Inclusion of Multi‐Omic Biomarkers Improves Prediction Accuracy of Response, Relapse, and Overall Survival in Acute Myeloid Leukemia Patients Receiving High‐Intensity Induction Chemotherapy

ABSTRACT

Background

Despite advancements in genetic markers for acute myeloid leukemia (AML) risk stratification, outcome prediction remains challenging due to disease heterogeneity and dynamic genetic changes, highlighting the need for reliable biomarkers to improve AML treatment strategies and patient outcomes. To refine outcome predictions, we investigated the use of microbial‐derived biomarkers to predict composite complete remission (CRc), relapse, and survival for patients on high‐ and low‐intensity regimens, and to integrate those variables into the widely clinically utilized European Leukemia Network (ELN‐2022) genetic risk classification model for high‐intensity‐treated patients.

Methods

We first developed machine learning models that integrate baseline fecal metabolomics, 16S rRNA‐based stool microbiome features, and clinical metadata (sex, antibiotic administration, AML somatic mutations, and cytogenetics) from two cohorts of AML patients (n = 83) undergoing remission induction chemotherapy. Univariate tests and sparse canonical correlation analysis were employed for variable selection and to explore fecal metabolite‐microbe relationships. A robust machine learning approach using XGBoost was employed, with 100 stratified data splits (80% training, 20% testing) and coarse‐to‐fine hyperparameter optimization. Variable importance was aggregated across all models to select key predictors.

Results

For high‐intensity‐treated patients, XGBoost models achieved aggregated AUROC scores of 0.719, 0.729, and 0.65 for CRc, relapse, and overall survival, respectively. For low‐intensity‐treated patients, these models achieved aggregate AUROC scores of 0.945, 0.724, and 0.768 for these same outcomes, respectively. Integrating the biomarkers identified in the high‐intensity machine‐learning models with the current ELN‐2022 AML risk stratification system effectively stratified patients into risk categories, which obtained higher concordance indices and likelihood ratios, demonstrating improved prognostic accuracy for each outcome compared to ELN‐2022 alone.

Conclusions

The inclusion of microbial‐derived biomarkers serves as a robust prognostic tool to improve outcome prediction in AML patients, highlighting the potential of its integration into AML risk assessment and paving the way for personalized treatment strategies and improved patient outcomes.

1. Introduction

Acute myeloid leukemia (AML) represents a heterogeneous group of aggressive hematopoietic neoplasms derived from myeloid precursors [1]. AML displays this heterogeneity in differences in pathophysiology, clinical, cytogenetic, and molecular profiles [2]. Despite high complete remission rates in AML patients treated with induction chemotherapy (60%–70%), relapse remains the primary cause of treatment failure in patients aged over 60 years, where the five‐year survival rate is particularly poor at ~30% [3, 4, 5]. Currently, there are several methods used in the prognosis of AML, most of which are based on patient cytogenetic and genetic factors [6].

The European LeukemiaNet (ELN) risk stratification classification system for adult AML is used to help in the diagnosis and management of AML by providing prognostic discrimination, aiding in determining treatment strategies, and is frequently updated to accommodate advances in the understanding of AML [7]. The most recent update to the ELN‐2022 risk classification system, while effective in younger patients treated with intensive chemotherapy, has shown limitations in older adults (≥ 60 years) receiving lower‐intensity therapies, as it often fails to distinguish between favorable and intermediate risk groups within this subset of patients [8, 9]. Thus, in December of 2024, the Beat‐AML refined risk stratification model was published to address these gaps through the addition of several additional genetic features for the risk stratification of older adults with AML receiving lower‐intensity therapies [9].

Both risk stratification systems are three‐tiered (favorable, intermediate, and adverse) but are limited to stratifying patients based only on cytogenetic aberrations and selected gene mutations. Given the complex nature of AML, the biomarkers serving as the foundation for treatment strategies should encompass its heterogeneity, incorporating variables intricately linked to the disease outcome [10, 11]. Biomarkers such as methylation, gene expression analysis, and immunological biomarkers have been explored. However, these suffer from various pitfalls, including tissue specificity, a limited selection of genomic regions, and a lack of reproducibility [12, 13].

Considering that inflammation and immune dysregulation play crucial roles in the development and advancement of leukemia, it is plausible that changes in gut microbiota composition and the metabolites they produce could contribute to leukemia progression as well as significantly influence treatment outcomes and patient response [14, 15, 16]. Microbial‐derived metabolites, such as short‐chain fatty acids (SCFA), secondary bile acids, and tryptophan metabolites, have been shown to impact cancer development and progression through various mechanisms, including modulation of the immune system and alteration of the tumor microenvironment [17, 18, 19, 20].

The gut microbiome has also been found to be associated with or contribute to the pathogenesis of multiple cancer types, including breast, colorectal, ovarian, lung, and prostate cancers [21, 22, 23, 24, 25]. Recent studies have expanded this understanding to hematologic malignancies, consistently showing profound dysbiosis in conditions such as AML compared to healthy individuals [26, 27, 28, 29]. In leukemia, specific compositional alterations have been identified as potential biomarkers of disease and treatment‐related comorbidities [30, 31, 32]. For instance, a recent study found that having a higher abundance of Faecalibacterium, Ruminococcus, Blautia, and Butyricimonas at diagnosis was predictive of improved hematologic recovery in AML patients [33]. The influence of gut microbiota extends beyond local effects, as studies have demonstrated the ability of microbial metabolites to drive malignant clonal hematopoiesis, as well as support myelopoiesis and lymphopoiesis post‐transplant, highlighting the far‐reaching impact of these microorganisms [20, 34].

Although microbial and metabolite variables have been shown to correlate with clinical outcomes, none are currently utilized in standard risk stratification systems for AML. In this study, we performed an integrated analysis of targeted fecal metabolomics, 16S rRNA amplicon sequencing of baseline stool samples, and clinical metadata collected from the electronic medical records to identify a set of biomarkers that could predict composite complete remission (CRc), relapse, and overall survival. Based on these biomarkers, we developed machine‐learning models that accurately discriminate AML patients who experienced CRc, relapse, and overall survival after induction chemotherapy.

2. Materials and Methods

2.1. Study Design and Participants

Baseline fecal samples were collected from 47 treatment‐naïve AML patients undergoing remission induction chemotherapy (RIC) from September 2013 to August 2015 at MD Anderson Cancer Center in Houston, TX. Aspects of this cohort were previously published, and 16S V4 rRNA sequences were deposited in the NCBI Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra) under the BioProject IDs PRJNA352060 and PRJNA526551 [32, 35, 36]. The study protocol was approved by the MD Anderson Cancer Center Institutional Review Board (IRB # PA13‐0339) and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants before enrollment. Additional baseline fecal samples were collected from a second cohort (PA15‐0780) of 38 AML patients from January 2015 to February 2020. This study followed the same approval, compliance, and consent standards. 16S V4 rRNA sequences were deposited under the BioProject ID PRJNA1124986 and also previously published [37, 38]. For both cohorts, baseline samples were collected on average 1.67 days within the start of RIC, with a range from 10 days before to 15 days after initiation. Eighty‐five patients were included in the final analyses, all of whom met the inclusion criteria of having a baseline stool sample, outcome data, and metaphase cytogenetic information available.

