This project presents an end-to-end analysis of Acute Myeloid Leukemia (AML) using molecular and clinical data from TCGA-LAML.
The work combines exploratory molecular profiling, multi-omics integration, machine-learning model comparison, candidate biomarker discovery, and survival analysis to investigate whether molecular patterns can support clinically relevant patient stratification.
The project was developed across two related analytical phases:
- Exploratory multi-omics analysis of mRNA expression, miRNA expression, and DNA methylation, including dimensionality reduction and multi-omics integration.
- Precision-medicine predictive modeling using matched clinical, mRNA, and miRNA data, followed by model evaluation, candidate biomarker identification, and survival analysis.
Important: MOFA-based multi-omics integration belongs to the exploratory phase of the project. The later machine-learning models were developed separately using matched mRNA and miRNA features and did not use MOFA factors as model inputs.
The project investigates several related questions:
- What major sources of molecular variation are present across AML patients?
- Can multiple molecular layers provide complementary information about AML heterogeneity?
- Can molecular profiles support prediction of patient outcome?
- Which molecular features contribute most strongly to predictive models?
- Do model-derived patient groups show differences in observed survival?
The goal is not to propose a clinically validated diagnostic or prognostic model, but to demonstrate an exploratory precision-medicine analytics workflow connecting molecular data with patient-level outcomes.
TCGA-LAML Data
│
├── Clinical Data
├── mRNA Expression
├── miRNA Expression
└── DNA Methylation
│
▼
Individual Omics Profiling
│
▼
PCA / Molecular Patterns
│
▼
Exploratory Multi-Omics Integration
(MOFA)
│
▼
Molecular Heterogeneity
────────────────────────────────────
Matched Clinical + mRNA + miRNA Data
│
▼
Data Preprocessing
│
▼
Feature Selection
│
▼
Predictive Modeling
RF / SVM / XGBoost
│
▼
Model Evaluation
│
▼
Candidate Biomarker Analysis
│
▼
Kaplan-Meier Survival Analysis
│
▼
Precision-Medicine Interpretation
The analysis uses preprocessed molecular and clinical information from the TCGA-LAML (Acute Myeloid Leukemia) cohort.
The earlier exploratory analysis examined:
- mRNA expression
- miRNA expression
- DNA methylation
- Clinical metadata
Individual molecular layers were investigated before exploratory integration using MOFA (Multi-Omics Factor Analysis).
For the later precision-medicine workflow, patient identifiers were harmonized across clinical, mRNA, and miRNA datasets.
The final matched modeling cohort contained:
- 162 patients
- 500 selected mRNA features
- 100 selected miRNA features
- 600 molecular features in total
Clinical outcome information was incorporated for supervised modeling and subsequent survival analysis.
The first phase focused on understanding molecular variation across AML patients.
Analyses included:
- Molecular feature distributions
- Patient-level molecular patterns
- Principal Component Analysis (PCA)
- Associations between molecular variation and clinical metadata
- Differential molecular feature exploration
- Multi-omics integration
MOFA was used as an exploratory integration method to investigate shared and modality-specific sources of variation across molecular layers.
This phase was designed to characterize AML molecular heterogeneity rather than build the later predictive classifier.
The heatmap illustrates variation across selected molecular features for a subset of AML patients, providing a visual representation of patient-level molecular heterogeneity.
PCA was used to examine dominant patterns of variation across the integrated molecular feature space.
The second analytical phase investigated whether matched molecular profiles could support patient outcome prediction.
Three supervised machine-learning algorithms were compared:
- Random Forest
- Support Vector Machine (SVM)
- XGBoost
The workflow included data preprocessing, feature selection, train/test evaluation, cross-validation, and comparison of predictive performance.
On the initial held-out split, XGBoost achieved a ROC-AUC of 0.726, outperforming the other evaluated models in that split.
The corresponding initial XGBoost performance included approximately:
- Accuracy: 70.7%
- Sensitivity: 65.0%
- Specificity: 76.2%
- ROC-AUC: 0.726
These results represent performance on the initial train/test split and should not be interpreted as externally validated clinical performance.
