AI Assists in Discovery of Potent Antibiotic Against Deadly Superbug
Researchers use AI for Antibiotic discovery
Created on May 26|Last edited on May 26
Comment
Researchers from Canada and the United States have leveraged AI to discover a new antibiotic capable of killing a particularly dangerous superbug. In the face of growing antibiotic resistance, finding new antibiotics is a pressing issue. Over a million people worldwide die annually from antibiotic-resistant infections. The problem is further compounded by the dearth of new antibiotic discoveries in recent decades. This work focuses on combatting Acinetobacter baumannii, a bacteria labeled a "critical" threat by the World Health Organization, known for its resistance to a range of antibiotics.
AI Assisted Discovery
The researchers used AI to filter through thousands of potential compounds. The AI was trained using known drugs, manually tested against A. baumannii to gauge their effectivity. This process enabled the AI to identify the chemical features necessary to combat the bacteria effectively. Subsequently, the AI was tasked with scrutinizing 6,680 compounds, the effectivity of which was unknown against A. baumannii [1]. This AI-facilitated process identified nine potential antibiotics, one of which was the highly potent abaucin. Initial laboratory tests on abaucin showed promising results. The antibiotic effectively treated infected wounds in mice and successfully eliminated A. baumannii samples from patients. Despite the initial success, more work is required. The team needs to refine the drug further in the laboratory and conduct clinical trials. They estimate that AI-aided antibiotics could become available for prescription by 2030 [1].
High Specificity
Interestingly, abaucin demonstrated precision in its effectiveness, working solely against A. baumannii. This specificity is unusual as most antibiotics affect a wide range of bacteria [1]. However, the team believes this precision could minimize the emergence of drug resistance and lead to fewer side effects. The AI method has the potential to screen tens of millions of compounds, a task practically impossible manually. AI's ability to enhance the rate and potentially reduce the cost of discovering new antibiotic classes underscores its value. Having previously tested AI principles in antibiotic discovery against E. coli, the researchers intend to apply this knowledge to tackle other problematic pathogens, such as Staphylococcus aureus and Pseudomonas aeruginosa.
Model Architecture
At the heart of the AI model utilized by the researchers is a graph encoder architecture, much like those employed in social networks and chemistry applications [2]. The researchers used a graph convolutional neural network, which learns feature representations directly from the data. Their model improves upon the existing Message Passing Neural Network (MPNN) framework. Instead of updating atom representations, it focuses on updating directed bonds. Additionally, it merges computed molecule-level features with the molecular representation taught by the MPNN. MPNNs work in two stages: a message-passing phase to build a neural representation of the molecule and a readout phase to make property predictions based on the final representation. During training, the network intakes molecular graphs and outputs predictions for each molecule, adjusting its approach based on the accuracy of its predictions.
Illustration of the MPNN from [2]
Overall, this research offers an encouraging step forward in using AI to combat antibiotic-resistant bacteria, demonstrating the technology's immense potential in accelerating the search for novel antibiotics.
Model paper:
Medical Paper:
Add a comment