Optimizer's dilemma: optimization strongly influences model selection in transcriptomic prediction.

Area of Science:

• Computational biology
• Machine learning
• Genomics

Background:

• Machine learning models are widely used in biological research, yet the choice of optimization algorithms is often not reported.
• Understanding the impact of different optimizers on model performance and interpretability is essential for reproducible scientific findings.

Purpose of the Study:

• To compare the performance and model sparsity of two optimization approaches, coordinate descent (liblinear) and stochastic gradient descent (SGD), for LASSO logistic regression.
• To predict mutation status and gene essentiality from gene expression data across pan-cancer driver genes.

Main Methods:

• Applied LASSO logistic regression using Python's scikit-learn package with two optimizers: liblinear (coordinate descent) and SGD.
• Evaluated model performance and sparsity across various regularization strengths for each optimizer.

Main Results:

• Both liblinear and SGD optimizers demonstrated comparable performance.
• Liblinear models required more regularization tuning but excelled at high model sparsity, while SGD models needed learning rate tuning but showed robustness across varying sparsities.
• Optimizer choice impacts model tuning and performance characteristics.

Conclusions:

• The choice of optimization algorithm significantly influences LASSO logistic regression model outcomes in gene expression-based cancer driver gene prediction.
• Clear reporting of optimizers is vital for the scientific community to ensure result transparency and reproducibility.