High temporal resolution prediction of mortality risk for single AML patient via deep learning

Summary

Acute myeloid leukemia (AML) is highly heterogeneous, necessitating personalized prognosis prediction and treatment strategies. Many of the current patient classifications are based on molecular features. Here, we classified the primary AML patients by predicted death risk curves and investigated the survival-directly-related molecular features. We developed a deep learning model to predict 5-year continuous-time survival probabilities for each patient and converted them to death risk curves. This method captured disease progression dynamics with high temporal resolution and identified seven patient groups with distinct risk peak timing. Based on clusters of death risk curves, we identified two robust AML prognostic biomarkers and discovered a subgroup within the European LeukemiaNet (ELN) 2017 Favorable category with an extremely poor prognosis. Additionally, we developed a web tool, De novo AML Prognostic Prediction (DAPP), for individualized prognosis prediction and expression perturbation simulation. This study utilized deep learning-based continuous-time risk modeling coupled with clustering-predicted risk distributions, facilitating dissecting time-specific molecular features of disease progression.

Graphical abstract

Highlights

• •Continuous time series survival model depicts dynamic risk of individual patient
• •Cluster patients by mortality risk curves and identify robust prognostic biomarkers
• •Subgroup with exceptionally poor prognosis identified from ELN2017 Favorable class
• •Web tool predicts death risk and simulates gene expression impact on prognosis

Artificial intelligence; Health sciences; Risk stratification

Introduction

Acute myeloid leukemia (AML) heterogeneity results in divergent pathogenesis, prognosis, and response to therapy, necessitating tailored treatment strategies. Accurate classification of AML patients is therefore critical to guide appropriate clinical management. Clinically, AML classification has primarily relied on genetic abnormalities. The European LeukemiaNet (ELN) 2017/2022 stratified patients into three risk groups (Favorable/Intermediate/Adverse) based on different genetic abnormalities.^1,2 However, this grouping was insufficient to depict the high heterogeneity of AML. The Cancer Genome Atlas (TCGA) grouped mutated genes into nine categories in a cohort of 200 de novo AML patients.³ A larger cohort of 1,540 AML patients was then grouped into 11 groups based on the co-occurrence patterns of genetic abnormalities (fusions and mutated genes), though around 15% of patients still lacked proper classification.⁴ Thus, information beyond genetics is required to categorize AML patients.

Gene expression profiling provides an unbiased overview of the transcriptional landscape by measuring the expression levels of thousands of genes in a single assay. This comprehensive profiling enhances the ability to capture a complete snapshot of the disease state. However, deriving biological insights from the vast complexity of expression data remains challenging. Advanced analytics are key to extracting signals relevant to pathogenesis from transcriptional noise.

Based on expression, researchers have developed signatures related to stemness and immune effector dysfunction to stratify AML prognosis.^5,6,7 Further integration of cytomolecular and leukemic stem cell signatures improved prognostic stratification.⁸ Another study used expression data to accurately predict AML cytogenetic risk via decision tree modeling.⁹ These demonstrated the potential of expression profiles for prognostic stratification. Other efforts had focused on unraveling the intricate molecular subtypes and pathogenic mechanisms driving AML. One study identified 14 gene expression clusters using WGCNA, exploring associations with mutations and drug sensitivity.¹⁰ Another recent study defined eight stable transcriptomic subtypes of AML, complementing previous genome-based classifications.¹¹ These analyses based on expression profiles revealed molecular mechanisms associated with prognosis. However, directly grouping samples by distinct prognostic features and extracting expression profiles for the prognostic groups can more directly identify prognosis-associated molecular mechanisms. In a recent study focusing on high-grade serous ovarian cancer, researchers examined the patients who survived more than 10 years and identified molecular characteristics of multiple alterations in DNA repair-related genes and more frequent somatic variants for the long-term survivors.¹² On the other hand, the prognosis for an individual patient is a composite of multiple independent risk factors over the interval from diagnosis to the latest follow-up. A risk distribution across the time interval enables us to identify molecular mechanisms associated with dynamic time-specific risks.

