همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

Encoding the gRNA-DNA Pairs with Deep Transformers to Predict off-target Effects of CRISPR

نویسندگان :

Roghayyeh Alipanahi¹ Leila Safari² Alireza Khanteymoori³

1- Department of Electrical and Computer Engineering University of Zanjan, 2- Department of Electrical and Computer Engineering University of Zanjan, 3- Department of Psychology, University of Freiburg, Freiburg, Germany

کلمات کلیدی :

CRISPR/Cas،gRNA Design،Off-target،Encoding،Deep Transformers

چکیده :

CRISPR is a revolutionary technology for genome editing, with significant potential in medicine. Improving the accuracy and precision of CRISPR is essential to the further development and widespread application of this novel genome editing technique. The accuracy of CRISPR-based genomic edition depends on two issues: the cutting power of Cas enzyme and the performance of the gRNA sequence. To achieve accurate edits, scientists must select the optimal gRNAs contains high on-target activity and low (no) off-target efficiency. To enhance the accuracy and precision of genome editing with CRISPR and make it practical, it is crucial to concentrate on predicting CRISPR off-target effects and reduce them. Although numerous deep learning-based models have been developed to predict off-target sites, current methods suffer from low precision and overfitting caused by insufficient data. Furthermore, the most of these algorithms use only gRNA sequences in on-hot vector form as input. However, recent research illustrated that both gRNA and DNA beside some epigenetic features strongly impact on the prediction precision of off-target sites. To address these challenges, we propose a novel multi-head attention based deep transformer model to encode both the gRNA and DNA sequences, and use them to predict off-target sites. Using multi-head attention–based transformer model, lead to capture any relationship between each nucleotide and k-mer with other nucleotides and K-mers within the gRNA and DNA sequences. This enhancement allows for a more comprehensive analysis of sequence characteristics and significantly improves the model's capacity to predict off-target sites. Furthermore, the utilization of multi-head attention architecture has enabled us to improve the accuracy and generalizability of our model across diverse CRISPR systems and cell lines. Comparison of off-target prediction results using our proposed gRNA-DNA encoding scheme, deployed in multi-head attention architecture with state-of-the-art models highlights the superior performance of our approach over multiple evaluation criteria.