’s partitioning approach and the biomarkers it identifies emerged as independent prognostic indicators of disease, characterizing subtypes associated with multiple clinical endpoints and unveiling previously unexplored phenotype heterogeneity across all cohorts.
Together, these results illustrate how MuTATE not only automates subtype discovery and risk classification, but also exposes molecularly defined subgroups missed by current expert paradigms. These insights may directly inform precision medicine strategies and guide post-operative management in LGG, GA, and EC.
