10.05.2024 Views

YSM Issue 97.1

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

FEATURE<br />

Computational Biology<br />

AI VS. SUPERBUGS<br />

CAN DEEP LEARNING BEAT<br />

ANTIBIOTIC RESISTANCE?<br />

BY MADELEINE POPOFSKY | ART BY MADELEINE POPOFSKY<br />

The great irony of recent advances<br />

in artificial intelligence (AI)<br />

is that even as programs have<br />

gained capabilities that ten years<br />

ago we could only dream of, most of<br />

the time, we have no idea how they<br />

do it. This lack of explainability has<br />

particularly frustrating consequences<br />

for the use of AI in drug discovery.<br />

Lacking the understanding of how AI<br />

generates predictions of successful drug<br />

candidates, scientists struggle to design<br />

drugs similar to those predicted. Felix<br />

Wong, a postdoctoral fellow at MIT,<br />

worked in collaboration with other<br />

researchers to solve this explainability<br />

problem in a recent Nature paper. The<br />

team applied newly explainable AI<br />

results to one of the most<br />

pressing medical crises<br />

of our age: antibiotic<br />

resistance. Currently,<br />

bacteria are<br />

developing resistance to antibiotics at<br />

a rate that outpaces researchers’ ability<br />

to design new ones. “Antimicrobial<br />

resistance is a public health crisis that<br />

is projected to kill ten million people<br />

worldwide per year by 2050,” Wong<br />

said. An especially deadly enigma is<br />

methicillin-resistant Staphylococcus<br />

aureus (MRSA), which already kills<br />

over ten thousand per year in the<br />

United States. For researchers, MRSA<br />

has proved to be an elusive target. As<br />

a Gram-positive bacterium, S. aureus<br />

lacks an outer membrane that helps<br />

many bacteria defend themselves from<br />

antibiotics, but it has nevertheless<br />

independently evolved resistance to<br />

many antibiotics. Thus, to find an<br />

antibiotic for a bacterium as unique<br />

as MRSA, researchers would need to<br />

identify a whole new structural class<br />

of drugs—a tall order, given that gaps<br />

between such kinds of discoveries have<br />

previously exceeded thirty-five years.<br />

This is where AI steps in. Over the<br />

years, researchers have trained deep<br />

learning models, a type of AI, to<br />

identify certain desirable molecules<br />

that are linked to antimicrobial activity.<br />

Such models have limited success when<br />

researchers cannot identify the models’<br />

reasoning—without it, researchers can’t<br />

point to specific chemical features that<br />

give the molecules their desired effects.<br />

These models are typically referred to<br />

as “black box” models.<br />

Equipped with these standard “black<br />

box” models and determined to shed<br />

light on their inner workings, Wong and<br />

his colleagues took on the challenge of<br />

creating an antibiotic for MRSA. Their<br />

crucial insight was to focus on specific<br />

chemical substructures as units for the<br />

deep learning models. If a model focuses<br />

on specific substructures, Wong and his<br />

colleagues thought, then maybe they<br />

could get the model to explain which<br />

substructures in its chosen molecules<br />

account for its antimicrobial activity.<br />

During preliminary testing, this idea<br />

was confirmed. “We noticed that<br />

compounds with similar structures have<br />

consistently similar model prediction<br />

28 Yale Scientific Magazine March 2024 www.yalescientific.org

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!