2.2. Specimen Processing, 16S RNA Sequencing, and Metabolomic Profiling by IC‐MS

Genomic DNA was extracted from 50 mg of baseline stool samples, and 16S V4 rRNA sequencing and analysis were performed as previously described [39]. Briefly, the V4 region of the 16S rRNA gene was sequenced using the Illumina MiSeq platform with a 2 × 250 bp paired‐end protocol. Sequences were processed with QIIME2 and VSEARCH, classified with the SILVA database version 138, and used to generate zOTU tables and calculate α‐ and β‐diversity [40, 41, 42]. Targeted metabolomics was performed to isolate the abundances of 10 metabolites (2‐hydroxyglutarate (2‐HG), acetic acid, butyric acid, glyceric acid, ketoleucine, lactate, malate, pentanoic acid, propionic acid, and succinate) in the baseline fecal samples. Approximately 100–200 mg of stool were snap‐frozen in liquid nitrogen, then homogenized with Precellys Tissue Homogenizer. Metabolites were extracted using 1 mL ice‐cold 0.1% ammonium hydroxide in 80/20 (v/v) methanol/water. Extracts were centrifuged at 17,000 g for 5 min at 4°C, and supernatants were transferred to clean tubes, followed by evaporation to dryness under nitrogen. Dried extracts were reconstituted in deionized water, and 10 μL was injected for analysis by ion chromatography (IC)‐MS. IC mobile phase A (MPA; weak) was water, and mobile phase B (MPB; strong) was water with 100 mM KOH. A Thermo Scientific Dionex ICS‐5000+ system included a Thermo IonPac AS11 column (4 μm particle size, 250 × 2 mm) with a column compartment kept at 30°C. The autosampler tray was chilled to 4°C. The mobile phase flow rate was 360 μL/min, and the gradient elution program was: 0–5 min, 1% MPB; 5–25 min, 1%–35% MPB; 25–39 min, 35%–99% MPB; 39–49 min, 99% MPB; 49–50, 99%–1% MPB. The total run time was 50 min. To assist the desolvation for better sensitivity, methanol was delivered by an external pump and combined with the eluent via a low dead volume mixing tee. Data were acquired using a Thermo Orbitrap Fusion Tribrid Mass Spectrometer under ESI negative ionization mode at a resolution of 240,000. Raw data files were imported into Thermo Trace Finder 5.1 software for final analysis. The relative abundance of each metabolite was normalized by sample weight.

2.3. Clinical Definitions and Features

This study focused on three independent outcome groups: overall survival after 2 years, relapse before 1 year, and composite complete remission (CRc; complete remission + complete remission with incomplete hematologic recovery + complete remission with incomplete platelet count recovery, defined by ELN‐2017 standards) [43]. CRc was defined as achieving remission within a 42‐day period, consistent with FDA regulatory standards [44]. However, a single outlier patient who achieved remission on day 46 was included in the CRc cohort based on their clear clinical response, despite occurring slightly outside the conventional timeframe.

Definitions of relapse and overall survival followed the ELN‐2017 standard guidelines [43]. Overall survival was calculated from the start of RIC to the time of death or last known follow‐up. The patient demographics, outcome classifications, chemotherapy type, antibiotic treatment information, and chemotherapeutic responses were collected from electronic medical records. Antibiotic administration was limited to antibiotics administered to the patient on the day of RIC start.

Patients were divided into two groups based on treatment intensity, which was defined according to the type of chemotherapy regimen they received. Patients who received intensive induction chemotherapy (defined as anthracycline plus cytarabine‐based regimens including 7 + 3, FLAG‐IDA, CPX‐351, and other high‐dose cytarabine‐containing intensive protocols) were assigned to the high‐intensity group, while those treated with lower‐intensity therapies (defined as hypomethylating agents, Venetoclax combination therapies, and low‐dose cytarabine regimens) were categorized in the low‐intensity group [9, 45]. Each group was analyzed separately to better account for differences in clinical factors, which would dictate the chosen treatment intensity (age, comorbidities, etc.) or known differences in clinical outcomes related to treatment intensity.

Metaphase cytogenetics was analyzed at MD Anderson, and patients were profiled for mutations in 65 genes commonly mutated in myeloid neoplasms via targeted sequencing from bone marrow or peripheral blood [46, 47, 48]. The variant allele frequency limit of detection was ≥ 2%. Risk categorization was performed for patients receiving high‐intensity RIC using the ELN‐2022 classification model [7]. A similar stratification could not be performed for low‐intensity patients using the Beat‐AML 2024 risk model, as data for the MLL2 mutation were not available for this patient cohort [9].

2.4. Patient Inclusion

Baseline fecal samples were collected from 85 patients. All 85 samples were included in the analysis for the overall survival outcome. Chemotherapy response data within the 42‐day period specified by FDA regulatory standards were available for 83 patients. The relapse outcome analyses included 64 patients, as only those who achieved recovery were eligible for this assessment.

2.5. Sparse Canonical Correlation Analysis

Sparse canonical correlation analysis (sCCA) identified strongly associated metabolite‐microbe groups by finding maximally correlated linear combinations of variables while promoting sparsity through variable‐specific weight thresholding. A centered log‐ratio (CLR) transformation of microbial abundance values was performed, with zero values replaced with a pseudocount of min (relative abundance)/2. sCCA was implemented using the ‘PMA’ (v. 1.2–2) package in R, correlating CLR‐transformed gut microbiome taxonomic abundances and fecal metabolome measurements [49]. Hyperparameters were tuned using the CCA.permute function (nperms = 20, niter = 5) before fitting the model. Lasso penalty was used to obtain the corresponding canonical vector to enforce sparsity by setting the parameters “typex” and “typez” to “standard.”

2.6. Machine Learning Modeling and Optimization

For each of the three outcome groups (overall survival, relapse, and CRc), we developed independent prediction models using a two‐step approach. The initial dataset included baseline metabolites and CLR‐transformed zOTU relative abundances, along with binary clinical variables such as AML patient mutations and karyotypes, sex, and antibiotics currently being administered on the starting day of RIC. Univariate analysis, including Chi‐squared, Fisher's Exact, and Mann–Whitney tests, was applied to screen potential prognosis‐related variables, using p < 0.2 as the cutoff, and these variables were utilized in the initial model‐building step of variable selection.

In the first step, we constructed a Variable Selection model using Extreme Gradient Boosting (XGBoost) to identify potential biomarkers using the R package ‘xgboost’ (v.1.7.5.1) [50]. To minimize variability present in the datasets and to promote the robustness of each outcome model, the modeling procedure was performed on 100 independent 80–20 stratified training–testing splits, and the outcomes were aggregated (Figure S1) [51]. The ‘caret’ package (v.6.0–94) in R was employed for model training and data splitting [52]. Models were trained with 5‐fold cross‐validation with 10 repeats implemented with ‘gbtree’ type booster to mitigate bias. Hyperparameters for each outcome model were tuned using a coarse‐to‐fine optimization method, involving three random hyperparameter searches on five randomly selected seeds for each outcome group (Table S1). Variable importance scores from these 100 independent models were aggregated for each outcome group, and variables with a mean importance score above a 0.03 threshold were selected for the second step.

In the second step, a Final Model was created using the selected variables following the same methodology as the Variable Selection model, where 100 independent 80–20 stratified training and testing splits were made. Hyperparameters for each outcome model were returned using the coarse‐to‐fine optimization method, and results from these 100 XGBoost models were aggregated. A SHAP model was fitted to each aggregate outcome to better understand how each feature contributed to the overall model performance and decision making.

Additionally, we developed a Clinical Model using only the mutations and karyotypes that are utilized in the ELN‐2022 and Beat‐AML risk stratification models (using all available cytogenetic markers) for the high and low‐intensity treated patients, respectively. We compared the performance of the Final Model (including both clinical and microbial variables) with the Clinical‐only Model using aggregated AUROC curves.

