Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Deep learning foundation models aim to interpret complex biological data, including single-cell transcriptomics.
• Predicting gene expression changes after genetic perturbations is crucial for understanding cellular function.

Purpose of the Study:

• To evaluate the performance of deep learning foundation models in predicting transcriptome alterations.
• To compare these advanced models against simpler baseline approaches for genetic perturbation prediction.

Main Methods:

• Assessed five foundation models and two other deep learning models.
• Benchmarked model performance against simple predictive baselines.
• Focused on predicting transcriptome changes following single and double genetic perturbations.

Main Results:

• No deep learning foundation models outperformed the simple baseline methods.
• The predictive accuracy for transcriptome changes was not improved by complex models.
• This suggests limitations in current deep learning approaches for this specific task.

Conclusions:

• Current deep learning foundation models do not surpass simple baselines for predicting genetic perturbation outcomes.
• Rigorous benchmarking is essential for evaluating and guiding the development of new computational tools in systems biology.
• Future research should focus on improving model interpretability and predictive power for complex biological systems.