How AI is decoding chemistry to revolutionize drug discovery

The search for new drugs begins in an ocean of uncertainty. Estimates suggest between 10²⁰ and 10⁶⁰ chemical molecules could hold potential as small-molecule therapies. Chemically evaluating even a fraction of these compounds would overwhelm traditional laboratory methods. Artificial intelligence is changing that equation by enabling researchers to rapidly sift through molecular possibilities, predict reaction pathways, and even design novel compounds.

At the forefront of this intersection is Connor Coley, an MIT associate professor who splits his time between chemical engineering and electrical engineering and computer science. His work merges computational modeling with real-world chemistry, aiming to automate and accelerate drug discovery. "Our approach isn’t limited to drug development," Coley explains. "It’s a framework that applies to any challenge involving organic molecules, though drug discovery remains our primary focus."

The making of a computational chemist

Coley’s passion for science runs in the family. His father, a radiologist, his mother—a molecular biophysicist turned business leader—and his grandmother, a math professor, collectively shaped his analytical mindset. By age 16, he had already graduated high school and competed in Science Olympiad, then enrolled at Caltech to study chemical engineering.

During his undergraduate years, Coley cultivated a parallel interest in computer science. He worked in a structural biology lab, using Fortran to help decode protein crystal structures. After graduating, he chose MIT for his PhD, drawn by the opportunity to blend chemistry with computation. "Chemical engineering provided the perfect bridge between my interests in science and mathematics," he says.

Under the guidance of professors Klavs Jensen and William Green, Coley focused on optimizing automated chemical reactions. His research combined machine learning and cheminformatics—computational tools for analyzing chemical data—to design reaction pathways for synthesizing new drug molecules. Simultaneously, he explored hardware innovations to automate these reactions, a direction furthered through his involvement in a DARPA initiative called Make-It.

"That program introduced me to the potential of machine learning in chemical synthesis," Coley recalls. "It reshaped my understanding of how models could map chemical feasibility and predict reaction outcomes."

From graduate student to faculty pioneer

While still completing his PhD, Coley applied for faculty positions, eventually accepting an offer from MIT at just 25 years old. The decision sparked debate—some advisors cautioned against staying at the same institution where he trained—but Coley was swayed by MIT’s unique ecosystem. "The cross-departmental collaboration here is unparalleled," he says. "The energy of the students and the strength of interdisciplinary research made the choice obvious."

To deepen his expertise, Coley deferred the faculty role for a year to complete a postdoctoral fellowship at the Broad Institute. There, he focused on identifying bioactive molecules from billions of candidates in DNA-encoded libraries, targeting proteins linked to disease mutations.

Returning to MIT in 2020, he established his lab with a bold mission: to use AI not only to synthesize promising compounds but to design entirely new molecules with tailored properties. Over the past few years, his team has developed computational tools that merge chemical intuition with algorithmic precision.

Teaching AI the language of chemistry

Coley’s lab has pioneered models that capture the subtleties of molecular interactions. One standout, ShEPhERD, evaluates potential drug molecules by analyzing their three-dimensional shapes and predicted interactions with target proteins. Pharmaceutical companies now deploy ShEPhERD to streamline early-stage drug discovery.

"We’re embedding medicinal chemistry principles into these models," Coley explains. "The goal is to give generative AI the same criteria chemists use—like binding affinity or metabolic stability—so it can make smarter design choices."

Another breakthrough came with FlowER, a generative AI system that predicts reaction outcomes from given chemical inputs. Unlike generic models, FlowER incorporates core physical laws, such as mass conservation, and evaluates the feasibility of intermediate reaction steps. "Chemists naturally think in terms of mechanisms and pathways," Coley notes. "We’re teaching models to do the same."

The road ahead: AI as a co-pilot for discovery

The next frontier lies in refining these models to handle even greater complexity. Coley envisions AI systems that don’t just suggest molecules but also propose synthetic routes tailored to available lab equipment or regulatory constraints. "The ultimate goal is to make AI a true collaborator in the lab," he says. "One that brings both speed and scientific rigor to the process."

As the boundaries between computation and chemistry continue to blur, Coley’s work stands as a testament to the power of interdisciplinary innovation. In a field where discovery often hinges on trial and error, AI is emerging as the ultimate shortcut—one that transforms impossibly vast chemical spaces into navigable terrain.