2.7. Prognostic Risk Model Construction and Validation

A prognostic risk model was developed for the high‐intensity treatment group by integrating the identified biomarkers from each ML model with the ELN‐2022 risk stratification model. A risk model was not developed for the low‐intensity treatment group as we were missing mutation data for the MLL2 mutation to be able to compare accurately to the Beat‐AML 2024 model. The least absolute shrinkage and selection operator (LASSO) Cox regression model was employed to select the optimal weighting coefficient for each variable. This was utilized to construct a continuous prognostic risk score, calculated using the following expression:

$$\text{Risk score}={\sum }_{i=0}^N\left({\text{Coef}}_i\times {\mathit{V}}_{\mathit{i}}\right)$$

N, Coef, and V represent the variable, correlation coefficient derived by the LASSO Cox regression model, and CLR‐transformed relative abundance of microbes, metabolite measures, or binary classification of clinical variables, respectively.

To define binary high and low microbial biomarker risk groups while avoiding perfect separation in this limited cohort, patients were ordered by descending risk score and the top 52 patients (equal to the total number of observed events) were assigned to the high‐risk group, while the remaining patients were assigned to the low‐risk group. This yields a single deterministic cutoff at a risk score of −0.067 (median = −0.067) for CRc, 1.071 (median = 0.005) for relapse, and 0.0034 (median = 0.0034) for overall survival. All patients with a risk score above the cutoff were classified as high‐risk and all others as low risk. Under this event‐balanced rule, both groups contained at least one event and one censored observation, which stabilized hazard ratio estimation and mitigated perfect separation in the Cox model [53].

The binary high/low microbial biomarker risk classification was then combined with the three‐tier ELN 2022 risk system to construct an integrated three‐tier risk model. The integrated risk classification was constructed thus: Favorable risk included ELN favorable risk patients with low microbial biomarker risk scores. Intermediate risk included ELN favorable risk with high microbial biomarker risk scores, and ELN adverse risk patients with low microbial biomarker risk scores. Finally, adverse risk included ELN intermediate‐risk patients with high microbial biomarker risk scores, and ELN adverse‐risk patients with high microbial biomarker risk scores. Survival curves for these risk groups and for the original ELN‐2022 risk stratification model were summarized using Kaplan–Meier curves generated with the R package ‘survival’ (v.3.5–7). Discriminatory performance was quantified using the concordance index, and model fit was compared using Cox proportional hazards regression, reporting the likelihood ratio statistic and corresponding p‐value for each model.

2.8. Statistical Analyses

All analyses used baseline stool samples with microbial features analyzed at the zOTU level unless stated otherwise. Following taxonomic classification, the microbial features present in less than 10% of patients and at < 0.1% abundance were pruned from the dataset. Microbiome abundance data were analyzed as relative abundances, and β‐diversity analyses used weighted and unweighted UniFrac distance matrices. Microbial and metabolite variables that exhibited a high correlation (> 0.9) with each other based on the Pearson correlation coefficient were removed. All analyses utilized Mann–Whitney U‐tests to compare α‐diversity (Shannon diversity and Observed OTUs) and bacterial taxa relative abundances between outcome groups. Fisher's Exact or Chi‐Squared Tests examined relationships between categorical clinical features and outcome groups. β‐diversity was assessed through PCoA of weighted and unweighted UniFrac distances, and the differences were assessed using PERMANOVA [54]. Unless otherwise stated, all hypothesis testing was two‐sided. For each analysis, the sample size (n) corresponds to the number of patients with available data for the variable and outcome of interest and is reported in the corresponding tables and figure legends. Findings were considered significant if the p‐values were < 0.05. All statistical analyses were performed using R version 4.3.1. For all analyses, patients were stratified into 12 experimental groups defined by treatment intensity (low vs. high) and outcome status (mortality yes/no, relapse yes/no, and CRc response yes/no), and the corresponding sample sizes (n) for each group are reported in Table 1.

TABLE 1.

Clinical characteristics of patients treated with low‐intensity therapies.

	CRc				Relapse				Overall survival
	Total	Yes	p a	Odds ratio	Total	Yes	p a	Odds ratio	Total	Yes	p a	Odds ratio
Patient count, N	38	18			25	11			39	20
Sex, N (%)			1	1.05			0.697	0.677			1	0.878
Female	15	7	—	—	10	5	—	—	15	8	—	—
Male	23	11	—	—	15	6	—	—	24	12	—	—
Time to event in days, median (range)	29.5 (20–40)				176 (91–360)				309 (31–683)
Chemo type			0.288	2.48			0.001	0			0.006	0.107
Hypomethylators	27	11	—	—	16	11	—	—	27	18	—	—
Other (Low)	12	8	—	—	9	0	—	—	12	2	—	—
AML somatic mutations, N (%)
TET2	15	8	0.741	1.47	15	3	0.677	0.514	15	8	1	1.05
RAS	12	6	1	1.16	12	4	1	1.41	12	8	0.307	2.28
DNMT3	10	5	1	1.15	10	4	1	1.03	10	4	0.468	0.509
ASXL1	13	10	0.015	6.69	13	5	1	1.11	13	5	0.307	0.427
FLT3‐ITD	3	1	1	0.538	3	0	0.487	0	3	1	0.595	0.431
NPM1	4	2	1	1.12	4	1	0.604	0.381	4	1	0.328	0.272
IDH2	7	3	1	0	7	3	0.288	4.57	7	5	0.41	2.60
TP53	8	3	0.697	0.608	8	4	0.133	6.84	8	7	0.045	8.68
CEBPA	5	4	0.169	5.20	5	2	1	0.821	5	2	0.652	0.564
RUNX1	5	2	1	0.715	5	0	0.487	0	5	3	1	1.40
IDH1	1	0	1	0	1	0	1	0	1	1	1	Inf
Fluoroquinolone administration, N (%) c	22	11	0.959	0.0027	17	5	0.081	2.93	22	9	0.25	1.32
Complex karyotype, N (%)	6	2	0.663	0.509	6	3	0.288	4.57	6	5	0.182	5.76
ELN2022, N (%) b			0.556	1.17			0.090	4.81			0.022	7.61
Favorable	1	0	—	—	1	0	—	—	1	0	—	—
Intermediate	8	5	—	—	7	1	—	—	8	1	—	—
Adverse	30	14	—	—	17	10	—	—	30	19	—	—

Note: Bolded values are those that are statistically signifiant (p < 0.05).

^a Fisher's exact test, p‐value, and odds ratio.

^b Chi‐squared test, p‐value, and Chi‐2 value are listed.

^c Fluoroquinolone prophylaxis administration on the day of baseline stool sample collection.

3. Results

3.1. Patient Characteristics

Eighty‐five patients, undergoing induction chemotherapy from September 2013 to January 2020, from two AML cohorts with baseline stool samples were analyzed. Of these, 39 received low‐intensity treatment, and 46 received high‐intensity treatment. Tables 1 and 2 present clinical characteristics for three outcome groups: CRc, relapse, and overall survival. Clinical patient characteristics, including chemotherapy type, host somatic mutations, sex, and antibiotic administrations occurring on the day of sample collection, were analyzed as potential factors associated with the three observed outcomes.

TABLE 2.

Clinical characteristics of patients treated with high‐intensity therapies.

