» Go to news main

Fighting against antimicrobial resistance in the artificial intelligence era

Jee In Kim's interdisciplinary PhD faces big questions in AI and machine learning (photo Daniel Abriel)
Jee In Kim's interdisciplinary PhD faces big questions in AI and machine learning (photo Daniel Abriel)

Artificial intelligence (AI) involves the learning and acquisition of intelligence by machines without humans’ continuous and explicit instructions. The discipline is expanding rapidly, and various AI technologies are starting to get incorporated into biological research. For my own research, I employ machine learning to predict a pathogen’s resistance to a specific antibiotic. Machine learning (ML) is a branch of AI, where algorithms mimic how humans learn using a large dataset to accurately predict an outcome. I feed various algorithms with pathogen genomes and the corresponding resistance or susceptibility profile to a specific antibiotic and create a model. The model can then predict the resistance status in a new but similar set of genomes and be further tweaked to improve the prediction performance.

I have described my research process as simply as possible to help readers understand the big picture. Reading the simple description of my research seems straightforward enough, but we need to consider that organisms evolve – and bacteria evolve fast. That is why we are in the antimicrobial resistance (AMR) apocalypse right now. Microorganisms are ubiquitous, densely populated, and capable of exchanging genetic materials with each other to help adapt to their changing surroundings more efficiently and quickly. With such knowledge, you can start to see why a prediction model might not be able to achieve 100% prediction performance. Maybe the pattern that the ML model learned from one dataset cannot perfectly predict the resistance outcome of another group of organisms of the same species because they are from different isolated locations, and the bacteria have adapted to their specific niches. Factors contributing to the emergence of AMR vary from one environment to another (re: my previous blogs), making it difficult to perfectly predict whether bacteria would be resistant to certain drugs with one general model.

My research goal is to maximize the amount of information that one can obtain through the use of ML models, not necessarily to develop a model with 100% prediction accuracy. Understanding which parts of the genomes the models are using to make predictions will help us to better understand the potential mechanisms bacteria use to exert AMR. The information has the potential to also assist in surveillance efforts for public health usage. If we notice an increasing trend in genetic patterns that are known to be involved with a specific resistance, we can better inform antimicrobial usage guides. However, ultimately, ML is a deductive tool used to develop hypotheses that should be validated experimentally. Without the bridge between ML predictions and experimental validation, we will always be stuck at the research phase of the technology. I hope to bridge the two sides of ML prediction and validation with my research and discover exciting results that I could share with you soon! In the era of AI, biologists and computer scientists need to work collaboratively to maximize the research impact that can potentially halt the emergence of AMR.