Fine-tuning-based Transfer Learning for Characterization of Adeno-Associated Virus.

Area of Science:

• Biotechnology
• Nanotechnology
• Bioinformatics

Background:

• Accurate identification of adeno-associated virus (AAV) vectors is crucial for dose-dependent gene therapy.
• Solid-state nanopore techniques offer potential for AAV characterization via ionic current, but current analysis methods are insufficient.
• Deep learning, specifically convolutional neural networks (CNNs), shows promise for classifying AAV vectors based on ionic current profiles.

Purpose of the Study:

• To develop an accurate and efficient method for classifying adeno-associated virus (AAV) vectors using ionic current data from solid-state nanopore measurements.
• To address the challenge of limited data for training deep neural networks from scratch by employing transfer learning techniques.

Main Methods:

• Utilized a pre-trained deep CNN (GoogleNet) for feature extraction from ionic current signals.
• Applied fine-tuning-based transfer learning to classify AAV vectors, requiring minimal preprocessing and no handcrafted features.
• Evaluated classification accuracy across different electric fields and time frame segmentations, and tested other pre-trained networks (ResNet50, InceptionV3).

Main Results:

• Achieved high average classification accuracy for AAV vectors, ranging from 90% to 99%, with minimal standard deviation across three independent trials.
• Demonstrated that classification accuracy is influenced by the applied electric field and the duration of data segments.
• Found that fine-tuning a deep network outperformed traditional feature extraction methods for classifying resistive pulse datasets.

Conclusions:

• Fine-tuning pre-trained deep networks is a highly effective and generic approach for classifying adeno-associated virus (AAV) vectors from nanopore ionic current data.
• This transfer learning strategy overcomes the limitations of small datasets, enabling accurate gene therapy vector identification.
• While GoogleNet offered efficient training, ResNet50 and InceptionV3 also proved effective, albeit with longer training times, highlighting the versatility of deep networks.