AI Finds New Antibiotic

AI Finds New Antibiotic

The last decade has seen a significant rise in antibiotic-resistant bacteria that has become a serious medical problem. Some bacterial strains are now resistant to multiple classes of antibiotics leaving little left to treat them with, and we need entirely new types of antibiotics to combat the surge in resistant bacteria. Unfortunately, the discovery of new antibiotics is decreasing as the common classes have already been found and widely used in medicine. Identification of new antibiotics typically requires screening large libraries of chemical compounds in hopes of finding some with antibacterial capability. Even with miniaturized, high-throughput technologies, brute force screening hundreds of thousands of compounds is an expensive and time-consuming process that all too often fails. To streamline the screening approach, researchers at Harvard-MIT used artificial intelligence (AI) to interrogate the compound libraries and pick out the most likely compounds for physical testing (A Deep Learning Approach to Antibiotic Discovery – published in the journal “Cell”).

Artificial intelligence (AI) has become a common feature in our lives, from chatbots that direct our telephone inquiries to Siri and Alexa that respond to our every request. A key feature of most AI is machine learning, the ability of the software to learn from experience and become more proficient at its tasks. This requires that the software be “trained” with a data set. For example, to produce an AI that can recognize dogs, the program would be trained with a data set consisting of pictures of dogs and non-dogs (other animals). Given sufficient training, when given a picture that it has never seen before the program should be able to classify the image as dog or non-dog. The Harvard-MIT researchers used this same approach to train their software to identify chemical compounds that could potentially stop bacterial growth. They used a training set of 2335 compounds of which 120 inhibited E. coli growth (antibiotic) and the remainder had no growth effect (non-antibiotic). After training the software, a library of 6,111 investigational human drugs was examined with 99 compounds identified as potential antibiotics. Testing of these 99 drugs found 1 that had potent antibacterial action against several human pathogens, including M. tuberculosis and C. difficile. This new antibiotic, named Halicin (after HAL 9000, the sentient computer in 2001 A Space Odyssey), has a novel mechanism of action, has low toxicity in animal studies, and appears very difficult for bacteria to develop resistance against. If additional testing and human trials are successful then this could represent the first generation of an entirely new class of antibiotics that are effective against some of our most challenging bacterial diseases.

To further test their AI, the Harvard-MIT researchers next had it examine a database of over 100,000,000 molecules, a quantity of compounds that would be nearly impossible to physically produce and evaluate. Ultimately the software found around 7000 compounds with potential anti-bacterial action. Using additional criteria, the researcher reduced this to 23 top candidates that were physically tested for antibiotic activity. Eight of the 23 compounds had significant antibiotic activity, with 2 compounds showing broad-spectrum activity and novel structural features only distantly related to known antibiotics. These exciting results demonstrate the amazing power of AI to identify novel antibacterial compounds out of massive compound databases, an advance that should greatly facilitate future screening efforts. Hopefully, this combination of AI prediction and physical testing will provide the next generation of antibiotics to protect us all from the expanding realm of drug- resistant bacteria.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: