Adapting Generative Genome Foundation Model Evo for Functional Genomics Prediction via Progressive Fine-Tuning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomic sequence data modeling is vital for understanding gene function, mutation impacts, and advancing precision medicine.
• Large language models (LLMs) offer new approaches for biological sequence analysis.
• The Evo foundation model shows promise for generative tasks but requires adaptation for specific supervised predictions.

Purpose of the Study:

• To propose Evo-TSFT, a progressive two-stage fine-tuning strategy to adapt the Evo model for DNA functional genomics classification.
• To enhance the applicability of foundation models for supervised prediction tasks in genomics.

Main Methods:

• Developed Evo-TSFT, a novel progressive two-stage fine-tuning strategy.
• Integrated LoRA-based fine-tuning and selective layer unfreezing with the pre-trained Evo model.
• Evaluated performance across 7 DNA functional genomics classification tasks and 24 datasets.

Main Results:

• Evo-TSFT achieved strong overall performance across diverse DNA functional genomics classification tasks.
• Demonstrated the effectiveness of the proposed fine-tuning strategy in adapting Evo for downstream prediction.
• Showcased competitive results compared to existing methods.

Conclusions:

• Evo-TSFT is an effective and competitive strategy for adapting foundation models like Evo to specific DNA functional genomics prediction tasks.
• The fine-tuning approach significantly improves the utility of pre-trained genomic models for supervised learning.
• This work facilitates advancements in functional genomics analysis and precision medicine.