Predicting sequence-specific amplification efficiency in multi-template PCR with deep learning

Area of Science:

• Molecular Biology
• Bioinformatics
• Genomics

Background:

• Multi-template polymerase chain reaction (PCR) enables parallel DNA amplification for diverse applications.
• Non-homogeneous amplification leads to skewed data, reducing accuracy and sensitivity in complex amplicon libraries.

Purpose of the Study:

• To develop a method for predicting sequence-specific amplification efficiencies in multi-template PCR.
• To enable the design of homogeneous amplicon libraries for improved data accuracy.

Main Methods:

• Employed one-dimensional convolutional neural networks (1D-CNNs) to predict amplification efficiencies based on DNA sequence information.
• Trained models on synthetic DNA pools and introduced the CluMo deep learning interpretation framework.

Main Results:

• Achieved high predictive performance (AUROC: 0.88, AUPRC: 0.44) in predicting amplification efficiencies.
• Identified specific motifs near adapter priming sites associated with poor amplification.
• Elucidated adapter-mediated self-priming as a major cause of low amplification efficiency.

Conclusions:

• Deep learning models can predict and mitigate non-homogeneous amplification in multi-template PCR.
• The findings challenge existing PCR design assumptions and reveal adapter-mediated self-priming as a key mechanism.
• This approach reduces sequencing depth requirements and enhances DNA amplification efficiency for various biological applications.