Classification of pediatric brain tumors

Semestral project for Machine Learning in Bioinformatics course at MFF, Charles University.

Authors: Kostejn Vit, Mihal Filip, 2024.

The work is summarized in the slides

Task

Classify pediatric brain based on Gene Expression data.

Interpret the model and identify the most important genes for each class (binary classificator).

Data

We used a dataset from a study that analyzed frozen or formalin-fixed paraffin-embedded (FFPE) tissue from 571 patient samples, including 559 pediatric brain tumors and 12 non-tumor brain control samples. The dataset was downloaded from the Gene Expression Omnibus (GEO) database.

Data preparation & exploration

import GEOparse

# study accession
GeO_key = 'GSE227756'
gse = GEOparse.get_GEO(geo=GeO_key, destdir="./", silent=True)

# get the data
df = gse.pivot_samples('VALUE').T

# get the labels
labels = gse.phenotype_data[['source_name_ch1','characteristics_ch1.0.tumor type']].values

Data: 571 samples

Features: 103 features (gene or protein names) represented by positive integers

Targets: Binary classification of pediatric brain tumors + 5 classes of pediatric brain tumors

Main challenge: Imbalanced classes

Data preprocessing

Normalization: StandardScaler

Dimensionality reduction: PCA

Oversampling: SMOTE

Model selection

In this part we selected multiple classifiers with different hyperparameters, trained them using the train data and evaluated them using F1 score and 5-fold cross-validation.

Models used:

• Logistic regression
• Perceptron
• Decision tree
• Random forest
• Gradient boosting
• SVM
• K neighbours
• MLP
• XGB
• LGBM

Methodology

Is SMOTE oversampling necessary?

Train-test split: 80-20

Grid search with CV: 5 folds

Scoring: F1 macro

Results

The best performing model was MLP with 64 neurons and 0.01 learning rate.

Model: MLP, 64 neurons, 0.01 learning rate

Data: StandardScaler, 44 principal components, SMOTE oversampling

Analysis of the confusion matrix

Our model primarily struggles with differentiating between high-grade glioma (HGG) and low-grade glioma (LGG), which are similar types of tumors, with HGG being more aggressive. Additionally, the model often confuses MMRD-HGG with HGG. While they are the same type of tumor, MMRD-HGG has a slightly higher mutation rate compared to HGG.

Feature importance

Feature interpretability for tumor and non-tumor samples

Approach

1. Take scaled data (StandardScaler) without dimensionality reduction
2. Train Logistic Regression
3. Coefficients represent feature importance
   • Positive: higher values are more likely to be tumor
   • Negative: lower values are more likely to be non-tumor