Background

DNA language models trained on large-scale genomic corpora have emerged as powerful tools for the prediction of functional sequence elements [1–6]. DNABERT-2 is a foundation model pretrained on data comprising multi-species genomes using byte-pair encoding tokenization and an efficient transformer architecture [6]. It has demonstrated state-of-the-art performance in diverse genome-understanding tasks.

Cytosine base editors (CBEs) introduce C-to-U (T) substitutions without generating double-strand breaks [7–9]; however, some base editors can also cause unintended C-to-U (T) substitutions on RNA, called RNA off-target [10–12]. RNA off-target sequences are often enriched in canonical ACW motifs (W = A or T/U) [10]; however, non-canonical events are also observed and are more difficult to characterize [13,14]. To systematically profile such off-targets, we previously developed PiCTURE (Pipeline for CRISPR-induced Transcriptome-wide Unintended RNA Editing) (https://github.com/KazukiNakamae/PiCTURE) [15], which processes RNA-seq data from CBE-transfected samples and outputs high-confidence C-to-U (T) substitution sites and their local sequence contexts [13].

In our recent study, PiCTURE outputs were used to fine-tune DNABERT-2 into a binary classifier, the SNL (RNAoffScan v2) model (https://github.com/KazukiNakamae/RNAOffScan), which distinguishes CBE-induced RNA substitutions from background substitution sites and outperforms motif-only baselines in terms of accuracy, precision, recall, and F1 score [13]. Hence, DNA foundation models can capture both canonical and noncanonical sequence signals associated with off-target editing.

In addition to genome editing safety, DNABERT-2 can be adapted to a broad range of regulatory genomics tasks [6]. In this protocol, eukaryotic promoter databases, such as EPDnew [16], provide experimentally validated transcription start sites (TSSs) and flanking promoter sequences for multiple species. These resources enable the systematic construction of a positive dataset for promoter prediction and allow the evaluation of model performance [17,18].

However, the practical steps required to construct training datasets from PiCTURE outputs or the EPDnew database, fine-tune the official DNABERT-2 implementation, and evaluate the classifier performance have not been described in a protocol format. The following protocol provides a reproducible, end-to-end workflow for (i) generating labeled DNA sequence datasets from PiCTURE outputs and the EPDnew database, (ii) formatting them for DNABERT-2, (iii) fine-tuning DNABERT-2 using the official pretrained model, and (iv) evaluating the performance relative to motif-based or other machine learning-based baselines. By following this protocol, researchers can adapt DNABERT-2 to new sequence-level classification tasks, quantify model performance, and leverage DNA language models for genome-editing risk assessment and eukaryotic regulatory genomics.