Here, we present a new classification of AML patients based on the predicted death risk distribution (derived from predicted survival probability distribution) to explore molecular mechanisms associated with various risk distributions. To obtain more refined death risk distributions, we developed a time-continuous survival probability model using deep learning approaches. We analyzed each risk group with expression and mutation profiles to explore key molecular mechanisms underlying the different death risk groups (Figure 1). Based on the death risk classification, we identified two biomarkers, the C1/C2 signatures, demonstrating stable and significant prognostic stratification capabilities. The C1 cluster identified a subset of patients with an exceptionally unfavorable prognosis within the ELN2017 Favorable group. To facilitate individualized risk prediction and simulate the impact of gene expression modulation on prognosis, we designed a web tool known as DAPP.

Figure 1.

Overview of the study design and methods

(A) Schematic of continuous-time series survival model.

(B) Schematic of evaluation of model performance.

(C) Schematic of risk of death curves clustering.

(D) Schematic of website DAPP.

Results

Time period-specific survival-related genes in AML

To explore time period-specific survival-related gene expression profiles, we collected gene expression at diagnosis and prognosis information from 1,668 primary AML patients across 6 published datasets for analysis. Additional clinical feature statistics for the six datasets are presented in Table S1.

Samples from the TCGA AML and BEAT AML 2018 datasets contained mutation information. We used LASSO (least absolute shrinkage and selection operator)^13,14 (family = binomial) for feature selection 10 times with the survival status at each monthly time point from 2 to 60 months as the response variable. This generated LASSO-selected survival-related features for each time point. We then analyzed how frequently each gene was selected as a predictive feature by LASSO across the 6-month bins (Figure 2A).

Figure 2.

Time-specific survival feature and evaluation of continuous-time series survival model

(A) Scatterplot of time-specific enrichment of survival-associated features, with the x axis representing months since diagnosis, size representing -log10 transformed p values, and color indicating median hazard ratio from Cox models in 6 datasets. Top 5 enriched genes per interval shown. p value was calculated using Fisher’s exact test with the alternative hypothesis set to “greater.”

(B) AUC statistics of 8 machine learning methods, with the x axis representing months 2–60, and the y axis showing mean AUC from 10-fold cross-validation per method and time point.

(C) Comparison of AUC from machine learning models using raw predictions versus monotonically decreasing spline smoothed predictions.

(D) AUC for AutoGluon models using cross-validation and independent datasets. For independent validation, only months with a sample size greater than 10 for either survival or death outcomes were retained for analysis.

We found that the same survival-related features were enriched in adjacent time bins, such as BSPRY being significantly enriched in months 24–30, 30–36, and 36–42. In contrast, no significant overlap was observed between distant bins. Among genes enriched in the first 6 months, IL1R2 has been previously reported to be associated with poor prognosis in AML,¹⁵ while DSG2 was negatively correlated with differentiation of human pluripotent stem cells.¹⁶ Our analysis indicates their unfavorable effects on survival were mainly concentrated within the first 6 months after diagnosis. In previous studies, high expression and low methylation of F2RL1 were associated with poor prognosis in AML patients.¹⁷ Our analysis further revealed that F2RL1 expression was enriched in influencing long-time (5-year) survival of AML patients. Since 5-year survival status reflects cumulative death risk over 5 years, the detrimental impact of high F2RL1 expression likely either broadly distributes over the 5 years or specifically concentrates in year 5, leading to lower survival rates in year 5.

Compared to the features related to overall survival, features associated with the distribution of death risk across time may better reveal molecular mechanisms underlying different developmental stages of AML. Death risk is the decrement of survival probability per unit of time, and accurate death risk values require first building an accurate continuous-time survival probability prediction model.

