Predicting Gene Essentiality from Gene Expression
Comparing interpretable regression strategies for high-dimensional genomic prediction.
2019 · University of Cambridge, Wellcome Sanger Institute
A comparative machine-learning study investigating whether gene-expression profiles can predict gene essentiality across cancer cell lines.
The Biological Question
This project was completed during a computational-analysis placement at the Wellcome Sanger Institute. The research question: can the expression level of one or multiple genes predict the essentiality of a target gene across cell lines?
Dataset Construction
Gene-expression data (CCLE) and gene-essentiality data (Achilles gene-effect scores) were aligned by gene and cell line, then filtered against reference sets of common-essential and non-essential genes to produce a consistent modeling set.
- Gene expression — genes
- 19,144
- Gene expression — cell lines
- 1,210
- Gene essentiality — genes
- 18,333
- Gene essentiality — cell lines
- 625
- Common-essential reference genes
- 1,247
- Non-essential reference genes
- 710
- After alignment & filtering — genes
- 896
- After alignment & filtering — cell lines
- 622
High-Dimensional Data Challenges
With thousands of genes and hundreds of cell lines, the feature space vastly exceeds sample size for any multi-gene model. This motivated regularized approaches — Lasso in particular — over unconstrained multi-feature regression.
Model Strategies
Three strategies were compared: linear regression as a same-gene baseline, spline regression to capture non-linear same-gene relationships, and Lasso regression to use multiple genes' expression with L1-regularized feature selection.
Linear regression
A simple baseline using the expression of the same gene to predict its essentiality.
Spline regression
A non-linear model using piecewise polynomial basis functions; the project compared four, five, and six knots.
Lasso regression
A high-dimensional model using the expression of multiple genes, with L1 regularization for coefficient shrinkage and feature selection.
low alpha → many active coefficientshigh alpha → sparse coefficients
Validation Protocol
Data was split 80/20 into train and test sets, with cross-validation used for model selection. All three model families were evaluated under the same protocol for a fair comparison.
- MSE
- R²
- Adjusted R²
- 5-fold Cross-Validation
Results
- Lasso produced the strongest overall performance across the evaluated MSE, R², and adjusted R² metrics.
- Spline regression performed slightly better than standard linear regression.
- Changing the number of spline knots between four, five, and six had limited impact.
- Best overall
- Lasso
- Best non-linear baseline
- Spline
- Core benefit
- Prediction plus feature selection
Feature Selection
Lasso's L1 penalty shrinks irrelevant coefficients toward zero, which made it possible to read off which genes' expression contributed to predicting a target gene's essentiality — prediction and feature selection from the same model.
Limitations and Future Work
- Reported gene relationships are predictive associations, not claims of biological causation.
- Individual high-R² gene models are not representative of overall dataset performance and are not presented as such.
- No average performance values are stated here beyond the qualitative ranking reported by the original study (Lasso > Spline > Linear).
- Any matrix, chart, or shrinkage visual on this page is an illustrative reconstruction of the method, not a rendering of the original results.
Technologies
- Python
- scikit-learn
- Linear Regression
- Spline Regression
- Lasso Regression
- Cross-Validation
- Feature Selection
- Genomic Data Processing
- MSE
- R²
- Adjusted R²