Predicting gene expression levels from DNA sequences and post-transcriptional information with transformers

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Predicting gene expression levels is vital for clinical applications.
• Existing methods like Xpresso use Convolutional Neural Networks and Transformers but suffer from sparse matrices due to one-hot encoding.
• These methods often neglect crucial post-transcriptional regulation processes.

Purpose of the Study:

• To introduce Transformer DeepLncLoc, a novel method for predicting mRNA abundance (gene expression levels).
• To address the limitations of sparse matrices and the exclusion of post-transcriptional regulation in current models.

Main Methods:

• Utilizes a transformer-based architecture for gene expression prediction.
• Employs the DeepLncLoc embedding method, based on the word2vec algorithm, to avoid sparse matrices.
• Integrates post-transcriptional information, including mRNA stability and transcription factors.

Main Results:

• Transformer DeepLncLoc achieved a R² of 0.76, surpassing the state-of-the-art method Xpresso (R² of 0.74).
• The inclusion of transcription factor data significantly enhanced predictive performance.
• Transformer architecture effectively models interactions between DNA locations, outperforming recurrent models.

Conclusions:

• Transformer DeepLncLoc offers a more effective approach to gene expression level prediction.
• The method's ability to handle sparse data and integrate regulatory information represents a significant advancement.
• The findings highlight the suitability of transformer models and the importance of incorporating transcription factor data for improved predictive power.