Building a continuous-time survival probability model

We developed a continuous-time survival probability model using deep learning approaches to obtain accurate continuous survival probability predictions for AML samples. We performed survival probability prediction from 2 to 60 months after diagnosis at monthly intervals, selecting genes most relevant for survival using LASSO Cox regression, and building prediction models using 8 machine learning methods separately for each of the 59 time points. We smoothed the predicted survival probabilities for each sample across all the time points using monotonically decreasing spline fitting to obtain a complete survival probability curve over 5 years.

We performed 10-fold cross-validation on the 1,668 samples with gene expression data, evaluating prediction performance across the 8 machine learning methods (Figures 1B and 2B). Smoothing improved performance for AdaBoost, DecisionTree, MLP, RandomForest, and KNeighbors (Figure 2C). Smoothing did not substantially change the results for AutoGluon and LinearSVM, which already had good prediction performance without smoothing. Meanwhile, the smoothing process led to decreased performance for the NaiveBayes model. Among all the methods, AutoGluon showed the best prediction accuracy. Thus, we selected the AutoGluon for subsequent analyses (Figure 2B). In two independent validation sets, we calculated the AUC for the AutoGluon model and observed that its measurement accuracy closely paralleled that achieved during cross-validation (Figure 2D).

Clustering death risk curves

The continuous-time survival probability model yielded complete survival probability curves over 5 years. The survival probability at a given time point reflects the cumulative death risk from diagnosis up to that particular time point. To obtain time period-specific death risk-related genes, we calculated death risk curves as the decrement in survival probability per unit time (10 days) (Table S2).

We converted the survival probability curves over 5 years for the 1,668 samples to death risk curves and then used unsupervised clustering to group samples into 7 clusters ranked by median survival (C1–C7) (Figures 1C, 3A–3C, and S1). C1 showed the most severe death risks, with risk peaking immediately after diagnosis. Death risk peaked around 6 months after diagnosis for C2, and years 1, 1.5, and 2 after diagnosis for C3, C4, and C5, respectively. Mortality exceeded 95% within 3 years after diagnosis for C1–C5. C6 exhibited a relatively broad peak of death risk that was distributed over an extended period of time around year 4 after diagnosis. C7 lacked a distinct death risk peak within 5 years after diagnosis and had the most favorable prognosis among all clusters (Figure 3D).

Figure 3.

Clustering of death risk curves

(A) Consensus matrices of 1,668 patients.

(B) Forest plot of median survival time with 95% CI.

(C) Kaplan-Meier plot for overall survival among patients in 7 death risk clusters.

(D) Individual death risk curves by cluster, with dashed lines indicating the median per cluster. p values for the Kaplan-Meier plots were computed using the log rank test.

Clinical characteristics of the death risk clusters

To investigate the clinical characteristics of the seven death risk clusters, we analyzed several clinical features, including FAB classification, ELN2017 Cytorisk classification, chemotherapy response, tumor purity, and treatment outcomes across these clusters (Figures 4 and S2). Based on the ELN2017 classification, we observed a successive decrease in the percentage of samples classified as Good or Favorable from C7 to C3 (Figures 4A and 4B), consistent with a corresponding deterioration in the prognoses of the clusters. However, the Favorable percentage for C1 and C2 unexpectedly increased, particularly for C1, despite their poor response to chemotherapy (Figure 4C), aligning with their markedly unfavorable survival outcomes. Upon segregating samples from C1 within the Favorable/Good categories, we discovered a subgroup with extremely poor prognoses (Figures 4D and 4E). This phenomenon was further validated in the independent BEAT_AML_2022 dataset (Figures 4F and S3). Consequently, we posit that the stratification provided by ELN2017 could be improved with the incorporation of C1. Additionally, we performed a multivariable Cox analysis for overall survival, incorporating C1–C7 classifications along with various clinical features such as age and gender, as well as prognostic indicators like LSC17 (a 17-gene leukemia stem cells score)⁵ and ELN2017 (Figure S4). The results demonstrated the independent prognostic value of the death risk clusters.

Figure 4.

Characteristics in clustered patients

(A) Cytorisk for patients in the TCGA AML dataset.

(B) ELN2017 for the BEAT AML 2018 dataset.

