Transfer learning for cross-context prediction of protein expression from 5'UTR sequence

Area of Science:

• Synthetic Biology
• Computational Biology
• Genomics

Background:

• Model-guided DNA sequence design accelerates cellular reprogramming and engineering of complex biological systems.
• Deep learning models offer accurate predictions for gene expression but struggle with generalization across contexts.
• Lack of model generalization limits the application of data-centric approaches in diverse genetic and experimental settings.

Purpose of the Study:

• To address the generalization limitations of deep learning models in DNA sequence design.
• To demonstrate a transfer learning procedure for adapting pre-trained models to new contexts.
• To enable accurate prediction of protein translation rates from 5' untranslated region (5'UTR) sequences in diverse Escherichia coli contexts.

Main Methods:

• Utilized a pre-trained deep learning model for sequence analysis.
• Applied a simple transfer learning procedure to fine-tune the model.
• Employed a small number of new measurements for calibration in diverse E. coli contexts.
• Leveraged massively parallel reporter assays (MPRAs) for feature learning.

Main Results:

• Successfully adapted a deep learning model to predict protein translation rate from 5'UTR sequence with high accuracy in new contexts.
• Demonstrated effective transfer of learned model features to different genetic and experimental settings.
• Showcased the utility of a small number of new measurements for efficient model calibration.

Conclusions:

• Transfer learning provides an effective solution for generalizing deep learning models in DNA sequence design.
• The developed approach facilitates accurate prediction of translation rates, accelerating biological engineering.
• The released model and calibration procedure serve as a foundation for future model-based sequence design efforts.