	CRc				Relapse				Overall survival
	Total	Yes	p a	Odds ratio	Total	Yes	p b	Odds ratio	Total	Yes	p a	Odds ratio
Patient count, N	45	32			39	8			46	20
Sex, N (%)			0.043	0.21			0.694	1.56			0.769	0.782
Female	26	22			23	4			26	12
Male	20	11			16	4			20	8
Time to event in days, median (range)	28.5 (15–46)				139.5 (80–228)				219 (29–606)
Chemo type b			0.322	0.446			0.218	4.88			0.219	2.52
Fludarabine	16	13			14	1			17	5
Non‐fludarabine (High)	29	19			25	7			29	15
AML somatic mutations, N (%)
TET2	23	16	1	0.86	19	2	0.235	0.283	23	10	1	1
RAS	16	11	1	1.17	14	4	0.424	2.06	16	6	0.756	0.691
DNMT3	15	10	0.733	0.733	12	4	0.221	2.79	15	10	0.0551	4.06
ASXL1	8	3	0.168	0.242	5	1	1	0.965	8	5	0.267	2.50
FLT3‐ITD	14	11	0.724	1.72	14	5	0.109	3.91	14	6	1	0.965
NPM1	15	14	0.034	8.93	14	2	0.686	0.536	15	5	0.365	0.541
IDH2	9	4	1	1.26	6	1	0.583	2.20	9	6	0.267	2.50
TP53	5	3	0.617	0.577	3	0	1	0	5	4	0.151	6.01
CEBPA	7	6	0.654	2.72	6	1	1	0.748	7	2	0.446	0.474
RUNX1	6	0	NA	NA	4	4	NA	NA	6	6	NA	NA
IDH1	8	6	1	1.26	6	2	0.583	2.20	8	5	0.267	2.50
Fluoroquinolone Administration, N (%) c	28	21	0.690	0.160	25	4	0.603	0.270	29	12	0.9466	0.00449
Complex Karyotype, N (%)	12	8	0.721	0.755	12	1	1	0.498	12	7	0.314	2.22
ELN2022, N (%) b			0.25	2.77			0.272	2.61			0.0971	4.66
Favorable	9	8			8	0			9	2
Intermediate	13	10			12	3			13	4
Adverse	23	14			19	5			24	14

Note: Bolded values are those that are statistically signifiant (p < 0.05).

^a Fisher's exact test, p‐value, and odds ratio.

^b Chi‐squared test, p‐value, and Chi‐2 value are listed.

^c Fluoroquinolone prophylactic administration on the day of baseline stool sample collection.

Significant associations between chemotherapy type and patient outcomes were revealed in patients treated with low‐intensity treatment, with hypomethylating agents associated with lower odds of death compared with other low‐intensity regimens (p = 0.006) but higher odds of relapse (p = 0.001; Table 1). Among the genetic variables assessed, ASXL1 mutation was associated with greater odds of achieving CRc (p = 0.015) compared to individuals with wild‐type ASXL1, while TP53 mutation was associated with increased odds of overall survival relative to TP53‐wild‐type individuals in this cohort (p = 0.045).

In high‐intensity treated patients, female sex was significantly associated with a higher odds of achieving CRc compared with males (p = 0.043). Furthermore, NPM1 mutation was likewise associated with higher CRc odds compared with NPM1‐wild‐type disease (p = 0.034; Table 2). Notably, no additional clinical variables demonstrated significant associations with overall survival or relapse in the high‐intensity treatment group.

3.2. Gut Microbiome Profiles of AML Outcome Groups

Two α‐diversity metrics, observed OTUs and Shannon diversity index, were used to compare the differences in microbial diversity within samples between outcome groups. However, neither metric showed significant differences at baseline between patients who experienced CRc, relapse, or overall survival and those who did not in either low or high‐intensity treated patients (Figure 1A; Figure S2A).

FIGURE 1.

Characterizing the fecal microbiome diversity of AML patient outcome groups. (A) Comparison of Shannon index alpha diversity values between affected and unaffected patients in response (CRc), relapse, and overall survival groups among patients treated with low‐intenisty (left) and high‐intensity regimens (right). Box‐and‐whisker plots show the median value with the box denoting the interquartile range. Whiskers indicate the largest and smallest value within 1.5 times the interquartile range. Each dot represents a baseline stool sample from an individual patient. p‐values were derived from Mann–Whitney U tests. Group‐wise sample sizes (n) for all outcome and treatment‐intensity combinations are provided in Tables 1 and 2 and apply to panels (A–C). (B) Beta diversity of the microbiome was compared using Principal Coordinate Analysis (PCoA) plots based on the weighted UniFrac distances for low‐intensity treated patients (left) and high‐intensity treated patients (right). Each dot represents a single baseline stool sample, red represents baseline stool samples from patients who did not experience the outcome, blue represents baseline stool samples of patients who did experience the outcome. For each axis, in parentheses, the percent of variation explained is reported. Ellipses indicate a 95% confidence interval. p‐values are derived from PERMANOVA test. (C) Composition of the bacterial community at the order level for patients who experienced and did not experience the specified outcomes with low‐intensity treated patients on the left, and high‐intensity treated patients on the right. Averages of the relative abundance of the top 20 orders are shown for each of the 6 outcome groups.

β‐diversity analysis using weighted UniFrac distances revealed no significant differences in microbial composition across CRc, relapse, and overall survival in either treatment group (Figure 1B). In contrast, analysis of unweighted UniFrac distances identified a significant difference in microbial composition between patients who achieved CRc and those who did not, specifically within the low‐intensity treatment group (p = 0.023; Figure S2B).

When observing differences among gut microbiome composition at the order level for each outcome group (Figure 1C) among patients receiving low‐intensity treatment, Methanobacteriales abundance was increased in those who experienced mortality compared to the overall survival group. Erysipelotrichales and Clostridiales were more abundant in those who did not experience CRc, whereas Acidaminococcales had higher abundance in those who achieved CRc (Table S2). In those treated with high‐intensity chemotherapy, Enterobacteriales was the only bacterial order found to differ significantly among outcome groups, where patients who did not achieve CRc had a higher abundance compared to those who achieved CRc (Table S3).

3.3. Investigation of Fecal Metabolite and Taxonomic Associations and Signatures Among Outcome Groups

To assess baseline metabolite differences within outcome groups, we performed targeted metabolomics and visualized the log2 fold changes using hierarchical clustering heatmaps for patients receiving low‐intensity and high‐intensity treatments (Figure 2A; Tables S4 and S5).

FIGURE 2.

Fecal microbiome and metabolome associations with outcome. (A) Heatmap representing the Log₂ fold‐change for baseline metabolites for those experiencing each outcome in low‐intensity treatment group (top) and high‐intensity treatment group (bottom). Colors range from blue to red indicating a higher or lower baseline metabolite abundance in each outcome group. Hierarchical clustering was performed on the metabolites, and the dendrogram is shown in the columns. (B–D) Volcano plot illustrating the pattern of baseline microbiome differences in terms of the logarithmic fold‐change in relative proportion of bacterial taxa in patients who experienced CRc (B), relapse (C), and overall survival (D). Green dots represent taxa found to have a significant unadjusted p‐value < 0.05 (green), and blue dots a p‐value of < 0.2. p‐values were derived from Mann–Whitney U‐tests. (E) Heatmap correlogram showing Pearson correlation coefficient between gut microbiota genera and fecal metabolites remaining after sparse canonical correlation analyses in patients who experience low‐intensity treatment (top) and high‐intensity treatment (bottom). Group‐wise sample sizes (n) for all outcome and treatment‐intensity combinations are provided in Table 1 and apply to all panels.

In the low‐intensity treatment group, hierarchical clustering identified two primary metabolite clusters. The first cluster included six metabolites that showed decreased fold changes in patients who relapsed but increased fold changes in patients with overall survival. Notably, this cluster comprised three short‐chain fatty acids: butyrate, pentanoic acid, and propionate. The second cluster contained four metabolites, including 2‐hydroxyglutarate (2HG), which was found to have decreased log₂ fold change in relapse and CRc, and ketoleucine, which had a decreased log₂ fold change in overall survival.

