Deep Learning AI has Discovered a Promising New Antibiotic

Finding new antibiotics is like a game of hide-and-seek. Even finding penicillin was due to random chance — bacteriologist Alexander Fleming returned from vacation to discover penicillin growing in his bacterial colonies. If Fleming hadn’t gone on vacation for so long, penicillin may not have been discovered for another few decades.

Chemists and medicine developers often find antibiotics that are related to each other, but none are novel. Pathogens keep improving their resistance to antibiotics, and it has become clear that we desperately need to find new antibiotics to counteract the pathogens’ growing drug resistance.

Recently, a team of researchers from Harvard and MIT announced that they have developed a medical application to find new antibiotics.

The Promise of Deep Learning

Antibiotics are expensive and time-consuming to find; they’re not economical for drug companies to invest in, but there are more and more cases of super-resistant bacteria that aren’t affected by one or more antibiotics. Without new antibiotics being discovered, experts forecast that the current rate of deaths due to drug resistance (700,000 per year) could rise to 10 million in 2050.

Enter deep learning and artificial intelligence (AI) applications.

The Harvard-MIT software uses deep learning, a small subset of the field of AI, to create a very niche algorithm that fuels drug discovery. Deep learning works best when it follows the “No Free Lunch” theorem of AI: it states that there is no universally-superior algorithm; if a task needs to be done, then an algorithm built from the ground-up, focusing its training and efforts on that one type of task, would have the best potential for success.

With the Boston-based collaboration at Harvard and MIT, the researchers are using a subset of deep learning called graph neural networks. Previous algorithms for drug discovery used long-winded text descriptions of chemicals, but this new method describes chemicals as a network of atoms, which allows the algorithm to more clearly “understand” the chemical make-up of antibiotics. And with this “understanding” comes a more accurate pattern-finder and drug discovery generator.

AI versus Humans

The team at Harvard and MIT trained the algorithm on data surrounding drugs that are effective and ineffective. They also made sure to include drugs that are known to be safe for human consumption. Lastly, the team included data on molecular make-ups of 2,500 compounds; half were FDA-approved drugs, and 800 are naturally-occurring. After this training period, the researchers felt that the algorithm was ready to start finding safe but potent antibiotics using combinations of millions of chemicals.

As promising as this all sounds, it turns out that deep learning alone isn’t good enough to find new antibiotics; researchers need to have a deep knowledge base of infections. This means AI can’t be left to its own devices. Humans need to stay involved and keep tweaking the algorithm as needed to improve results. But unlike humans, AI doesn’t have any bias unless it’s given biased data by its human developers. That means AI doesn’t have any preconceived idea about what an antibiotic should look like.

With this algorithm, the Harvard-MIT team was able to find many new antibiotics that don’t look anything like existing antibiotics. One, in particular, looks very promising: it’s named Halicin, and it’s being looked at for use in treating diabetic patients. Halicin is very effective against E. coli, in addition to many more deadly pathogens, like those that cause tuberculosis and colitis (inflammation of the colon).

Perhaps most promising is the fact that Halicin works nicely against the drug-resistant pathogen Acinetobacter baumanni. According to the CDC (Centers for Disease Control and Prevention), this bacteria is one of the deadliest pathogens in the world.

The Future of Drug Discovery

One issue that Halicin could cause is the potential to destroy “good” and harmless bacteria in our body. It could have metabolic side effects too, but experts aren’t worried about these issues until we know more from testing and clinical trials.

The Harvard-MIT team tested the efficacy of Halicin on mice that were infected with a strain of bacteria that is resistant to all known antibiotics. Halicin performed extraordinarily well: it cleared up the infections within a day.

With the promising results of Halicin, should we stop looking for new antibiotics? The long and short answers are “no”. We need to continue searching for antibiotics through our old method and any new methods that will spring up from innovative technologies like AI and MedTech. It won’t be long before Halicin is being used in hospitals, and soon enough, we’ll hear about a new strain of Halicin-resistant bacteria infections.

Tomorrow’s Hope

As for the team of researchers at Harvard and MIT, they are already working on expanding their algorithm by feeding it 100 million more chemical molecules. Using this new giant subset of training data, the team has discovered 23 more antibiotics that don’t resemble our current group of antibiotics. Two of these seem to be as promising as Halicin, so they’ll require more testing and trials.

It turns out that using AI to fuel drug discovery could very well give humanity the final leg-up in the ongoing fight against antibiotic-resistant pathogens.