Project Lead: Coleman Martin, UW Chemical Engineering
Data Science Lead: Noah Benson
In 2014, the United Nations set a goal to end the HIV/AIDs endemic by 2030, aiming for 90% of people living with HIV know their HIV status, be on retroviral treatment, and be virally suppressed. As of 2019, the most recent data, the United Nations program on HIV/AIDs reports that only 59% of people living with HIV are virally suppressed. HIV viral load testing is the key diagnostic in this plan and, to achieve these goals, 30 million diagnostic tests will be needed each year. A rapid, quantitative, and accessible diagnostic HIV test is not currently available.
The Posner Research Group has developed a novel rapid quantitative isothermal nucleic acid test that relies on discrete puncta counting in fluorescence microscopy images at a single time point. This test quantifies nucleic acids at low viral loads; however, this method loses accuracy at high viral loads within the clinically relevant range because the spacing between puncta decreases and they merge. This results in undercounting and poor accuracy and ultimately hinders the clinical utility of the diagnostic.
In this work, we addressed this limitation by applying a ResNet convolutional neural network to both spatial and temporal image data from time-resolved fluorescence microscopy. Here we use seven time points that are significant over the course of the reaction as temporal data in the model input channels. We then used OPTUNA (https://optuna.org/) to optimize model architecture and hyperparameters, including the learning rate and ResNet complexity.
Our final model achieves 95% prediction accuracy across four clinically relevant HIV viral load classifications and can be extended to 91% accuracy across five orders of magnitude of input DNA concentration. This approach significantly extends the dynamic range of the original system and represents a promising step toward accessible and scalable HIV diagnostics.
Image Caption: Figure 1 (A) Time-resolved fluorescence microscopy images were acquired every 5 seconds over a 15-minute period (B) Images were preprocessed with background subtraction and resized to 500x500px. Seven key time points were selected and fed into a ResNet model. The image shown corresponds to 100,000 DNA copies. (C) The ResNet processes image sequences through layered convolutions, extracting features and classifying HIV viral loads. (D) The model was evaluated for accuracy on four clinical viral load categories and across a 5-log DNA concentration range.