Similarly, the high‐intensity group heatmap also revealed two main clusters. In this group, the first cluster was characterized by four metabolites, 2HG, lactate, malate, and succinate, all of which had a decreased log2 fold change in the relapse outcome and a positive log2 fold change for CRc and overall survival. Interestingly, all other changes across the other metabolites were minimal. Despite these differences, Mann–Whitney U‐tests identified only one statistically significant finding. Among high‐intensity treated patients, those who relapsed had significantly higher levels of baseline acetic acid compared to those who did not relapse (p = 0.025; Figure 2A).

When examining taxonomic patterns, specific zOTUs were significantly associated with each outcome group. Within the low‐intensity treatment group, 12 zOTUs differed in the overall survival group, 26 within the CRc group, and eight within the relapse group (Figure 2B–D; Tables S6–S8). For the high‐intensity group, 15 zOTUs were significant in overall survival, 13 in CRc, and 10 in relapse (Figure 2B–D; Tables S9–S11). Notably, very few zOTUs were shared between treatment intensities within any given outcome. Only zOTU_256 (Eggerthellaceae genus) in CRc and zOTU_225 (Coprococcus genus) in overall survival were common to both treatment groups.

To further explore links between the gut microbiome and fecal metabolites, we performed sparse canonical correlation analysis (sCCA) and visualized the results as heatmaps of pairwise Pearson correlation coefficients (Figure 2E; Tables S12 and S13). The overall correlation between the microbial and metabolite datasets was 0.781 (99% CI: 0.599–0.844) for the high‐intensity group and 0.799 (99% CI: 0.616–0.872) for the low‐intensity group.

Strikingly, in both treatment groups, a single metabolite (pentanoic acid) emerged as the primary driver of the association between the metabolites and microbial datasets. This finding is consistent with the objective of sCCA, which seeks the sparsest set of features that maximally explain the shared variance between the two data sets. Among the taxa identified, several zOTUs belonging to the Eubacterium, Lachnospiraceae, and Blautia genera were positively correlated with pentanoic acid in both treatment groups. Taxa within these genera are recognized producers of SCFA. In contrast, a zOTU mapping to Lactococcus was found to be negatively correlated with high‐intensity treated patients, and one zOTU mapping to Enterococcus was found to be negatively correlated with low‐intensity treated patients.

3.4. Identification of Potential Microbial and Metabolite Biomarkers for Leukemia Clinical Outcomes Using XGBoost Variable Selection

To identify candidate microbial and metabolite biomarkers associated with clinical outcomes, univariate filtering was first performed to exclude features not significantly related to outcomes. For metabolite and microbiome data, Mann–Whitney U‐tests were used (Tables S6–S11), while clinical characteristics were assessed using Chi‐squared or Fisher's Exact tests (Table 1). Variables with p‐value ≤ 0.20 were retained for subsequent machine learning analyses. In the low‐intensity treatment group, 82 variables were selected from univariate statistical analyses for the CRc outcome (81 microbial and one metabolite), 53 for relapse (52 microbial and 1 metabolite), and 61 for overall survival (58 microbial and 3 metabolites). In the high‐intensity group, univariate filtering yielded 55 variables for CRc (53 microbial, 1 metabolite, and one clinical), 38 for relapse (37 microbial and 1 metabolite), and 53 for overall survival (52 microbial and 1 metabolite). Each set of filtered variables was used to build a predictive model for its respective clinical outcome.

For each outcome, an aggregate XGBoost model was constructed from 100 independent runs, and the resulting aggregate AUROC curves were plotted. In the low‐intensity group, relapse had the lowest aggregate AUROC (0.598, 95% CI: 0.0554), while CRc had the highest (0.778, 95% CI: 0.0355), and overall survival was intermediate (0.714, 95% CI: 0.0374) (Figure S3). In the high‐intensity group, CRc again had the highest aggregate AUROC (0.717, 95% CI: 0.0404), followed by relapse (0.702, 95% CI: 0.0583), with overall survival lowest (0.602, 95% CI: 0.037) (Figure S4).

Potential microbial, metabolite, or clinical biomarkers for each outcome were identified based on mean non‐zero importance scores across aggregated models, retaining those with a score above 0.03. In high‐intensity treatment, this yielded 14 variables for CRc, eight for relapse, and 10 for overall survival. For low‐intensity treatment, 9 variables were selected for CRc, 15 for relapse, and 13 for overall survival (Tables S14 and S15).

Across microbial features identified by the classification models, distinct taxonomic patterns emerged between treatment intensities and clinical outcomes. In high‐intensity treatment models, taxa from the genera Ruminococcus, Eubacterium, Bacteroides, and Blautia recurred as variables of importance across multiple outcomes, suggesting their broad relevance to RIC outcomes. Outcomes in the high‐intensity group were further differentiated by taxa such as Acutalibacter, Dorea, and Enterococcus for CRc, and Parabacteroides and Fusicatenibacter for overall survival and relapse, respectively. In low‐intensity treatment models, biomarkers were more taxonomically diverse, with Bacteroides representing a considerable proportion of relapse‐associated features, while Blautia, Eubacterium, and Lachnoclostridium were frequently selected across outcomes. Notably, several metabolites emerged as variables of importance alongside microbial variables. Acetic acid was identified in both relapse and overall survival models for the high‐intensity group, while ketoleucine, pentanoic acid, and succinate were identified in overall survival for the low‐intensity group. Additionally, sex was selected as a potential biomarker in the high‐intensity CRc group.

3.5. Comparison of Biomarkers Against Clinical Variables

To assess the predictive efficacy and clinical relevance of the identified biomarkers against those variables within the ELN‐2022 and Beat‐AML 2025 risk models, a second machine‐learning model was developed using the same XGBoost framework previously described (Figure 1). For each outcome, an XGBoost model was trained using either the identified selection of microbial, metabolite, and clinical variables (combined variable model) or was trained using only those variables included in the published clinical models of ELN‐2022 for patients receiving high‐intensity treatment and Beat‐AML 2024 for those receiving low‐intensity treatment [7, 9]. These variables included karyotype and mutation data. Due to missing data, the Beat‐AML clinical variable XGBoost models lacked data for the MLL2 mutation.

The performance of each aggregated model was determined using an AUROC curve and DeLong test (Table S16). Each of the proposed models trained on microbial, metabolite, and clinical features had a higher AUC compared with models trained only on clinical features present in the published Beat‐AML 2024 and ELN‐2022 models for both low and high‐intensity treatment groups (Figures 3 and 4A–C). For patients receiving low‐intensity treatment, our combined CRc model achieved an AUC of 0.945 compared to 0.645 (p < 2.2e‐16) for the model including clinical variables from the Beat‐AML 2024, while our relapse model reached 0.724 versus 0.464 (p = 1.53e‐11), and the overall survival model reached 0.768 versus 0.659 (p = 6.04e‐05) (Figure 3A–C). Similarly, in the high‐intensity treatment group, our combined feature models yielded higher AUCs for CRc (0.719 vs. 0.583; p = 1.34e‐05) and relapse (0.730 vs. 0.565; p = 1.50e‐04) compared to models trained only on clinical features present in the published ELN‐2022 model, while the overall survival model (0.650 vs. 0.629; p = 0.414) did not significantly differ from models trained on the standard clinical variables (Figure 4A–C).

FIGURE 3.

