The chemical universe is staggering in its scale. Estimates suggest that between 1020 and 1060 compounds could theoretically serve as small-molecule drugs, a number so vast that it dwarfs the grains of sand on every beach on Earth. For chemists, evaluating even a fraction of these candidates through traditional laboratory experiments would take millennia. This is the precise bottleneck that a new generation of artificial intelligence models, developed at MIT, is designed to break. By embedding fundamental chemical principles directly into machine learning architectures, researchers are not just accelerating drug discoveryaaa—they are building models that genuinely understand the rules of chemistry.
At the center of this effort is MIT Associate Professor Connor Coley PhD ’19, the Class of 1957 Career Development Associate Professor with joint appointments in the departments of Chemical Engineering and Electrical Engineering and Computer Science, and the MIT Schwarzman College of Computing. His work sits at the intersection of chemical engineering and computer science, developing computational models that can analyze vast libraries of potential compounds, design entirely new molecules, and predict the reaction pathways needed to synthesize them. While the approach is general enough to apply to any organic molecule, the primary target is small-molecule drug discovery.
The Scale of the Chemical Search Space
The numbers involved in chemical space are almost incomprehensible. The range of 1020</ to 1060 potential small-molecule drugs represents every possible combination of atoms and bonds that could theoretically exist. This is not a problem that can be solved by brute force experimentation. Traditional high-throughput screening, which tests thousands of compounds against a biological target, only scratches the surface. The challenge is further compounded by the fact that a promising candidate must not only bind to a target protein but also be synthesizable, stable, non-toxic, and able to reach its intended site of action in the body.
This is where AI models offer a transformative advantage. Instead of synthesizing and testing every candidate, machine learning models can learn the underlying patterns of chemical reactivity, binding affinity, and molecular properties. They can then screen billions of virtual compounds in silico, ranking them by their likelihood of success. The key question, however, has always been whether these models are truly learning chemistry or simply memorizing statistical correlations. Coley’s research suggests that the answer is increasingly the former.
From Chemistry Olympiad to MIT Faculty
Coley’s path to becoming a leader in AI-driven chemistry was shaped by a family steeped in science. His father is a radiologist, his mother earned a degree in molecular biophysics and biochemistry before attending the MIT Sloan School of Management, and his grandmother was a math professor. Growing up in Dublin, Ohio, he participated in Science Olympiad competitions and graduated from high school at the age of 16.
He headed to the California Institute of Technology, where he chose chemical engineering as a major because it offered a way to combine his interests in science and mathematics. During his undergraduate years, he also pursued computer science, working in a structural biology lab where he used the Fortran programming language to help solve the crystal structure of proteins. After graduating from Caltech, he came to MIT in 2014 to start a PhD under the advisement of professors Klavs Jensen and William Green.
His doctoral work focused on optimizing automated chemical reactions, combining machine learning with cheminformatics—the application of computational methods to analyze chemical data—to plan reaction pathways for new drug molecules. He also worked on designing the hardware needed to perform those reactions automatically. This was partly conducted through a DARPA-funded program called Make-It, which aimed to use machine learning and data science to improve the synthesis of medicines and other useful compounds from simple building blocks.
That program was the real entry point for Coley into thinking about how models could understand the making of different chemicals and the possibilities of reactions. He began applying for faculty positions while still a graduate student and accepted an offer from MIT at age 25. Despite the conventional wisdom against staying at the same institution for graduate school and faculty appointments, Coley decided that the resources, interdisciplinary fluidity, and collaborative ecosystem at MIT were too compelling to pass up.
He deferred the faculty position for one year to complete a postdoc at the Broad Institute, seeking more experience in chemical biology and drug discovery. There, he worked on identifying small molecules from billions of candidates in DNA-encoded libraries that might have binding interactions with mutated proteins associated with disease.
Building Models with Chemistry Intuition
Since returning to MIT in 2020, Coley has built a lab group with a clear mission: deploy AI not only to synthesize existing compounds with therapeutic potential but also to design new molecules with desirable properties and invent new ways to make them. The lab has developed a range of computational approaches to tackle these goals, each grounded in a philosophy of pairing a specific chemical challenge with a potential computational solution.
ShEPhERD: Evaluating Drugs Through 3D Shape
One of the most notable models to emerge from Coley’s lab is ShEPhERD. This model was trained to evaluate potential new drug molecules based on how they will interact with target proteins, using the three-dimensional shapes of the drug molecules as the primary input. The model is designed to give generative AI a medicinal chemistry intuition, making it aware of the right criteria and considerations that a human chemist would apply.
Pharmaceutical companies are already using ShEPhERD to help them discover new drugs. The model’s ability to assess molecular interactions in three dimensions represents a significant advance over earlier approaches that relied on simpler, two-dimensional representations of molecules. By incorporating spatial awareness, ShEPhERD can better predict whether a candidate molecule will fit into a protein’s binding pocket and form the necessary interactions.
FlowER: Predicting Reactions with Physical Constraints
Another major model from the lab is FlowER, a generative AI approach to predicting chemical reactions. FlowER is designed to predict the reaction products that will result from combining different chemical inputs. What sets FlowER apart from other reaction prediction models is that the researchers built in an understanding of fundamental physical principles, such as the law of conservation of mass. They also compelled the model to consider the feasibility of the intermediate steps—the reaction mechanisms—that need to take place on the pathway from reactants to products.
These constraints improved the accuracy of the model’s predictions. This is a critical insight: by forcing the model to respect the laws of physics and chemistry, rather than simply learning statistical patterns from data, the researchers created a system that generalizes better to novel reactions.
Thinking about those intermediate steps, the mechanisms involved, and how the reaction evolves is something that chemists do very naturally. It is how chemistry is taught, but it is not something that models inherently think about. Coley’s group has spent considerable time figuring out how to ensure that their machine-learning models are grounded in an understanding of reaction mechanisms, in the same way an expert chemist would be.
