Workflow for Fine-Tuning and Evaluating DNA Language Models for Specific Genomics Issues.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• DNA language models (e.g., DNABERT-2) show promise for predicting functional genomic elements.
• Practical protocols for preparing data, fine-tuning, and evaluating these models are lacking.

Purpose of the Study:

• To present a step-by-step computational protocol for preparing training data, fine-tuning DNABERT-2, and evaluating binary sequence classifiers.
• To enable researchers to adapt DNA foundation models for new genomic applications.

Main Methods:

• Developed a command-line protocol for creating DNABERT-2 compatible datasets from positive and negative sequence sets.
• Fine-tuned the official DNABERT-2 model using prepared datasets.
• Demonstrated protocol application using RNA off-target sites (PiCTURE pipeline) and core promoter prediction (EPDnew database).

Main Results:

• Successfully prepared labeled DNA sequence datasets for DNABERT-2 fine-tuning.
• Showcased reproducibility of DNABERT-2 implementation for binary classification.
• Applied the protocol to RNA off-target prediction and core promoter identification.

Conclusions:

• The protocol facilitates the adaptation of DNA foundation models for diverse genomic tasks.
• Enables safety assessment of genome editing tools and functional annotation of regulatory sequences.
• Provides a reproducible framework for applying DNABERT-2 to new biological problems.