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

Modeling genomic sequence data is crucial for identifying functional genomic elements, predicting mutation effects, and driving advancements in precision medicine and bioengineering. However, the inherent complexity of genomic information poses significant challenges for functional analysis and structural prediction of genomic sequences. Recent developments in large language models (LLMs) have introduced powerful new paradigms for modeling biological sequences. The genomic foundation model Evo, trained on vast multi-species DNA sequence data, has demonstrated remarkable capabilities in generative tasks across molecular to genomic scales. However, Evo cannot be directly applied to specific supervised functional genomics prediction tasks, such as core promoter detection. To address this limitation, we propose Evo-TSFT, a novel progressive two-stage fine-tuning strategy that adapts Evo for DNA functional genomics classification tasks. Evo-TSFT integrates LoRA-based fine-tuning and selective layer unfreezing with the pre-trained Evo model, achieving strong overall performance across 7 DNA functional genomics classification tasks spanning 24 datasets. Experimental results show that Evo-TSFT is an effective and competitive strategy for adapting Evo to downstream DNA functional genomics prediction tasks.