(C) Response to chemotherapy for GSE106291.

(D and E) Kaplan-Meier plot of overall survival among patients categorized by ELN2017 classification (Favorable class is divided into two categories, Favorable in C1 or not) in the BEAT AML 2018 and TCGA AML datasets.

(F) Kaplan-Meier plot of overall survival among patients categorized by ELN2017 classification (Favorable class is divided into two categories, Favorable in predicted C1 or not) in the independent BEAT_AML_2022 dataset. p values for the Kaplan-Meier plots were computed using the log rank test.

Mutation analysis of death risk clusters

To examine mutation features within each cluster, we first selected patients with mutation samples (n = 335 from TCGA AML and BEAT AML 2018) and used Fisher’s test to identify enriched mutated genes within each cluster (Figure 5A). The clusters with the poorest prognosis, C1 and C2, were enriched for TP53 mutations, with C1 also enriched for ASXL1 and RUNX1 mutations and C2 for del(5) and del(17). TP53 significantly co-occurred with del(5) and del(17). Among patients with TP53 mutations, those in C1 demonstrated significantly poorer prognoses compared to those in C2, aligning with the overall survival trends of C1 and C2 (Figure 5B). C2 had a higher proportion of del(5) mutations among TP53-mutated cases (Figure 5C). There were no significant differences in age and gender demographics between patients with TP53 mutations in C1 and C2 (Figures S5A and S5B). Differential expression analysis between C1 and C2 patients with TP53 mutations revealed higher expression levels of mitochondrial apoptosis genes in C1 (Figures 5D and 5E). For clusters C3–C5 with death risk peaks shifted progressively later, C3 was enriched for TET2 and SRSF2 mutations, C4 for IDH2 and NRAS mutations, which were mutually exclusive, and C5 for PTPN11 mutations. The cluster associated with good prognosis, C6, showed enriched mutations in CBF-AML, KIT, and CEBPA, with KIT significantly co-occurring with inv(16) and t(8; 21), which was consistent with previous studies.^{18,19,20,21,22} C7, the cluster with the most favorable prognosis, was enriched for the t(15; 17) translocation and corresponding PML-RARA fusion gene. Since effective therapeutic strategies exist for AML with PML-RARA fusion, it is associated with better survival. Notably, clusters C1–C5 all had higher FLT3 mutation frequencies but with death risk peaks shifted progressively later.

Figure 5.

Mutation profiles of risk curve clusters

(A) Left: percentage of mutations per gene in TCGA AML & BEAT AML 2018 cohorts (n = 335), with asterisks indicating significant enrichment in the cluster (∗p value<0.05, ∗∗p value<0.01, ∗∗∗p value<0.001). Right: p values for co-occurrence and mutual exclusivity between mutations.

(B) Kaplan-Meier curves for overall survival of TP53-mutated patients in clusters C1 and C2.

(C) Mutation heatmaps of TP53-mutated patients in clusters C1 and C2.

(D) Differentially expressed genes between TP53-mutated patients in clusters C1 and C2.

(E) Enrichment analysis for genes with higher expression in TP53-mutated cluster C1 patients.

Unveiling signature patterns of the death risk clusters

To unveil distinctive features within each cluster, we first identified marker genes based on the training datasets (Table S3). Subsequently, the enrichment of the activity of these cluster signatures in various cell types in AML, including malignant and immune cells, was examined using the published AML single-cell dataset GSE116256 (Figures 6A–6C and S6).²³ C1 demonstrates heightened activity in cDC-like/Mono-like/ProMono-like malignant cells and normal Mono/ProMono cells, whereas C2 displays elevated signature scores in hematopoietic stem cells (HSCs) and HSC-like/progenitor-like tumor cells. A notable correlation exists between the activities of C2 and LSC17 across various cell types.⁵ LSC17 demonstrates a preference for HSC-like tumor cells, whereas the C2 signature favors healthy HSC (Figure 6D). Clusters C3–C5 exhibited activity in B/T/natural killer cells of the lymphoid lineage, while C6 showed activity in HSC-like/GMP-like tumor cells. C7 demonstrated high activity in plasma/GMP/GMP-like cells. The aforementioned results indicated that the poorly prognostic C1/C2 clusters were associated with two distinct malignant cell clusters, with C2 characterized by high stemness, and C1 encompassing multiple AML leukemia cell types, including cDC-like, ProMono-like, and Mono-like malignant cell types.