Clinical comparison XGBoost models for each outcome group in low‐intensity treatment group. (A–C) Aggregate AUROC curves for predicting each of the three outcome groups, CRc (A), relapse (B), and overall survival (C) in the low‐intensity treatment group. The proposed model (black line), trained on microbial, metabolite, and clinical variables identified by the variable selection XGBoost model, is compared to the XGBoost model using standard clinical variables (red line), which was trained using only karyotypic and mutation data included in the published Beat‐AML criteria for low‐intensity treated patients. Shaded regions represent the 95% confidence intervals for each model's ROC curve. The dashed line depicts the predictive performance of chance. (D‐F). SHAP beeswarm plots showing feature importance for the XGBoost model distinguishing CRc (D), relapse (E), and overall survival (F). Each row represents a feature, ordered by average absolute SHAP value (feature importance). Each point on a row corresponds to an individual observation in the dataset; its horizontal position indicates the SHAP value (impact of the feature on the prediction for that instance), while its color reflects the feature value (from low, yellow, to high, purple). Dots to the right (positive SHAP values) indicate that the feature increases the model's predicted probability for the outcome, while dots to the left (negative SHAP values) indicate that the feature decreases the predicted probablity for the outcome. Group‐wise sample sizes (n) for all outcome and treatment‐intensity combinations are provided in Tables 1 and 2 and apply to all panels.

FIGURE 4.

Clinical Comparison XGBoost models for each outcome group in high‐intensity treatment group. (A–C) Aggregate AUROC curves for predicting each of the three outcome groups, CRc (A), relapse (B), and overall survival (C) in the high‐intensity treatment group. The proposed model (black line), trained on variables identified by the variable selection XGBoost model, is compared to the XGBoost model utilizing only standard clinical variables (red line), which was trained using only karyotypic and mutation variables included in the published ELN‐2022 criteria for high‐intensity treated patients. Shaded regions represent the 95% confidence intervals for each model's ROC curve. The dashed line depicts the predictive performance of chance. (D–F). SHAP beeswarm plots showing feature importance for the XGBoost model distinuishing CRc (D), relapse (E), and overall survival (F). Each row repsents a feature, ordered by average absolute SHAP value (feature importance). Each point on a row corresponds to an individual observation in the dataset; its horizontal position indicates the SHAP value (impact of the feature on the prediction for that instance), while its color reflects the feature value (from low, yellow, to high, purple). Dots to the right (positive SHAP values) indicate that the feature increases the model's predicted probability for the outcome, while dots to the left (negative SHAP values) indicate that the feature decreases the predicted probablity for the outcome. Group‐wise sample sizes (n) for all outcome and treatment‐intensity combinations are provided in Table 2 and apply to all panels.

We used SHapley Additive exPlanations (SHAP) analysis to interpret the final XGBoost models and identify the most influential features associated with outcomes in each intensity cohort. SHAP values quantified the contribution of individual features to model predictions, providing detailed insight into model behavior. Within the patients receiving low‐intensity chemotherapy, the variables show promise for integration with the Beat‐AML risk stratification system. Among the nine features in the CRc model, higher abundances of three taxa (Sellimonas_unclass, Parabacteroides_distasonis, and Lachnoclostridium_unclass) were found to be linked to a higher probability of CRc (Figure 3D). Among the other six variables, higher abundances of Coprococcus_unclass, Fusicatenibacter_saccharivorans, Phascolarctobacterium_faecium, Eubacterium_hallii, Blautia_unclass, and Collinsella_aerofaciens were correlated with a lower probability of CRc.

Among the potential biomarkers selected for the relapse model in patients receiving low‐intensity chemotherapy, higher abundances of nine features were found to be associated with a decreased predicted probability of relapse (Lactococcus_lactis, Oscillibacter_unclass, and Parasutterella_excrementihomonis, and six zOTUs mapping to Bacteroides) (Figure 3E). High abundances of the other six features were found to be associated with an increased predicted probability of relapse (Clostridium_aldenense, Ruminococcus_gnavus, Bifidobacterium_gallicum, Anaerostipes_caccae, and two zOTUs mapping to Blautia).

Finally, in the overall survival model for patients on low‐intensity therapy, three features (Lachnoclostridium_unclass, Veillonella_tobetsuensis, and Pentanoic Acid) were associated with decreased predicted probability of overall survival, while the rest except succinate, which was ambiguously associated, was found to be positively associated with overall survival prediction (Schallia_odontolytica, Ketoleucine, Methanobrevibacter_smithii, Parasutterella_excrementihominis, Gemmiger_formicilis, and Clostridium_innocuum, and three zOTUs mapping to Bacteroides; Figure 3F).

Models trained on available Beat‐AML 2024 clinical variables found that Complex karyotype, monosomal karyotype, and Mutated ASXL1, BCOR, EZH2, RUNX1, SF3B1, SRSF2, STAG2, USAF1, and/or ZRSR2, Mutated TP53, and KRAS are significant contributors to the discriminatory ability of models predicting CRc, relapse, and overall survival in low‐intensity treated patients (Figure S5).

In the high‐intensity treatment group, 14 features contributed to the prediction of CRc, where Sex and Ruminococcus faecis exhibited the greatest overall contribution to the CRc model, as reflected by their larger mean absolute SHAP values (Figure 4D). High abundances of five taxa, mapping to Enterobacter_cloacae, Dorea_unclass, Clostridium_aldenense, Akkermansia_muciniphila, and Enterococcus_faecalis, had a positive impact on the prediction of CRc. In contrast, male sex and higher abundances of six taxa (mapping to Ruminococcus_faecis, Bacteroides_caccae, Acutalibacter_muris, Eubacterium_rectale, Flavonifractor_plautii, and Incertae_Sedis_unclass) were associated with a lower predicted probability of CRc. Finally, two taxa, belonging to Oscillibacter_unclass and Bacteroides_thetaiotamicron, showed both positive and negative SHAP contributions across individual cases, suggesting an ambiguous role in outcome prediction. Interestingly, no common variables emerged between high‐ and low‐intensity groups when predicting CRc.

Among the nine variables contributing to the prediction of relapse for patients receiving high‐intensity chemotherapy, high abundances of five taxa (mapping to Bacteroides_thetaiotamicron, Butyricicoccus_faecihominis, Streptococcus_thermophilus, Moryella_unclass, and Anaerostipes_caccae) were found to be correlated with an increased predicted probability of relapse (Figure 4E). The other four variables, acetic acid, Fusicatenibacter_saccharivorans, Bacteroides_vulgatus, and Ruminococcus_faecis, were found to be linked to decreased probability of relapse when in higher abundance. Notably, zOTUs mapping to Bacteroides_thetaiotaomicron and Butyricicoccus_faecihominis were found to be the highest contributors to model decision making. Similar to the low‐intensity group, Anaerostipes_caccae emerged as a predictor that was associated with an increased probability of relapse.

In the overall survival model for patients on high‐intensity chemotherapy, we found only two taxa (Acutalibacter_muris and Parabacteroides_merdae) associated with a decreased probability of overall survival, while five (Incertae_Sedis_unclass, Dorea_formicigenerans, DTU089_unclass, Ruminococcus_gnavus, and Eubacterium_rectale) were found to be associated with an increased probability of overall survival (Figure 4F). The other three taxa had ambiguous associations (Acetic Acid and two zOTUs mapping to Blautia_unclass). Among these, Acutalibacter_muris and Parabacteroides_merdae were two of the highest contributors to model decision‐making.

Within the SHAP analysis of the XGBoost models trained solely on clinical variables (ELN‐2022 for high‐intensity treated patients) we revealed limited feature contribution to model decision making. Of the included clinical variables, only three demonstrated SHAP values greater than zero (Mutated ASXL1, BCOR, EZH2, RUNX1, SF3B1, SRSF2, STAG2, USAF1, and/or ZRSR2, Complex karyotype, monosomal karyotype, and Mutated NPM1, without FLT3‐ITD), suggesting that a small subset of clinical features contributed to prediction of clinical outcomes utilizing XGBoost (Figure S6).

