Artificial intelligence has discovered a surprising new antibiotic

The latest revolution in artificial intelligence (AI) offers new hope. In a study published February 20 in the journal Cell, scientists from MIT and Harvard used a type of artificial intelligence called deep learning to discover new antibiotics

By Sriram Chandraskaran, Assistant Professor of Biomedical Engineering, University of Michigan

Imagine you are a fossil hunter. You spend months in the heat of Arizona digging up bones only to discover that what you have discovered is the bones of a previously discovered dinosaur. This is how the search for antibiotics has been going until recently. Very few new antibiotics are discovered, and in all searches the same types of antibiotics come up again and again.

With the rapid increase in drug resistance among many bacteria, new antibiotics are desperately needed. It may only be a matter of time before a wound or scratch becomes life-threatening. However, few new antibiotics have entered the market, and even these are only mild versions of old antibiotics.

While the prospects look bleak, the latest revolution in artificial intelligence (AI) offers new hope. In a study published on February 20 in the journal Cell, scientists from MIT and Harvard used a type of artificial intelligence called deep learning to discover new antibiotics.

The traditional way of discovering antibiotics - from soil or plant extracts - has not revealed new candidates, and there are many social and economic obstacles to solving this problem as well. Some scientists have recently tried to tackle this by searching the DNA of bacteria for new genes that produce antibiotics. Others look for antibiotics in exotic places like our noses.

Medicines found through such unorthodox methods face a long and difficult road to market. The drugs that are effective in petri dishes may not work well in the body. They may not be absorbed well or they may cause side effects. Producing these drugs in large quantities is also a significant challenge.

Deep learning

These deep learning algorithms power many of today's facial recognition systems including autonomous cars. They mimic the way neurons in our brain work by learning patterns from data. An individual artificial neuron may detect simple patterns such as lines or circles. By using these thousands of artificial neurons, deep learning can perform particularly complex tasks such as identifying cats in videos or locating tumors in biopsy images.
Given the power and success of the method, it may not be surprising to learn that researchers looking for new drugs are embracing deep learning. However, building an artificial intelligence method to discover new drugs is no small task. This is mostly because there is no such thing as a free lunch in the AI ​​field. This means that if an algorithm performs spectacularly at one task, say facial recognition, then it will fail spectacularly at another task, such as drug discovery.
The Harvard-MIT team used a new type of deep learning called "Graphic Neural Networks" for drug discovery. Back in the "Stone Age" of AI in 2010, AI models were built for drug discovery using text descriptions of chemicals. It's like describing a person's face using words like "dark eyes" and "long nose". These text descriptions are helpful but obviously don't paint the whole picture. The AI ​​method used by the Harvard-MIT team describes chemicals as a network of atoms, allowing the algorithm a more complete picture of the chemicals than text descriptions can provide.

Human knowledge and AI layers

However, deep learning alone is not enough to discover new antibiotics. This should be coupled with a thorough biological knowledge of infections. The Harvard-MIT team rigorously trained the AI ​​algorithm using examples of effective and ineffective drugs. In addition, they used drugs known to be safe to identify potentially safe but potent antibiotics among millions of chemicals.
Unlike people, AI has no preconceived notions, especially about what antibiotics should look like. In an old-school AI lab, my lab recently discovered some surprising candidates for TB treatment, including an antipsychotic drug. In a study conducted by the Harvard-MIT team, they found a gold mine of new candidate antibiotics. These drug candidates do not look like existing antibiotics. One promising candidate is lysine, a drug being tested for the treatment of diabetes.
Surprisingly, lysine was potent not only against E. coli, the bacteria on which the AI ​​algorithm was trained, but also against more deadly pathogens, including those that cause tuberculosis and colitis. It is worth noting that lysine was potent against drug-resistant Acinetobacter baumanni. This bacterium is at the top of the list of deadliest pathogens compiled by the Centers for Disease Control and Prevention.
Unfortunately, the strong power of lysine causes the killing of harmless bacteria in our bodies. There may also be metabolic side effects, as it was originally designed as an antidiabetic drug. Given the dire need for new antibiotics, these may be small sacrifices to pay to save lives.

Continuing the evolution

Given the promise of lysine, should we stop looking for new antibiotics?

Lycin may overcome all obstacles and eventually reach the market. But he still has to overcome a constant enemy that is the main cause of the drug resistance crisis: evolution. Humans have overmedicated pathogens over the past century. However, the pathogens have always developed resistance. So it probably won't be long before we encounter a lysine-resistant infection. However, with the power of deep learning, we may now respond quickly with new antibiotics.
Potential AI-discovered antibiotics face many challenges before they reach the clinic. The conditions under which these drugs are tested are different from those inside the human body. My lab and others are building new AI tools to simulate the body's internal environment to assess the strength of the antibiotic's effect. AI models can now also predict drug toxicity and side effects. Together, these AI technologies may soon give us an edge in the never-ending battle against drug resistance.

For an article in The Conversation