Cross-validation produced more conservative estimates than the initial held-out evaluation.
This difference highlights an important modeling consideration: performance from a single split can overestimate generalization, particularly when working with relatively small biomedical cohorts and high-dimensional molecular data.
Feature-importance analysis from the predictive workflow was used to identify molecular variables contributing strongly to model predictions.
The analysis generated a ranked set of candidate mRNA and miRNA biomarkers for further investigation.
These features should be interpreted as model-derived candidate biomarkers, not clinically validated biomarkers.
The corresponding results table is available here:
results/top20_candidate_biomarkers.csv
To investigate whether model-derived patient stratification was associated with observed survival, predicted prognosis groups were evaluated using Kaplan-Meier analysis.
The predicted groups showed a statistically significant difference in observed survival distributions, with a reported:
Log-rank p-value = 0.0044
This suggests that the molecular prediction groups captured survival-related structure within the analyzed cohort.
However, this finding remains exploratory and requires validation in an independent patient cohort before clinical interpretation or application.
The project demonstrates several important analytical findings:
- AML patients showed substantial molecular heterogeneity across the investigated molecular profiles.
- Multi-omics exploration provided a framework for examining shared patterns across molecular layers.
- XGBoost achieved a ROC-AUC of 0.726 on the initial held-out split.
- Cross-validation produced more conservative performance estimates, emphasizing the importance of robust model evaluation.
- Feature-importance analysis identified a set of candidate mRNA and miRNA biomarkers.
- Model-derived patient groups demonstrated significantly different observed survival distributions (log-rank p = 0.0044).
- The analysis illustrates how molecular profiling, machine learning, and survival analysis can be connected within a precision-medicine workflow.
aml-multiomics-precision-medicine/
│
├── README.md
│
├── notebooks/
│ └── AML_MultiOmics_Precision_Medicine_Analysis.ipynb
│
├── scripts/
│ ├── mrna_analysis.R
│ ├── mirna_analysis.R
│ ├── methylation_analysis.R
│ └── multiomics_integration.R
│
├── results/
│ ├── figures/
│ │ ├── integrated_multiomics_feature_heatmap.png
│ │ ├── pca_integrated_multiomics.png
│ │ ├── roc_curves_model_comparison.png
│ │ ├── model_evaluation_cross_validation.png
│ │ ├── top20_candidate_biomarkers.png
│ │ └── kaplan_meier_predicted_risk_groups.png
│ │
│ └── top20_candidate_biomarkers.csv
│
└── docs/
├── methodology.md
└── results.md
Programming & Analysis
- Python
- R
- Pandas
- NumPy
- Scikit-learn
- XGBoost
Statistical & Biomedical Analysis
- Principal Component Analysis
- Multi-Omics Factor Analysis (MOFA)
- Feature importance analysis
- Kaplan-Meier survival analysis
- Log-rank testing
Visualization
- Matplotlib
- Seaborn
- R visualization packages
Detailed documentation is available in:
The analysis notebook is available in:
This project should be interpreted as an exploratory biomedical data-analysis study.
Important limitations include:
- Relatively small matched patient cohort
- High dimensionality compared with sample size
- No independent external validation cohort
- Potential instability of feature importance across different training samples
- Simplification of survival information for the classification component
- Candidate biomarkers require independent biological and clinical validation
Accordingly, the predictive results should not be interpreted as a clinically validated AML prognostic model.
This repository consolidates related academic analyses of the TCGA-LAML cohort into a reproducible portfolio project demonstrating skills in:
Biomedical Data Analysis · Multi-Omics Analysis · Machine Learning · Statistical Analysis · Data Visualization · Biomarker Discovery · Survival Analysis · Precision Medicine
Some analyses originated from collaborative academic coursework. This portfolio repository reorganizes and documents the analytical workflow for professional presentation while preserving the distinction between exploratory multi-omics integration and the later predictive modeling phase.
Nada Emad
Data Analyst | Business, Operations & Healthcare Analytics
B.Sc. Information Technology & Biomedical Informatics
Nile University