Tuberculosis (TB) is one of the world’s deadliest diseases, according to the U.S. Centers for Disease Control and Prevention. Close to 10 million people were infected with it in 2017, a year during which there were 1.3 million reported TB-related deaths. Making matters worse, the bacterium that causes TB — mycobacterium tuberculosis — is difficult to target, due to its ability to develop a resistance to certain drugs.

Fortunately, researchers at Harvard Medical School’s Blavatnik Institute have devised a computational approach capable of detecting resistance to commonly used TB drugs pre-treatment. In experiments, they managed to accurately anticipate a TB strain’s resistance to 10 first- and second-line drugs in a tenth of a second and with greater precision than similar models.

The method is described in the journal EBioMedicine and will be added to Harvard Medical School’s genTB tool, which analyzes TB data and predicts TB drug resistance.

“Drug-resistant forms of TB are hard to detect, hard to treat and portend poor outcomes for patients,” said Maha Farhat, senior study author and assistant professor of biomedical informatics at Harvard Medical School, in a statement. “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.”

As Farhat and colleagues explain, of the millions of new cases of TB diagnosed each year, roughly 4% are resistant to at least two drugs and 1 in 10 shows resistance to multiple medications. Drug-sensitivity testing equipment can be difficult to procure in the developing world, and even in well-equipped laboratories, it takes weeks for the results to be validated. Newer exams that scan sample DNA for resistance genes have their limitations, too, chiefly an inability to spot resistance for more than a handful of drugs or to detect the presence of rare resistance-promoting genetic variants. As for whole-genome sequencing tests, they frequently underperform on detection of resistance to second-line drugs.

The researchers’ method, by contrast, leverages machine learning algorithms to capture the effects of multiple mutations. It incorporates two models: a statistical model and a “wide-and-deep” system that codes each mutation as a variable that either confers resistance or doesn’t.

“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,” said study first author Michael Chen, who began developing the models as a research as a freshman at Harvard. “The … neural network interlaces two forms of machine learning to identify the combined effects of genetic variants on antibiotic resistance.”

The two AI systems were trained on 3,601 TB strains resistant to first- and second-line drugs, including 1,228 multidrug-resistant strains with results from drug-sensitivity testing. To test their performance, the paper’s coauthors fed them samples from a test corpus of 792 fully sequenced TB genomes.

The wide-and-deep AI system predicted resistance to first-line and second-line drugs with 94% accuracy and 90% accuracy, respectively, while the statistical model predicted resistance to first-line drugs with 94% accuracy and to second-line therapy with 88% accuracy. Both models were capable of predicting resistance in a tenth of a second to first- and second-line therapies, while the wide-and-deep model showed an aptitude for predicting the effects of extremely rare genetic mutations.

The researchers claim that if incorporated in clinical tests, the models could make drug resistance detection both faster and more accurate.

“The wide-and-deep model is a decision-making tool that combines all of the data with prior biological knowledge that resistance is caused both by large individual mutations and the interactions between many different mutations,” Andy Beam, study coauthor and faculty member in biomedical informatics at Harvard Medical School, said. “Our model highlights the role of artificial intelligence in the case of TB, but its importance goes well beyond TB.”