Breaking: Deep learning multi-dataset training improves CRISPR base-editing predictions

The research team addressed this by generating substantial new experimental data and developing a novel training strategy. Using their previously established SURRO-seq technology (see Figure 1), which creates libraries pairing gRNAs with their target sequences integrated into the genome, they measured base-editing efficiency for approximately 11,500 gRNAs each for ABE7.10 and BE4-Gam base editors in HEK293T cells. After quality filtering, they obtained robust measurements for over 11,000 gRNAs per editor.

Analysis revealed that ABE7.10 exhibited highly specific adenine-to-guanine transitions at 97%, whilst BE4 showed 92% cytosine-to-thymine specificity. Both editors displayed peak activity at positions four through eight in the protospacer sequence. Notably, the team found positive correlations between base-editing efficiency and standard Cas9-induced indel frequency at the same sites, suggesting that Cas9 activity predictions could inform base-editing models.

The key innovation lies in how they integrated their data with existing published datasets. Rather than simply pooling data, they developed an architecture that trains simultaneously on multiple datasets whilst explicitly labelling each data point's origin. For adenine base editors, they incorporated five datasets, including their SURRO-seq data and published data from Song, Arbab, and two Kissling datasets, examining both ABE7.10 and the engineered ABE8e variant. For cytosine base editors, they used three datasets from SURRO-seq, Song, and Arbab studies. The model architecture employs deep convolutional neural networks with multiple filter sizes to process the 30-nucleotide target sequence, alongside molecular features including gRNA-DNA binding energy and predicted Cas9 efficiency.

The researchers acknowledge limitations in their current models. They have incorporated only three base editor variants – ABE7.10, ABE8e, and BE4 – and all training data derive from HEK293T cells, except for datasets that include mouse embryonic stem cells and U2OS cells. Different deaminases exhibit distinct sequence motif preferences, meaning predictions for base editors not represented in the training data remain challenging. The dataset integration strategy could apply to other CRISPR technologies where data heterogeneity poses challenges. The models are available both as a web server and as standalone software under an academic use licence.

»We hope CRISPRon-ABE and CRISPRon-CBE will help other researchers select gRNAs giving the highest editing efficiency and the intended outcome,« says Yonglun Luo.

The study was led by Jan Gorodkin and Yonglun Luo at the University of Copenhagen and Aarhus University, respectively, in Denmark. The study was published today, 7 November 2025, in Nature Communications under the title "Deep learning models simultaneously trained on multiple datasets improve base-editing activity prediction".

The CRISPRon-ABE and CRISPRon-CBE tools can be accessed via https://rth.dk/resources/crispr/. Sequencing data is deposited in the China National GeneBank Database under accession CNP0001031 and NCBI Sequence Read Archive under accession PRJNA1321836.