Figure 6.

Signatures for cluster C1–C7

(A) Uniform manifold approximation and projection (UMAP) for AML single-cell dataset GSE116256.

(B) Percent of origin cell types labels for malignant cells in Tisch.

(C) Mean signature score for cluster C1–C7 in each cell type.

(D) Scatterplot for mean signature score of C2 and LSC17. Cell types with LSC17 score >0 in more than half of their cells are shown. Pearson correlation coefficient (R) was calculated and the p value determined by t test.

(E) Heatmap plot of correlation between C1/C2/cell-type-specific signatures in the three validation datasets.

(F) Kaplan-Meier estimates of overall survival in patients from three independent datasets with C1/C2 signature above and below the optimal cut point. The cut point was determined using the maxstat package in R, and survival curves were compared using a log rank test.

(G) Enrichment analysis for cluster C1/C2 marker genes. p values for the Kaplan-Meier plots were computed using the log rank test.

Further analysis of the validation datasets (GSE71014, GSE10358, and BEAT_AML_2022, n = 307) examined the correlation between C1/C2 signatures and various cell-type-specific signatures (Figure 6E; Table S4). Consistent with the findings mentioned earlier, C1 exhibited a robust correlation with cDC-like/Mono-like/ProMono-like malignant cells, while C2 consistently demonstrated a strong correlation with HSC and HSC-like/Prog-like signatures. Interestingly, C1/C2 showed superior survival stratification compared to individual cell-type-specific signatures (Figure 6F and S7). The C1/C2 signatures exhibited a more significant and robust ability to stratify survival than individual AML cell-type-specific signatures. This suggested that the C1/C2 signatures represented a core molecular feature that is associated with prognosis more directly.

In-depth functional enrichment analysis revealed intriguing associations of these marker genes with biological processes (Figures 6G and S8). C1 was linked to ketone and lipid biosynthesis, while C2 showed associations with angiogenesis and endocytosis.

Development of a web prediction tool

We developed DAPP (https://xgaoo.github.io/DAPP), a web server application for prognosis prediction using the continuous-time survival probability model. A total of 1,668 expression samples and 1,647 mutation samples were collected. Expression-based and mutation-based models were built separately. DAPP provides survival prediction and death risk prediction. Users can input the expression matrix or mutations of an individual patient to obtain continuous survival probability and death risk curves over 5 years for the patient.

DAPP enables users to simulate the effect of gene expression perturbation on the patient’s prognosis using the expression samples. Users can load their own samples or the samples in DAPP and modify the expression values of genes of interest to predict the patient’s prognosis. By comparing to the original samples, the impact of simulated changes on the predicted death risks can be observed (Figure 7).

Figure 7.

Screenshot of DAPP prediction result

(A) The interface for submitting data to DAPP.

(B) The display of DAPP prediction results.

Discussion

Various classifications of patients based on the transcriptome features have been conducted. In this study, we classified the AML patients according to their death risk curves, which allowed the investigation of the molecular features directly linked to survival. To model the death risk curve for individual AML patients, we built the continuous time series survival model. By utilizing predicted death risk curves for patient classification and biomarkers extracted from the C1/C2 classes, associated with the worst survival outcomes, we demonstrated that these biomarkers exhibit consistent and significant prognostic stratification across multiple validation datasets.

Within the continuous time series survival model, in order to accurately and meticulously depict the risk changes over time, independent machine learning survival prediction models were built monthly. Since survival curves represent a cumulative effect over time, we transformed the predicted survival curves into death risk curves to capture detailed events for patient classification. This model, established across several large cohorts of AML patients, provided predicted death risk distributions for individual AML patients throughout the entire diagnostic period, facilitating the development of precise treatment strategies.