3.6. Development of Risk Stratification and Comparison to ELN‐2022

To further evaluate the prognostic value of each proposed outcome model within the high‐intensity treatment group, final combined models were created based on the ELN‐2022 risk stratification algorithm which included microbial, metabolite, and clinical variables discerned to be variables of importance in the XGBoost model. A similar analysis for low‐intensity treated patients could not be performed due to missing mutation data. Risk scores were calculated as described in Methods and patients were stratified into favorable, intermediate, and adverse risk groups. Comparison of patient assignments revealed that 76%, 69%, and 50% of patients were regrouped from the ELN‐2022 to the proposed CRc, relapse, and overall survival groups utilizing microbial, metabolite, and clinical variables defined by the XGBoost models, respectively (Figure S7). The ELN‐2022 model's effectiveness was assessed, achieving a c‐index (concordance) of 0.556, 0.638, and 0.602 for CRc, relapse, and overall survival outcomes, respectively (Figure 5A–C). The cumulative hazards predicted by the proposed models were determined through univariate Cox proportional hazards tests. Compared with the ELN‐2022 model, all three ELN‐integrated models for CRc, relapse, and overall survival (ELN algorithm models integrating microbial, metabolite, and clinical variables defined in XGBoost models) showed a higher concordance, likelihood ratio, and more significant p‐value than the ELN‐2022 model alone, indicating a stronger ability to stratify patients according to risk (Figure 5D–F). The integrated CRc model showed a likelihood ratio of 13.64 and c‐index of 0.654, compared to 2.22 and 0.556 for the ELN‐2022 risk model alone. The integrated relapse model had a likelihood ratio of 12.75 and a c‐index of 0.806, compared to 4.15 and 0.638 for the ELN‐2022 risk model alone. Similarly, the likelihood ratio of the integrated survival model was 5.73, with a c‐index of 0.631, compared to 3.87 and 0.602 for the ELN‐2022 risk model alone. Overall, demonstrating the potential of integrating microbial and metabolite biomarkers to provide more accurate prognostic information compared to the ELN‐2022 model alone for high‐risk patients.

FIGURE 5.

Comparative analysis of standard ELN‐2022 risk model versus integrated ELN‐Microbiome model for high‐intensity treated patients. CRc (A), relapse (B), and overall survival (C) probability of high‐intensity treated patients classified according to the ELN‐2022 genetic risk classification model. Kaplan–Meier analysis for AML patients classified by the proposed integrated ELN‐microbiome model system for CRc (D), relapse (E), and overall survival (F). Kaplan–Meier curves are shown for each group of patients; C‐statistics are based on Cox proportional hazards models. Group‐wise sample sizes (n) for all outcome and treatment‐intensity combinations are provided in Tables 1 and 2 and apply to all panels.

4. Discussion

Accurate prediction of patient outcomes is fundamental to optimizing therapy in AML, a disease characterized by remarkable heterogeneity in onset, progression, and treatment response. Current approaches such as the ELN‐2022 and Beat‐AML 2024 recommendations only focus on genetic alterations, potentially not capturing AML's complex nature. There is a growing recognition of the need to incorporate multi‐omic data and other co‐variates to better reflect disease diversity in prognostic models. This study addresses this knowledge gap by establishing a machine‐learning model integrating clinical factors, fecal metabolites, and gut microbiome data to predict CRc, relapse, and overall survival in AML patients who are treated with high‐ and low‐intensity treatments. Our approach advances risk stratification by offering a multi‐omic view of prognostic biomarkers. The resulting models, combined with the ELN‐2022 recommendations for high‐intensity treated patients, demonstrate remarkable efficacy in accurately categorizing AML patients into distinct risk groups, potentially paving the way for more personalized and effective treatment strategies.

While traditional statistics excel at identifying individual variable relationships, complex diseases like AML require more sophisticated approaches. Machine learning offers a solution by providing the capability to analyze the entire structure of variables collectively, while simultaneously identifying correlations between these variables and the disease state. This approach allows for a more comprehensive understanding of intricate systems and their outcomes. This study optimized AML risk stratification models by integrating multi‐omic biomarkers identified through an XGBoost model pipeline developed by Topcuoglu et al. in 2020 [51]. We analyzed gut microbiome composition across outcomes and found significant differences in specific families and zOTUs, despite little baseline α‐ and β‐diversity differences (Figures 1 and 2; Figure S2; Tables S6–S11). However, given the high dimensionality of microbiome and metabolomic features relative to the sample size, there is an inherent risk of model overfitting despite the use of internal resampling approaches, and these findings should be interpreted with appropriate caution.

Our analyses revealed pronounced differences in baseline fecal metabolite abundances associated with clinical outcomes in low‐intensity versus high‐intensity treatment groups. Distinct clustering and log2 fold change patterns (Figure 2A) indicate that baseline metabolite profiles differ by treatment intensity. Where, in the low‐intensity group, certain metabolites such as succinate and acetic acid were elevated among patients with favorable outcomes, whereas metabolites like 2HG and ketoleucine were reduced. The high‐intensity groups exhibited slightly different clustering, with metabolites such as 2HG, succinate, lactate, and malate more strongly associated with the positive outcomes of CRc and overall survival. These differences should also be interpreted in the context of temporal heterogeneity in treatment (2013–2020), during which evolving chemotherapy protocols and supportive care practices may have influenced both metabolic profiles and clinical outcomes, representing a potential confounding factor.

In the biomarkers identified by the machine learning models, acetic acid emerged as a significant biomarker for overall survival and relapse in high‐intensity treatment patients (Figure 4; Table S15). In contrast, the low‐intensity models identified ketoleucine, pentanoic acid, and succinate as potential biomarkers within the overall survival model (Figure 3; Table S14). While the impact of metabolites on cancer progression and treatment efficacy has been extensively studied, the distinct differences in metabolite levels between treatment intensities have not been well documented, particularly in AML, a disease characterized by highly divergent therapeutic approaches [14, 55, 56, 57]. Interestingly, increased levels of acetic acid were associated with favorable outcomes in the high‐intensity treatment group. This observation aligns with prior reports showing that acetate can reduce cell proliferation in colon cancer cells through suppression of glycolytic metabolism, and that elevated fecal acetate concentrations correlate with longer progression‐free survival in patients with solid tumors treated with PD‐1 inhibitors [58, 59]. While these parallels are biologically suggestive, the associations observed in this study are correlative and do not establish a causal role for these metabolites in AML outcomes. Together, these findings suggest a potentially beneficial role of metabolites in modulating cancer response and highlight the need for further investigation into their mechanistic contribution within AML treatment paradigms.

Divergent microbial signatures emerged clearly when comparing high and low‐intensity treatment groups, highlighting the impact of chemotherapy intensity on gut ecosystem structure in AML patients. In high‐intensity cohorts, taxa such as Enterobacter, Dorea, Clostridium, Akkermansia, and Enterococcus were associated with the predicted probability of CRc, while a separate set, Ruminococcus, Bacteroides, Acutalibacter, Eubacterium, and Flavonifractor, were linked to decreased likelihood of the outcome. Conversely, in low‐intensity treatment groups, genera traditionally implicated in mucosal repair and immune modulation, including Sellimonas, Parabacteroides, and Lachnoclostridium, displayed potentially protective associations with CRc [60, 61, 62]. These SHAP‐derived associations should be interpreted as indicators of model behavior and predictive contribution rather than direct evidence of biological effect. These observations suggest that chemotherapy intensity not only alters microbial community composition but also modulates the prognostic relevance of individual taxa, with high‐intensity regimens favoring profiles associated with dysbiosis and potential pathogen expansion and low‐intensity regimens maintaining a greater proportion of health‐associated commensals.

