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

DNA language models, such as DNABERT-2, have recently enabled the accurate prediction of functional sequence elements across species. However, the practical, protocol-style steps needed to transform these resources into training datasets, fine-tune the official DNABERT-2 model, and evaluate classifier performance have not been explicitly described. Herein, we present a step-by-step computational protocol for preparing training data, fine-tuning DNABERT-2, and evaluating sequence-level binary classifiers using readily available command-line tools. The protocol has been demonstrated using RNA off-target sites induced by cytosine base editors, detected by our PiCTURE pipeline from RNA sequencing (RNA-seq) data, and extended to core promoter prediction using the EPDnew database. We describe how to derive positive and negative sequence sets into DNABERT-2 compatible datasets, and fine-tune the official pretrained model of DNABERT-2 using the datasets. We also demonstrate how to compute the standard performance metrics and compare the model outputs with the baselines. This protocol will help researchers adapt DNA foundation models to new genomic tasks, including the safety assessment of genome editing tools and the functional annotation of regulatory sequences. Key features • A step-by-step pipeline has been provided for preparing labeled DNA sequence datasets for DNABERT-2 fine-tuning. • The reproducibility of the official DNABERT-2 implementation and scripts for binary classification is shown. • Application of PiCTURE-derived C-to-U (T) substitution sites for off-target RNA risk prediction. • An extension to core promoter prediction using promoter datasets from EPDnew is described.