Semi-quantitative Classification of HIV-1 Nucleic Acids Using ResNet Image Analysis of Discretized Isothermal Amplification Reactions in a Microfluidic Chip

Technical Summary: Semi-quantitative Classification of HIV-1 Nucleic Acids via ResNet and Microfluidic Isothermal Amplification

Problem Statement
While isothermal nucleic acid amplification tests offer significant advantages for rapid and decentralized molecular diagnostics, they frequently suffer from a lack of robust quantitative readouts when compared to quantitative PCR (qPCR). Existing quantitative isothermal molecular tests often struggle to maintain accuracy across both high and low nucleic acid concentration regimes, limiting their utility in settings requiring precise viral load estimation or semi-quantitative assessment outside centralized laboratories.

Methodology
The authors propose a semi-quantitative measurement approach that leverages machine learning to analyze spatiotemporal features derived from real-time fluorescence imaging. The experimental workflow involves:

• Platform: A microfluidic chip facilitating rapid recombinase polymerase amplification (RPA) reactions.
• Data Acquisition: Multiple images are sampled throughout the amplification process to capture the development of fluorescence over time.
• Model Architecture: A Convolutional Neural Network (CNN), specifically utilizing a ResNet architecture, is trained to process these image sequences.
• Training Strategy: Instead of relying on explicit feature extraction, the model learns spatiotemporal patterns of fluorescence development directly from the raw image data.
• Classification Targets: The model is trained to classify samples into two distinct schemes:
  1. Clinically relevant categories.
  2. Logarithmically spaced concentration ranges spanning five orders of magnitude.

Key Contributions

• Integration of Deep Learning with Isothermal Amplification: The study demonstrates the feasibility of using deep learning to interpret complex spatiotemporal fluorescence data from isothermal reactions, bypassing the need for manual feature engineering.
• Extended Dynamic Range: The approach successfully addresses the limitations of current isothermal tests by maintaining accuracy across a broad concentration spectrum (five orders of magnitude), specifically where other methods typically fail at extreme high or low concentrations.
• Automated Interpretation: The system provides a framework for the automated interpretation of amplification reactions, reducing the reliance on specialized laboratory expertise for data analysis.

Results
The study reports high performance metrics for the trained models:

• Clinical Classification: The model achieved 94.6% accuracy in categorizing samples into clinically relevant groups.
• Logarithmic Classification: The model achieved 92.7% accuracy in classifying samples into logarithmically spaced concentration bins.
• Error Analysis: The majority of classification errors occurred between adjacent concentration categories, indicating that the model effectively distinguishes between significantly different viral loads even if fine-grained differentiation between similar concentrations remains challenging.

Significance
The paper posits that integrating machine learning with image-based amplification methods can support rapid semi-quantitative molecular testing. By extending the usable dynamic range of the assay and enabling automated interpretation, this approach may facilitate the broader deployment of nucleic acid diagnostics in decentralized settings, moving beyond the constraints of centralized laboratory environments. The work suggests that machine learning can effectively bridge the gap between the speed of isothermal methods and the quantitative rigor traditionally associated with PCR.