Abstract by Brian Brown
Fighting Germs with Machine Learning
Antibiotics have saved untold millions of lives since their emergence in the early 20th century. However, microbes can develop resistance to antibiotics, rendering antibiotics less effective. This potentially serious problem has been remediated by the consistent development of new classes of antibiotics over time. Unfortunately, the last few decades have seen a decline in new antibiotic discovery, and antibiotic resistance looms as a greater and greater threat. We present a new method for antibiotic discovery that leverages machine learning and high-throughput laboratory techniques to identify common themes in antibiotic substances. By training an artificial neural network to recognize toxic and non-toxic substances, our method was able to classify substances as antibiotic, neutral, or probiotic with an accuracy of 72%, or approximately two and one-fifth times better than random guessing. Our success in predicting antibiotic properties indicates common features among antibiotic substances that can be identified and exploited for future antibiotic development.