Strategies for effectively modelling promoter-driven gene expression using transfer learning.

Area of Science:

• Molecular Biology
• Bioinformatics
• Genomics

Background:

• Designing cell-type-specific synthetic promoters is crucial for molecular medicine and gene therapy but is challenging due to limited data for model training.
• Existing large datasets offer potential for transfer learning, but effective strategies for promoter-driven expression remain underexplored.

Approach:

• Investigated various pretraining tasks, transfer strategies, and model architectures for modeling promoter-driven expression.
• Developed two benchmarks to evaluate methods in data-constrained and large dataset scenarios.
• Assessed the effectiveness of transfer learning in improving model performance for predicting gene expression.

Key Points:

• In data-constrained settings, pretraining and transfer learning boosted performance by 24-27%.
• In large dataset settings, transfer learning provided more modest gains of up to 2%.
• Identified optimal model architectures for both training from scratch and transfer learning scenarios.

Conclusions:

• The proposed methods are broadly applicable for modeling gene expression in understudied cell types.
• Findings guide the selection of models for optimizing synthetic promoter design in gene delivery.
• This work facilitates model-based optimization for precise gene delivery applications.