A Word2Vec-ResNet Transfer Learning model for promoter prediction with dimensionality reduction and cross-domain...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Promoter prediction is essential for understanding gene regulation.
• Traditional methods like one-hot encoding face challenges with high dimensionality and limited knowledge integration.
• Single-domain knowledge constraints hinder the predictive performance of existing models.

Purpose of the Study:

• To propose an advanced promoter prediction method, Word2Vec-ResNet.
• To address the limitations of traditional encoding strategies in promoter prediction.
• To enhance model generalization by integrating natural language processing (NLP) and cross-domain transfer learning.

Main Methods:

• Utilized Word2Vec for pretraining nucleotide sequence embeddings on source domain data.
• Implemented a ResNet architecture for promoter prediction.
• Applied cross-domain transfer learning by transferring pretrained embeddings to the target domain.
• Evaluated the method on promoter datasets from four diverse organisms: Bacillus subtilis, Escherichia coli, Saccharomyces cerevisiae, and Drosophila melanogaster.

Main Results:

• Achieved a 97.6% average reduction in encoding dimension compared to one-hot encoding.
• Demonstrated an average increase of 18.12% in prediction accuracy compared to baseline methods.
• Showcased significant performance improvements across multiple species, highlighting model generalization.

Conclusions:

• Word2Vec-ResNet effectively reduces encoding dimensionality while improving promoter prediction accuracy.
• The integration of NLP and cross-domain transfer learning offers a powerful approach for deciphering transcriptional regulatory mechanisms.
• This method provides a more efficient and accurate tool for genomic analysis and gene regulation studies.