Harvard undergraduate creates a computer program which detects when one of the world’s top infectious killers becomes resistant to common drugs.
Tuberculosis (TB) remains one of the world’s deadliest diseases. In 2017, close to 10 million people were diagnosed, with up to 1.3 million TB-related deaths according to the World Health Organization [1]. One of the greatest challenges in treating TB is the bacterium’s ability to adapt and become drug resistant. Of the 10 million new detections detected, 4% were resistant to at least two drugs, while 1 in 10 of those exhibited extensive drug resistance to multiple medications.
Identifying resistant strains quickly remains challenging. Current drug resistance testing methods are laborious, due to the bacterium’s propensity to grow slowly in the lab, delaying drug-sensitivity test results by as much as six weeks after patients are diagnosed. “The ability to rapidly detect the full profile of resistance upon diagnosis is critical both to improving individual patient outcomes and in reducing the spread of the infection to others”, argues Maha Farhat, an assistant professor of biomedical informatics at Harvard Medical School [2].
Longevity.Technology: TB is a killer – we improve our medicine and it improves its resistance. This diagnostic tool could turn the tide as well as having implications for treating other diseases in the future.
However, Michael Chen, a Harvard undergraduate applied math student working with Harvard Medical School’s Blavatnik Institute, has created a machine learning model which predicts TB strain’s resistance to first-line drugs with 94% accuracy and with 90% accuracy for second-line drugs, on average, in a tenth of a second.
Chen’s innovative new model outperforms existing approaches for detecting drug resistance, which either take too long or are not as precise. New molecular models that scan the DNA of TB strains are being increasingly used in clinical settings but only detect resistance for up to four drugs. The most comprehensive method for detecting all resistance-conferring mutations is genome sequencing. However, while they can detect resistance to first-line drugs, they do not adequately detect resistance to second-line drugs.
“Our goal was to develop a neural network model, which is a type of machine learning that loosely resembles how connections between neurons are formed in the brain”, reports Chen in EBioMedicine [3]. The neural network approach combines two modes of analysis—wide learning and deep learning. The wide-learning function resembles a statistical model, where each mutation is either coded as a variable that confers resistance or does not. Deep-learning incorporates hidden layers which assess how multiple genes and mutations interact with each other. According to the researchers, the resulting model behaves just as a diagnostic tool that can evaluate all available information against prior knowledge, to determine a strain’s resistance.
Chen’s model will be available online soon through Harvard Medical School’s genTB tool. Nevertheless, in the developing world were TB instances are highest, modern diagnostic capabilities cannot support AI drug-sensitivity testing. In which case, a recent study published in The Journal of Infectious Diseases which argues widely available and inexpensive anti-platelet drugs, such as aspirin, could be used to treat patients with severe TB holds more promise [4]. However, as co-author in the Harvard study, Andy Beam notes, their model’s “importance goes well beyond TB” and indicates “AI will help guide clinical decision-making by rapidly synthesizing large amounts of data to help clinicians make the most informed decision in many scenarios and for many other diseases” [3].
[1] https://www.who.int/news-room/fact-sheets/detail/tuberculosis
[2] https://hms.harvard.edu/news/got-resistance
[3] https://www.ebiomedicine.com/article/S2352-3964(19)30250-6/fulltext
[4] https://www.sciencedaily.com/2019/03/190325101349.htm
