Evaluating the Transferability of a CNN-Based HIV Drug Resistance Prediction Model Across Subtypes and Continents
Estefanos Kebebew, Takkudzwa Chirenje, Tarnampreet Kaur, Kaylee Chow
Department of Biology
Faculty Supervisor: Pleuni Pennings
Human immunodeficiency virus drug resistance (HIVDR) poses a significant challenge to the efficacy of antiretroviral therapy (ART), particularly in low- and middle-income countries (LMICs). HIVDR testing is pertinent in determining which medications are needed for the infected individual. However, it is very costly. Sub-Saharan Africa (SSA) is the HIV epidemic epicenter, where HIV-1 subtype C is most prevalent. Available machine learning prediction models have been developed using subtype B sequences from high-income countries, and there are concerns about their transferability to other subtypes. This study uses a deep learning model called Convolutional Neural Network (CNN) to determine if a model trained on HIV-1 sequences from Kaiser Permanente Northern California (KPNC) (n = 3399) could accurately predict M184V and K103N drug resistance in an SSA dataset (n = 4366) obtained from the Stanford HIV Drug Resistance Database. Accuracy, confusion matrix, precision, and F1-score will be used to assess model performance. We will also plot the learning curve to determine if the model is generalizing effectively. Next, we will use the saved trained model, initially trained on subtype B, to test its adaptability on HIV subtype C using transfer learning. This model will demonstrate the effectiveness of the multi-continental approach.