Using transfer learning models for DNA sequence similarity via fCGR method

Similarity analysis of DNA sequences is a critical issue for understanding evolutionary relationships and identifying genetic mutations. Since traditional alignment-based methods have high computational costs, this study investigated the applicability of transfer learning models for alignment-independent DNA similarity analysis. DNA sequences were visualized with the Frequency Chaos Game Representation (fCGR) method and feature extraction was performed with ResNet50, EfficientNetB0, and MobileNet models. Three similarity metrics such as cosine similarity, Euclidean distance, and correlation and four different hierarchical clustering methods were compared. The results show that cosine similarity metric reflects genetic similarities better. MobileNet provided the highest accuracy rate with its lightweight structure and efficient feature extraction. Feature vectors visualized with PCA exhibited strong clustering tendencies and were in agreement with reference phylogenetic trees. The study demonstrates the applicability of transfer learning in genetic analyses and shows that scalable and biologically meaningful analyses can be performed.