Machine Learning Modeling for Gene Expression: From Transcriptome Data to Biological Interpretation

Core Project Information

• Original Author/Maintainer: sidhikh0409-wq
• Source Platform: GitHub
• Release Date: June 2, 2026
• Original Link: https://github.com/sidhikh0409-wq/Gene-expression-modelling-with-ML-

Core Insights

This project innovatively converts differential gene expression analysis from traditional binary classification to a regression problem. It uses machine learning models to quantitatively predict gene expression changes and significance scores, compares multiple algorithms, and supports biological interpretation. By retaining more quantitative information, it provides finer biological insights.

Traditional differential gene expression analysis often uses binary classification (significant/non-significant), which loses quantitative information about expression changes. The core objectives of this project are:

• Treat differential expression analysis as a regression problem to quantitatively predict gene expression changes (logFC) or significance scores (B statistic)

Advantages of this method:

1. Retain continuous numerical information of expression changes
2. Predict the probability of gene significance (B statistic)
3. Enhance biological interpretation ability

Dataset Features

The project uses differential gene expression statistics from transcriptome analysis. Key columns include:

• logFC: log fold change of gene expression
• AveExpr: average expression level
• t: adjusted t-statistic
• P.Value: raw p-value
• B: log odds of differential expression The target variable is logFC or B statistic (continuous) instead of binary labels.

Preprocessing Steps

1. Remove non-numeric identifier columns (e.g., gene names)
2. Select the target variable
3. Check for missing values and outliers
4. Perform feature standardization

Train-Test Split

Adopt an 80% training set (for training, tuning, cross-validation) + 20% test set (for independent evaluation of generalization ability).

Evaluated Regression Models

Compare 6 algorithms:

1. Linear Regression (baseline, high interpretability)
2. Random Forest Regression (ensemble, handles non-linear relationships)
3. Decision Tree Regression (single tree, easy to visualize)
4. Gradient Boosting Regression (sequential ensemble, excellent performance)
5. Support Vector Regression (kernel method, suitable for high-dimensional data)
6. K-Nearest Neighbors Regression (instance-based, non-parametric)

Evaluation Metrics

• R² Score: proportion of explained variance
• MSE: Mean Squared Error
• RMSE: Root Mean Squared Error

Analysis Workflow

Import preprocessed data → select target variable → split dataset → train multiple models → evaluate performance → compare results → select best algorithm → interpret feature contributions

Bioinformatics Value

• Predictive modeling: not only identify significant genes but also predict new gene expression changes
• Feature importance: identify key influencing features via models like Random Forest
• Model comparison: systematically evaluate the performance of different algorithms on biological data

Methodological Insights

1. Problem reframing: converting classification to regression retains more information
2. Multi-model comparison: avoid dependency on a single algorithm
3. Interpretability: provide biological insights through feature analysis

Extended Applications

• Drug response prediction: predict drug sensitivity based on gene expression
• Disease progression modeling: predict molecular changes in disease stages
• Biomarker discovery: identify expression patterns related to clinical outcomes
• Personalized medicine: predict treatment response based on individual transcriptomes

Current Limitations

• Small dataset size (typical transcriptome data)
• Lack of advanced feature selection methods (e.g., LASSO, Elastic Net)
• No comparison with deep learning methods

Improvement Directions

• Integrate multi-omics data (genomics, proteomics, etc.)
• Apply regularization methods for feature selection
• Explore the application of neural networks in gene expression prediction
• Add time-series modeling to capture dynamic expression changes

This project demonstrates the innovative application of machine learning in bioinformatics. By replacing traditional classification with regression modeling, it extracts richer biological information from transcriptome data. It embodies interdisciplinary thinking—combining machine learning methodologies with bioinformatics problems, providing a reference framework for researchers in bioinformatics, computational biology, and machine learning applications in life sciences.