Predicting DNA toehold-mediated strand displacement rate constants using a DNA-BERT transformer deep learning model.

Area of Science:

• Biotechnology and Nanotechnology
• Computational Biology
• Machine Learning

Background:

• Dynamic DNA nanotechnology enables applications in molecular computing, cargo delivery, and sensing.
• Machine learning integration can enhance the design of complex DNA machinery.
• Toehold-mediated strand displacement is a fundamental process in DNA nanotechnology.

Purpose of the Study:

• To develop a novel deep learning framework for predicting DNA strand displacement rate constants.
• To leverage transformer architecture and DNA-BERT for enhanced prediction accuracy.
• To analyze the impact of local features and sequence representation on model performance.

Main Methods:

• Generated an in-silico dataset of 4450 DNA sequences and rate constants using KinDA.
• Trained a 1D convolutional neural network utilizing local features and DNA-BERT sequence embeddings.
• Compared DNA-BERT with the One-hot encoder for sequence representation.

Main Results:

• The deep learning model achieved a root mean square error of 0.76 in predicting rate constants.
• DNA-BERT demonstrated improved prediction accuracy compared to traditional methods.
• Analysis provided insights into the influence of local features on model training.

Conclusions:

• The developed deep learning framework effectively predicts DNA strand displacement kinetics.
• DNA-BERT offers a powerful tool for accelerating the design and optimization of DNA nanomachines.
• This approach minimizes reliance on computationally intensive simulations and experimental validation.