Biomarkers of Acute Myeloid Leukemia From RNA-seq Expression and Feature Selection with Machine Learning

Jenny Smith (lead)

Vikas Peddu

David Lee (technical)

Sean Madden (Scribe)

Overview

• Leukemia the most common pediatric cancer

• AML has high molecular heterogeneity

• Recent interest in comparing pediatric and adult cancers to find age-related and –unrelated markers

• Interest in applying machine learning for feature selection, inc. ”old” and newer techniques

• Need for well-packaged data forms for TARGET

​

Variable Selection

1. 60, 488 Genes Initially x 478 Samples
1. Filtered to remove low count genes - at least 1 CPM in 5% of Samples
2. 21,407 Genes with Adequate expression levels x 478 Samples
3. Split into 3 Data sets
• David - merged Clinical Annotations and Normalized counts for each cohort separately
• AML: 187 cases; NBL: 150 cases; WT: 132 cases
• Biological Samples used in RNAseq from various time points (eg diagnosis, remission (?), and relapse)
4. Batch Effect Investigation
• PCA w/ different covariates

​

​

5A. Differential Expression Analysis

• Limma voom for each cohort comparing groups of interest (eg all patients who relapsed, patients with mutation X )

​

5B. No additional filters

​

​

Machine Learning

Data Access, Normalization, and Summaries of TARGET AML Samples

• Download Illumina HiSeq RNA-seq gene counts by samples
• Convert to TMM-normalized gene expression (scaled to library size)
• Samples: N = 119 primary marrow, 36 primary peripheral blood
• Approx. balanced survival time between ‘old’ and ‘young’ age groups (based on median)
• Approx. balanced sex terms between old/young

Feature