Predicting the spread of antimalarial drug resistance using deep learning surrogates - ABSTRACT Malaria represents a major global public health challenge, killing 619,000 people each year, predominantly children in Africa. The recent emergence of artemisinin partial resistance (ArtR) in Africa has the potential to lead to a global public health crisis and may result in the failure of test-and-treat strategies that serve as a cornerstone of malaria control. If it follows the historical example of Southeast Asia, where ArtR parasites are now widespread, millions of additional malaria cases could occur each year. A major challenge for disease ecology is understanding and predicting how resistance to interventions emerges and spreads. For malaria, transmission models in popular use today are simulation based, allowing for rich representations of transmission cycles and also capturing the true random nature of infection. This creates significant problems for model-based inference. Traditional inference methods like Bayesian Markov chain Monte Carlo (MCMC) cannot be used in most cases, and more advanced methods are limited in scope or power. Deep learning offers a new and general-purpose solution to this problem. Deep neural networks (DNNs) trained on large amounts of simulation output can learn the complex relationships between parameter inputs and outputs. The trained “surrogate” model can then be used to produce outputs that closely resemble the original simulation model in a fraction of the run-time, or they can be used within traditional Bayesian MCMC to estimate parameters. This approach, referred to as “deep learning surrogates” (DLS), is gaining traction in many areas of research but has not yet penetrated infectious disease modeling. Most applications of DLS have focused on speeding up prediction. While this is an important use-case, in infectious disease modeling we are equally interested in learning from observed data. Here, mathematical models can provide estimates of crucial parameters like the basic reproductive number of a pathogen, or more detailed outputs like the complete surface of disease prevalence over a geographical region. Thus, DLS has enormous potential by linking complex simulation models to traditional parameter estimation methods. The goal of the proposal is to apply DLS methods to the urgent problem of antimalarial resistance in Sub-Saharan Africa. Specifically we will: 1) develop a temporal DLS model to predict the impact of drug policy interventions in Kinshasa Province, DRC and 2) develop spatial DLS models to predict the spread of artemisinin resistance in the Great Lake region of Africa and Ethiopia. In both cases, we will leverage some of the most comprehensive epidemiological and genetic datasets available.
