Novel Biomarker Prediction for Lung Cancer Using Random Forest Classifiers.

Area of Science:

• Genomics and Bioinformatics
• Cancer Research
• Computational Biology

Background:

• Lung cancer is a leading cause of cancer-related mortality, with non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC) as the primary types.
• Functional genomics and RNA-Seq have advanced the study of gene expression in tumors, but biomarker discovery remains challenging.
• Machine learning offers a promising approach to classify biomarkers based on gene expression patterns in different lung cancers.

Purpose of the Study:

• To compute transcript statistics and identify differential gene expression between lung cancer samples and the reference genome.
• To develop and evaluate machine learning models for classifying genes associated with NSCLC, SCLC, or both.
• To identify potential gene expression biomarkers for NSCLC and SCLC using transcriptome analysis.

Main Methods:

• Transcriptome data analysis, including normalized fold change and gene expression levels.
• Application of exploratory data analysis to understand feature distributions.
• Implementation and comparison of supervised (Logistic Regression, KNN, SVM, Random Forest) and ensemble (XGboost, AdaBoost) machine learning algorithms.
• Utilized Near Miss under-sampling to address dataset imbalance.

Main Results:

• The Random Forest classifier achieved 87% accuracy, identifying it as the best-performing model.
• BRAF, KRAS, NRAS, and EGFR were predicted as potential biomarkers for NSCLC.
• ATF6, ATF3, PGDFA, PGDFD, PGDFC, and PIP5K1C were predicted as potential biomarkers for SCLC.
• Common biomarkers for both NSCLC and SCLC included CDK4, CDK6, BAK1, CDKN1A, and DDB2.
• The fine-tuned model achieved 91.3% precision and 91% recall.

Conclusions:

• Machine learning, particularly the Random Forest classifier, is effective for identifying lung cancer biomarkers from gene expression data.
• Specific gene sets were identified as potential biomarkers for NSCLC and SCLC.
• Dataset limitations, including feature scarcity and imbalance, restrict further accuracy improvements but highlight the potential of this approach.