Recurrent identification of SCFA‐producing taxa across multiple outcome models supports the role of microbial metabolism in shaping AML prognosis. Genera such as Bacteroides, Anaerostipes, Eubacterium, and Lachnoclostridium, recognized for their capacity to produce SCFAs like acetate and butyrate, were repeatedly implicated as key features in all models [63, 64, 65]. The directionality of their associations, sometimes differing depending on the treatment group or outcome, points to complex, context‐dependent contributions of these microbes. Nevertheless, their prominence underscores the importance of fermentative microbial metabolism and its metabolic products as modulators of host response to both disease and therapy. This finding aligns with a study that reported that intensive chemotherapy reduces the abundance of SCFA‐producing bacteria and fecal SCFA levels, an effect associated with gastrointestinal toxicity, loss of barrier integrity, and poorer overall outcomes in hematologic malignancies [14].

Due to incomplete mutational data for the genetic markers incorporated in the Beat‐AML 2024 risk model, an accurate estimation of the impact of the addition of microbial‐derived biomarkers into this model could not be made. To approximate this impact, we compared machine learning models predicting the outcomes to those trained using only features from the ELN‐2022 risk stratification model or all (but one feature) from the Beat‐AML 2024 risk stratification model. Across both high and low‐intensity groups, all microbial and metabolite‐integrated models demonstrated an enhanced ability to predict all three outcomes compared to the models trained on only standard clinical features (Figures 3 and 4).

The integration of microbial, metabolite, and other clinical variables with the ELN‐2022 risk stratification model outperforms the ELN‐2022 model in terms of both the likelihood ratio test and the concordance index for all three outcomes (Figure 5). The significant improvement in both tests suggests that the integrated model is more effective at predicting patient outcomes and has better prognostic significance. For overall survival and CRc, the integrated models demonstrate a stronger ability to stratify according to risk; however, when looking at CRc, the integrated model does not effectively differentiate the adverse risk group, highlighting a potential area for model improvement or the need for additional data.

In this study, certain limitations need to be considered. First, no external, publicly available dataset matched the disease type and included all variables used here, limiting our ability to perform external validation. As a result, all performance metrics, including those for the integrated ELN‐microbiome models, were derived from internal resampling, which may overestimate model performance and limit generalizability. Additionally, this dataset was retrospective, where certain demographic indicators, such as age, race, and diet, which would have been beneficial to include, were not available. Furthermore, the study cohort spans patients treated between 2013 and 2020, a period during which evolving clinical practices may introduce temporal heterogeneity that could affect the results. Although the sample size was sufficient for model development and statistical analysis, it remains relatively small, especially when subdivided by treatment intensity and outcome, and is therefore a significant limitation for applying complex machine learning models to high‐dimensional microbiome and metabolomic data. This imbalance between feature dimensionality and cohort size further increases the risk of overfitting, even with cross‐validation strategies. As such, the findings should be interpreted as exploratory and hypothesis‐generating rather than definitive. Despite these limitations, our findings highlight the importance of incorporating a wide range of biological data, including microbial composition and fecal metabolites, to enhance the predictive accuracy of AML clinical outcomes.

In conclusion, this study proposes a multi‐omic approach to biomarker identification and risk stratification, integrating microbial and metabolite biomarkers to better predict overall survival, relapse, and CRc in AML. However, before clinical implementation, these models will require rigorous external validation in independent AML cohorts that capture diverse patient populations and treatment regimens. Such validation should include prospective, multi‐center studies with standardized protocols for sample collection, microbiome and metabolome profiling, and clinical annotation to ensure reproducibility. Model performance should be evaluated not only for discrimination but also for calibration and clinical utility across distinct subgroups, including different treatment intensities and risk categories. Ultimately, demonstration of consistent performance and added predictive value over existing clinical risk frameworks (ELN‐2022, Beat‐AML 2024) across cohorts will be necessary before these models can be considered for integration into clinical decision‐making. Moreover, further investigation into the functional implications of these microbial signatures could provide valuable insights for developing microbiome‐based prognostic tools or therapeutic interventions. In summary, these models offer the potential to enhance current prognostic assessments and guide individualized therapy decisions in AML.

Author Contributions

Tomo Hayase: data curation, formal analysis. Samantha Franklin: conceptualization, investigation, methodology, writing – original draft, writing – review and editing, visualization, formal analysis. Jessica Galloway‐Peña: conceptualization, funding acquisition, investigation, writing – original draft, writing – review and editing, methodology, supervision, project administration. Robert R. Jenq: supervision, resources. Phillip Lorenzi: resources, formal analysis, investigation. Ivan Ivanov: formal analysis, investigation, methodology. Tapan Kadia: supervision, resources. Sai Prasad Desikan: data curation, investigation. Eiko Hayase: data curation, formal analysis. Samuel Shelburne: supervision, resources, funding acquisition. Pranoti Sahasrabhojane: data curation, project administration, investigation. Chia‐Chi Chang: data curation, formal analysis. Jayastu Senapati: data curation, investigation.

Funding

This work was supported by the Division of Intramural Research, National Institute of Allergy and Infectious Diseases, K01AI143881.

Ethics Statement

The study protocols were approved by the MD Anderson Cancer Center Institutional Review Board (IRB # PA13‐0339 and PA15‐0780) and were conducted in accordance with the Declaration of Helsinki.

Consent

Written informed consent was obtained from all participants before enrollment.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: XGBoost hyperparameter search space explored with the optimized search space for each outcome group.

Table S2: Top 20 order abundances between outcome groups for the Low‐Intensity Treatment Group.

Table S3: Top 20 order abundances between outcome groups for the High‐Intensity Treatment Group.

Table S4: Raw Metabolite Measurements for low‐intensity treated patients.

Table S5: Raw metabolite measurements for high‐intensity treated patients.

Table S6: Significant zOTUs for Overall Survival for the Low Intensity Groups

Table S7: Significant zOTUs for CRc for the Low Intensity Groups.

Table S8: Significant zOTUs for Relapse for the Low Intensity Groups

Table S9: Significant zOTUs for Overall Survival for the High Intensity Groups

Table S10: Significant zOTUs for CRc for the High Intensity Groups

Table S11: Significant zOTUs for Relapse for the High Intensity Groups

Table S12: Comparison correlation values for low‐intensity treated patients.

Table S13: Comparison correlation values for high‐intensity treated patients.

Table S14: Variable importance scores of the top 15 features for each outcome model aggregated from 100 independent Variable Selection XGBoost models for low‐intensity models.

Table S15: Variable importance scores from the top 15 features for each outcome model aggregated from 100 independent Variable Selection XGBoost models for high‐intensity models.

Table S16: Test results for AUC and DeLong's test in comparing XGBoost models including proposed microbial and metabolite variables with standard clinical variables models.

Figure S1: XGBoost pipeline.

Figure S2: Fecal microbiome diversity of AML patient outcome groups.

Figure S3: ROC curve for XGBoost variable selection models for each outcome variable for low‐intensity treated patients.

Figure S4: ROC curve for XGBoost variable selection models for each outcome variable for high‐intensity treated patients.

Figure S5: SHAP summary plots for XGBoost models predicting clinical outcomes based on Beat‐AML 2024 risk stratification variables.

Figure S6: SHAP Summary plots for XGBoost models predicting clinical outcomes based on ELN‐2022 risk stratification variables.

Figure S7: Distribution of high‐intensity treated patients across risk groups.

Data Availability Statement

16S Sequencing data are under BioProject ID accession numbers: PRJNA352060, PRJNA526551, and PRJNA1124986. Metabolomic data may be found in a data supplement.