DNACSE: Enhancing Genomic LLMs with Contrastive Learning for DNA Barcode Identification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• DNA barcoding is crucial for biodiversity assessment.
• Existing DNA language models struggle with DNA barcoding due to data structural differences.
• There's a need for specialized models to improve DNA barcoding tasks.

Purpose of the Study:

• To introduce DNACSE (DNA Contrastive Learning for Sequence Embeddings), an unsupervised framework to fine-tune DNA language models for barcoding.
• To enhance the embedding space distribution for better sequence representation.
• To improve the performance of DNA language models in DNA barcoding and generalized genomic tasks.

Main Methods:

• Developed DNACSE, an unsupervised noise-contrastive learning framework.
• Applied DNACSE to fine-tune a DNA language foundation model.
• Evaluated performance on DNA barcoding tasks including fine-tuning, linear probe, and zero-shot clustering.

Main Results:

• DNACSE achieved 99.17% accuracy in fine-tuning and 98.31% in linear probe tasks, outperforming BarcodeBERT.
• Adjusted Mutual Information (AMI) score improved to 92.25% in zero-shot clustering.
• DNACSE demonstrated enhanced performance in generalized genomic tasks through benchmarking.

Conclusions:

• DNACSE significantly improves DNA barcode species classification by leveraging multispecies and barcode information.
• The framework offers a viable approach for biodiversity exploration and conservation.
• DNACSE enhances the utility of DNA language models for genomic applications.