The AML classification based on features related to death risk held significant clinical implications. Unlike traditional RNA expression matrix clustering, this classification avoided interference from prognostically irrelevant genes, demonstrating notable prognostic stratification. We describe the clinical and research implications of this classification at multiple levels. Mutational enrichment analysis revealed that C1/C2 with the worst prognosis exhibited enrichment in TP53 mutations, with C1 additionally enriched in ASXL1 and RUNX1 mutations, while C2 demonstrated an enrichment pattern for deletions on chromosomes 5/7. These mutations provided a genomic explanation for the divergent prognoses observed between C1 and C2. At the single-cell level, we observed that C1/C2 signatures were associated with distinct malignant clusters. This indicated that AML patients in the C1 and C2 classes, characterized by poorer prognoses, might have harbored different types of tumor cells. Furthermore, the C1 class could effectively distinguish a subgroup with an extremely poor prognosis from the Favorable class according to the ELN2017 classification. The identification of this subgroup might alter their unfavorable outcome if recognized early in AML diagnosis and treated with personalized interventions. Therefore, the classification based on death risk features helds greater clinical significance compared to traditional AML classifications.

Limitations of the study

Model training necessitates a substantial amount of samples sourced from diverse clinical centers, resulting in batch effects within the sequencing data of these samples. While this presents a challenge in achieving precise results, it also bestows greater robustness upon the model. To ensure the accuracy of our current findings, we employ multiple validation sets to corroborate the reliability of our conclusions.

In addition, clinical factors such as treatment protocols exert a notable influence on prognosis. However, standardizing these factors across different studies proves challenging, and, sometimes, the information may be incomplete, precluding its inclusion in our predictive model. In this study, we have focused solely on the statistical analysis of clinical features to mitigate inter-dataset disparities. In the future, enhancing our model could involve integrating treatment strategies and comprehensive clinical information, provided that datasets with more standardized treatment protocols and thorough information recording become accessible.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Microarray data	Verhaak et al.24	GEO: GSE6891
Microarray data	Li et al.25	GEO: GSE37642
RNA-seq data	Erickson et al.26	GEO: GSE106291
BEAT AML 2018	Tyner et al.10	https://www.cbioportal.org/study/summary?id=aml_ohsu_2018
BEAT AML 2022	Bottomly et al.27	https://www.cbioportal.org/study/summary?id=aml_ohsu_2022
TCGA AML	Cancer Genome Atlas Research et al.3	https://www.cbioportal.org/study/summary?id=laml_tcga_pub
Microarray data	Tomasson et al.28	GEO: GSE10358
Microarray data	Chuang et al.29	GEO: GSE71014
scRNA-seq data	van et al.23 Han et al.30	GEO: GSE116256
Software and algorithms
Scikit-learn 1.2.2	Pedregosa et al.31	https://scikit-learn.org/
AutoGluon 0.7.0	Erickson et al.32	https://auto.gluon.ai/
Python 3.9.16	Open Source	https://www.python.org/
R 4.0.3	Open Source	https://cran.r-project.org/
Glmnet 4.1–3	Simon et al.13	https://cran.r-project.org/web/packages/glmnet
Mgcv 1.8–33	CRAN	https://cran.r-project.org/web/packages/mgcv
ConsensusClusterPlus	Wilkerson et al.33	https://bioconductor.org/packages/release/bioc/html/ConsensusClusterPlus.html
Survival 3.2–13	CRAN	https://cran.r-project.org/web/packages/survival/
Survminer 0.4.9	CRAN	https://cran.r-project.org/web/packages/survminer/
Limma 3.46.0	Ritchie et al.34	https://bioconductor.org/packages/release/bioc/html/limma.html
GSVA 1.38.2	Hanzelmann et al.35	https://bioconductor.org/packages/release/bioc/html/GSVA.html
Metascape	Zhou et al.36	https://metascape.org/
DAPP	This Study	https://xgaoo.github.io/DAPP

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Xin Gao(gaoxin1@ihcams.ac.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •This paper analyzes existing, publicly available data. These accession numbers for the datasets are listed in the key resources table.
• •The web server DAPP is publicly available. DOIs are listed in the key resources table.
• •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Experimental model and study participant details

The data used in this study is downloaded from public available resources listed in key resources table. Study participant details are listed in Table S1. The patients comprised 532 men and 524 women, with ages from 18 to 88 years old.

