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

Promoter prediction is critical for deciphering transcriptional regulatory mechanisms. However, traditional one-hot encoding strategies suffer from dimensionality explosion with vocabulary expansion, while single-domain knowledge constraints limit predictive performance. Therefore, we propose a promoter prediction method (Word2Vec-ResNet) that innovatively integrates natural language processing (NLP) techniques with cross-domain transfer learning. By pretraining word embeddings on source domain data and transferring the pretrained embedding table to the target domain, the method effectively reduces the dimensionality of nucleotide sequence encoding while leveraging inter-domain knowledge to enhance model generalization. Comprehensive experiments on promoter datasets of four representative organisms (Bacillus subtilis, Escherichia coli, Saccharomyces cerevisiae, and Drosophila melanogaster) demonstrate that the proposed method achieves significant performance improvements: compared with one-hot encoding, its average encoding dimension is reduced by 97.6%; compared with baseline methods, the prediction accuracy is increased by an average of 18.12% (with the ratio of the training set to the test set being 8:2).