Characterizing DNA Origami Nanostructures in TEM Images Using Convolutional Neural Networks.

Area of Science:

• Nanotechnology
• Biomedical Engineering
• Computational Science

Background:

• Artificial intelligence (AI) models accelerate materials design.
• DNA origami nanostructures are crucial for programmable self-assembly in biomedicine.
• Characterizing nanostructures aids in understanding their function and application.

Purpose of the Study:

• To benchmark the performance of nine Convolutional Neural Network (CNN) models for characterizing DNA origami nanostructures.
• To evaluate CNN models' ability to determine the ligation number in transmission electron microscopy (TEM) images.
• To compare the accuracy of pretrained versus fine-tuned CNN models using simulation and experimental data.

Main Methods:

• Pretraining nine CNN models (AlexNet, GoogLeNet, VGG16, VGG19, ResNet18, ResNet34, ResNet50, ResNet101, ResNet152) on 720 coarse-grained molecular dynamics (MD) simulation images.
• Fine-tuning the pretrained CNN models using a dataset of 146 experimental TEM images.
• Benchmarking model performance based on accuracy, computational time, and model size.

Main Results:

• All CNN models exhibited similar computational time requirements.
• Among pretrained models, ResNet50 and VGG16 achieved the highest accuracies on test MD images.
• Fine-tuned VGG16 demonstrated the highest agreement with experimental TEM images, indicating superior characterization capability.

Conclusions:

• Fine-tuned VGG16 CNN models can efficiently and accurately characterize the number of ligation sites in DNA origami nanostructures from TEM images.
• This AI-driven approach offers a rapid method for analyzing nanostructures in large image datasets.
• The study highlights the potential of AI in advancing materials design and nanomedicine applications.