Fine-tuning a sentence transformer for DNA.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Sentence-transformers library enables text embedding for various applications.
• Vector space embeddings allow for similarity-based text analysis.
• Existing DNA transformers like DNABERT and Nucleotide transformer are domain-specific.

Purpose of the Study:

• Fine-tune a natural language sentence transformer for DNA sequences.
• Evaluate its performance on eight benchmark tasks.
• Compare its efficacy against DNABERT and the Nucleotide transformer.

Main Methods:

• Fine-tuning a sentence transformer model on DNA text.
• Evaluating the model across eight benchmark tasks.
• Comparative analysis with DNABERT and Nucleotide transformer.

Main Results:

• The proposed model's DNA embeddings surpassed DNABERT in several tasks.
• The Nucleotide transformer showed higher classification accuracy but with significant computational costs.
• The proposed model demonstrated competitive performance and efficiency, especially in retrieval tasks.

Conclusions:

• The fine-tuned sentence transformer provides a viable option for DNA sequence analysis.
• It balances performance and accuracy, offering a practical alternative for resource-constrained settings.
• The proposed model outperforms DNABERT and presents a more computationally feasible option than the Nucleotide transformer.