What Is the Difference Between Statistical Learning and True Chemical Understanding?
This is a question that many researchers in the field are actively debating. A model that simply memorizes reaction outcomes from a training set may appear to perform well on familiar chemistry but will fail when confronted with novel substrates or reaction conditions. A model that understands the underlying principles—such as conservation of mass, electron movement, and the energetic feasibility of transition states—can make reliable predictions even in uncharted chemical space.
FlowER and ShEPhERD represent a shift toward the latter approach. By incorporating physical constraints and mechanistic reasoning, these models are not just pattern-matching engines. They are learning the rules of chemistry. This distinction is crucial for drug discovery, where the most valuable predictions are often about molecules that have never been synthesized before.
How Do These AI Models Actually Work in Practice?
For a chemist working in the pharmaceutical industry, the workflow might look like this. First, a target protein associated with a disease is identified. The chemist then wants to find a small molecule that can bind to that protein and modulate its activity. Instead of manually designing molecules or screening a physical library of compounds, the chemist uses a model like ShEPhERD to generate candidate molecules that are predicted to have high binding affinity based on their three-dimensional shape complementarity to the protein.
Once a promising candidate is identified, the chemist needs to figure out how to actually make it. This is where FlowER comes in. The model can predict the likely products of various reaction pathways, helping the chemist choose a synthetic route that is both feasible and efficient. The model can also flag potential side reactions or decomposition pathways that might complicate the synthesis.
Throughout this process, the AI is not replacing the chemist. It is acting as a powerful assistant that can explore vastly more options than a human could, and it can do so in a fraction of the time. The chemist still makes the final decisions, using their expertise to interpret the model’s predictions and design experiments.
Broader Implications for Drug Discovery and the Pharmaceutical Industry
The impact of this work extends far beyond the MIT lab. The pharmaceutical industry is under constant pressure to reduce the time and cost of developing new drugs. On average, it takes more than a decade and over a billion dollars to bring a new drug to market. A significant portion of that cost comes from the many candidates that fail in clinical trials after years of development.
AI models that can more accurately predict which compounds are worth pursuing—and which synthetic routes are most viable—have the potential to dramatically reduce this attrition rate. By improving the quality of candidates that enter preclinical testing, these models could save pharmaceutical companies billions of dollars and, more importantly, deliver new therapies to patients faster.
The work at MIT is also helping to democratize access to advanced chemical design capabilities. Smaller biotech companies, which may not have the resources to maintain large teams of synthetic chemists, can use these models to design and evaluate molecules with a level of sophistication that was previously only available to large pharmaceutical firms.
Beyond drug discovery, the same principles apply to other areas of chemistry, including materials science, agrochemicals, and specialty chemicals. Any field that involves designing and synthesizing new molecules could benefit from AI models that understand the underlying chemistry.
The Role of Laboratory Automation and Experimental Design
Coley’s lab is not solely focused on computational models. Students in the group also work on computer-aided structure elucidation, laboratory automation, and optimal experimental design. These areas are closely connected to the AI work. For example, a model that predicts reaction outcomes is most useful when it can be rapidly tested experimentally. Automated synthesis platforms can execute the predicted reactions, and the results can be fed back into the model to improve its accuracy.
This closed-loop approach—where computation and experimentation are tightly integrated—is a hallmark of Coley’s research philosophy. It reflects a broader trend in the field toward what is sometimes called “self-driving laboratories.” These are systems that can autonomously design experiments, execute them, analyze the results, and update their models, all without human intervention. While full autonomy is still a goal for the future, the components are being developed and tested in labs like Coley’s.
Why MIT Is a Unique Environment for This Work
Coley’s decision to stay at MIT after his PhD was influenced by the institution’s unique strengths. The fluidity across departments allows researchers to combine expertise from chemical engineering, computer science, and the Schwarzman College of Computing in ways that are difficult to replicate elsewhere. The caliber and enthusiasm of the students, combined with the strength of collaborations, create an ecosystem where interdisciplinary work thrives.
MIT has also made significant institutional investments in the intersection of AI and science. The Schwarzman College of Computing, launched in 2018, was designed specifically to support this kind of cross-cutting research. For Coley, this environment has been essential for pursuing a research agenda that requires deep expertise in both chemistry and machine learning.
The Future of AI in Chemistry
Looking ahead, the trajectory of this field is clear. Models will continue to become more sophisticated in their understanding of chemical principles. The goal is not just to predict outcomes but to design molecules and reactions that have never been conceived of before. This requires models that can reason about chemistry in a way that is analogous to, and perhaps eventually superior to, human experts.
One of the key challenges is data. While there are large databases of chemical reactions and molecular properties, they are still sparse relative to the vastness of chemical space. Models that can learn from limited data, or that can generate their own training data through simulation, will have a significant advantage. Another challenge is interpretability. Chemists need to trust the models, and that requires understanding why a model makes a particular prediction. Work on mechanistic grounding, as demonstrated by FlowER, is an important step in this direction.
The convergence of AI, automation, and chemical engineering is creating a new paradigm for how molecules are discovered and made. The work at MIT, led by researchers like Coley, is not just advancing the frontier of AI in chemistry—it is redefining what is possible in molecular science. For the pharmaceutical industry, and for anyone who relies on the medicines that industry produces, this is a development of profound significance.
Through research threads that span generative models, reaction prediction, structure elucidation, and laboratory automation, Coley’s group is systematically building the tools that will allow chemists to navigate the vast chemical universe with unprecedented speed and precision. The models are learning the rules of chemistry, and that learning is opening doors to molecules that could become the next generation of life-saving drugs.