Method details

Data preparation

Mutation and karyotype data

Mutation and karyotype data were obtained from three sources. AMLSG data was downloaded from the AML-multistage GitHub repository (http://www.github.com/gerstung-lab/aml-multistage).³⁷ Mutation data for the BEAT AML 2018 and TCGA AML datasets were downloaded from cbioportal (http://www.cbioportal.org/).^3,10 Only de novo AML patients with overall survival information were retained, yielding sample sizes of n = 1234 for AMLSG, n = 217 for BEAT AML 2018, and n = 196 for TCGA AML. The AMLSG dataset in this article is only used to build the mutation prediction model of the web server.

Gene expression data

For microarray data, the series matrix files for GSE37642²⁵ from two platforms (GPL96/GPL570) were downloaded from the NCBI Gene Expression Omnibus (GEO). 9 samples with missing survival data were removed, leaving 553 cases. Raw CELL files for GSE6891²⁴ were downloaded from GEO. Overall survival data was obtained from the cytopred GitHub repository (https://github.com/narenschandran/cytopred).⁹ Samples missing age or survival data were excluded, leaving 505 cases. For RNA-seq data, raw read count files for GSE106291²⁶ were downloaded from GEO. RPKM values for all genes in all samples were calculated using the R/Bioconductor edgeR package (version 3.24.3). mRNA expression Z-Scores for the BEAT AML 2018 dataset and mRNA expression data for the TCGA AML dataset were obtained from cbioportal (http://www.cbioportal.org/). Only de novo AML cases with survival data were retained, yielding n = 250 for GSE106291, n = 187 for BEAT AML 2018, and n = 173 for TCGA AML. All gene expression data was normalized to z-scores across all genes for each sample.

Validation datasets

BEAT_AML_2022²⁷ dataset was downloaded from cbioportal, including only de novo AML cases while excluding duplicate patients from BEAT AML 2018. GSE10358²⁸ and GSE71014²⁹ were downloaded from GEO. Age information is not available for GSE71014, and therefore, it was not utilized in the model AUC assessment for this study. Data processing procedures were consistent with those applied to the preceding datasets.

Machine learning model building

We evaluated a total of 8 machine learning models using two Python modules: scikit-learn (version 1.2.2) for RandomForest, KNeighbors, LinearSVM, DecisionTree, AdaBoost, NaiveBayes and MLPClassifier; and AutoGluon (version 0.7.0) for the AutoGluon model.^31,32 Python version 3.9.16 was used. Scikit-learn is built on top of SciPy, while AutoGluon is an open-source autoML toolbox from Amazon.

LASSO feature selection

Feature selection was implemented using the "cv.glmnet" function in the glmnet package, with the parameter family set to cox/bin and the parameter type of measure set to deviance. The lambda value corresponding to "lambda.min" from "cv.glmnet" was extracted and used to obtain the selected gene set. For feature enrichment, we performed LASSO with binomial family 10 times (excluding 10% of samples each time) using survival status at each monthly time point from 2 to 60 months as the response variable. We divided the time points into 6-month bins. For LASSO selections in one 6-month bin, a total of n features (without removing duplicates) was selected, with gene i being selected m times. Across all the other bins, a total of N features (without removing duplicates) were selected, with gene i being selected M times. We constructed a 2x2 contingency table using m, n-m, M, and N-M and performed a Fisher’s exact test to obtain an enrichment p-value for gene i in the bin.

