BERT-siRNA: siRNA target prediction based on BERT pre-trained interpretable model.

Area of Science:

• Bioinformatics
• Molecular Biology
• Computational Biology

Background:

• RNA interference (RNAi) relies on small interfering RNA (siRNA) for mRNA silencing.
• Accurate siRNA selection is crucial but challenging due to limitations in current machine learning approaches, including large data needs and complex preprocessing.
• Existing methods often struggle with accuracy, hindering efficient RNAi applications.

Purpose of the Study:

• To develop an accurate and explainable computational method for predicting siRNA target gene knockdown efficiency.
• To overcome the limitations of small sample sizes and extensive data preprocessing in current siRNA selection models.
• To enhance the reliability and stability of siRNA prediction for biological researchers.

Main Methods:

• Proposed BERT-siRNA, a novel method utilizing a pre-trained DNA-BERT module and a Multilayer Perceptron (MLP) module.
• Employed transfer learning by pretraining DNA-BERT on extensive genomic data and fine-tuning on specific siRNA datasets.
• Integrated attention score analysis and hidden layer visualization for model explainability.

Main Results:

• BERT-siRNA demonstrated superior performance compared to all existing siRNA prediction models on an independent public dataset.
• The model accurately predicted high-efficiency siRNA knockdown for SARS-CoV-2 and aligned with experimental results for PDCD1, CD38, and IL6.
• Attention analysis revealed a predominant focus on the 5' end of the siRNA, and MLP layer visualizations clarified effective feature extraction.

Conclusions:

• BERT-siRNA offers a reliable, stable, and highly accurate solution for siRNA target gene knockdown efficiency prediction.
• The model's explainability, through attention scores and hidden layer analysis, enhances its utility for biological research.
• This advancement facilitates more effective and precise applications of RNA interference technology.