Cox proportional hazards for individual genes

Univariate Cox proportional hazards analyses were performed separately on the 6 datasets. Gene expression was dichotomized at the median value, and the median hazard ratio (HR) value across datasets was calculated for each gene.

Monotonic spline fitting of survival probabilities at multiple time points

Predicted survival probabilities for an individual sample at different time points were smoothed using a monotonically decreasing spline fit implemented using the mgcv package (version 1.8–33) in R. The "gam" function in the mgcv package was run with argument bs set to "cr", argument k set to 10, and monotonicity constraint enforced by function "mono.con" in the mgcv package. This yielded daily survival predictions from days 1–1800.

Clustering death risk curves

Survival probability predictions from days 1–1800 were binned into 10-day intervals. The average survival probability per 10-day bin was calculated, and the decrement between adjacent bin averages was computed to obtain the death risk value for each 10-day bin.

Death risk curves from day 60 onward were extracted, and consensus clustering was performed using the ConsensusClusterPlus R package.³³ The distance parameter was set to Euclidean and the clustering method to pam. Maximum number of clusters was 20 with 100 repetitions, each sampling 80% of the data.

The median survival for each cluster was estimated using the surv_median function in the "survminer" package and used to rank clusters from C1 to C7 based on decreasing median survival.

To assess the predictive significance of death risk clusters C1-C7, we performed a multivariable Cox regression analysis for overall survival. This analysis incorporated death risk clusters alongside diverse clinical characteristics (including age and gender) and other prognostic markers (such as LSC17 and ELN2017). The analysis was conducted utilizing the R package "survival".

Prediction of the sample classifications in the BEAT_AML_2022 dataset

To predict the cluster affiliation of the samples in the BEAT_AML_2022 dataset, death risk values were calculated for each 10-day bin for each BEAT_AML_2022 sample. Each sample was then classified based on its proximity to the nearest centroid of the clusters, which were previously defined using the training samples.

Mutation enrichment analysis of death risk clusters

For the 335 patients with both mutation and expression data, we analyzed the enrichment of gene g mutations in cluster c as follows. Let m be the number of mutations in gene g within cluster c, n be the number of wildtype samples in cluster c, M be the number of mutations in gene g outside cluster c, and N be the number of wildtype samples outside cluster c. We constructed a 2x2 contingency table using m, n, M, and N and performed Fisher’s exact test to obtain an enrichment p-value. Mutations significantly enriched in cluster c were defined as having a p-value <0.05.

C1-C7 signature score

The highly expressed genes in cluster C_n (n = 1:7) were identified through a differential gene expression analysis, comparing C_n to other clusters within each training dataset using the R limma package.³⁴ Genes exhibiting significantly higher expression in C_n (p-value <0.05) and having the highest expression in C_n across all clusters were chosen. Subsequently, genes were ranked in descending order based on the number of datasets showing significantly high expression and the average logFC. The top 50 genes were selected as marker genes for the respective cluster. The C_n signature score was then computed using ssGSEA based on the identified C_n marker genes.³⁵ The functional enrichment analysis for marker genes was conducted utilizing Metascape with the GO Biological Process ontology.³⁶ The significance threshold was established at a p-value <0.01.

Marker gene for cell types

The AML single-cell dataset GSE116256 was obtained from the Tisch2 database.^23,30 Cell type marker gene analysis was performed using the Seurat package’s FindMarkers function (p-value <0.05, fold change >2), and cell type labels were sourced from Tisch2.

Quantification and statistical analysis

Enrichment analysis in Figure 2A was performed using Fisher’s exact test with the alternative hypothesis set to "greater". P-values for the Kaplan-Meier plots were computed using the log rank test. Enrichment analysis in Figure 5A was conducted using Fisher’s exact test with the alternative hypothesis set to "greater". Significance levels are indicated as follows: ∗p-value <0.05, ∗∗p-value <0.01, ∗∗∗p-value <0.001. Enrichment p-values in Figure 5E were calculated based on the cumulative hypergeometric distribution.

Published: July 5, 2024