Intuition and AI make a strange couple.
Intuition is hard to describe. It’s that gut feeling that gnaws at you, even if you don’t know why. We naturally build intuition through experience. Gut feelings aren’t always right; but they often creep into our subconscious to supplement logic and reasoning when making decisions.
AI, in contrast, rapidly learns by digesting millions of cold, hard data points, producing purely analytical—if not always reasonable—results based on its input.
Now, a new study in Nature Communications marries the odd pair, resulting in a machine learning system that captures a chemist’s intuition for drug development.
By analyzing feedback from 35 chemists at Novartis, a pharmaceutical company based in Switzerland, the team developed an AI model that learns from human expertise in a notoriously difficult stage of drug development: finding promising chemicals compatible with our biology.
First, the chemists used their intuition to choose which of 5,000 chemical pairs had a higher chance of becoming a useful drug. From this feedback, a simple artificial neural network learned their preferences. When challenged with new chemicals, the AI model gave each one a score that ranked whether it was worthy for further development as medication.
Without any details on the chemical structures themselves, the AI “intuitively” scored certain structural components, which often occur in existing medications, higher than others. Surprisingly, it also captured nebulous properties not explicitly programmed in previous computer modeling attempts. Paired with a generative AI model, like DALL-E, the robo-chemist designed a slew of new molecules as potential leads.
Many promising drug candidates were based on “collative know-how,” wrote the team.
The study is a collaboration between Novartis and Microsoft Research AI4Science, the latter based in the UK.
Down the Chemical Rabbit Hole
Most of our everyday medicines are made from small molecules—Tylenol for pain, metformin for diabetes management, antibiotics to fight off bacterial infections.
But finding these molecules is a pain.
First, scientists need to understand how the disease works. For example, they decipher the chain of biochemical reactions that give you a pounding headache. Then they find the weakest link in the chain, which is often a protein, and model its shape. Structure in hand, they pinpoint nooks and crannies that molecules can jam into to disrupt the protein’s function, thereby putting a stop to the biological process—voilà, no more headaches.
Thanks to protein prediction AI, such as AlphaFold, RoseTTAFold, and their offshoots, it’s now easier to model the structure of a target protein. Finding a molecule that fits it is another matter. The drug doesn’t just need to alter the target’s activity. It also must be easily absorbed, spread to the target organ or tissue, and be safely metabolized and eliminated from the body.
Here’s where medicinal chemists come in. These scientists are pioneers in the adoption of computer modeling. Over two decades ago, the field began using software to sift enormously large databases of chemicals looking for promising leads. Each potential lead is then evaluated by a team of chemists before further development.
Through this process, medicinal chemists build an intuition that allows them to make decisions efficiently when reviewing promising drug candidates. Some of their training can be distilled into rules for computers to learn—for example, this structure likely won’t pass into the brain; that one could damage the liver. These expert rules have helped with initial screening. But so far, no program can capture the subtleties and intricacies of their decision-making, partly because the chemists can’t explain it themselves.
I’ve Got a Feeling
The new study sought to capture the unexplainable in an AI model.
The team recruited 35 expert chemists at various Novartis centers around the world, each with different expertise. Some work with cells and tissues, for instance, others with computer modeling.
Intuition is hard to measure. It’s also not exactly reliable. As a baseline, the team designed a multiplayer game to gauge if each chemist was consistent in their choices and whether their picks agreed with those of others. Each chemist was shown 220 molecule pairs and asked an intentionally vague question. For example, imagine you’re in an early virtual screening campaign, and we need a drug that can be taken as a pill—which molecule would you prefer?
The goal was to reduce overthinking, pushing the chemists to rely on their intuition for which chemical stays and which goes. This setup differs from usual evaluations, where the chemists check off specific molecular properties with predictive models—that is, hard data.
The chemists were consistent in their own judgment, but didn’t always agree with each other—likely because of differing personal experiences. However, there was enough overlap to form an underlying pattern an AI model could learn from, explained the team.
They next built up the dataset to 5,000 molecule pairs. The molecules, each labeled with information on its structure and other features, were used to train a simple artificial neural network. With training, the AI network further adjusted its inner workings based on feedback from the chemists, eventually giving each molecule a score.
As a sanity check, the team tested the model on chemical pairs different from those in its training dataset. As they increased the number of training samples, performance shot up.
While earlier computer programs have relied on rules for what makes a promising medicine based on molecular structure, the new model’s scores didn’t directly reflect any of these rules. The AI captured a more holistic view of a chemical—a totally different approach to drug discovery than that used in classic robo-chemist software.
Using the AI, the team then screened hundreds of FDA-approved drugs and thousands of molecules from a chemical databank. Even without explicit training, the model extracted chemical structures—called “fragments”—that are more amenable to further development as medicines. The AI’s scoring preferences matched those of existing drug-like molecules, suggesting it had grasped the gist of what makes a potential lead.
Novartis isn’t the first company to explore a human-robot chemical romance.
Previously, the pharmaceutical company Merck also tapped into their in-house expertise to rank chemicals for a desirable trait. Outside the industry, a team at the University of Glasgow explored using intuition-based robots for inorganic chemical experiments.
It’s still a small study, and the authors can’t rule out human fallacies. Some chemists might choose a molecule based on personal biases that are hard to completely avoid. However, the setup could be used to study other steps in drug discovery that are expensive to complete experimentally. And while the model is based on intuition, its results could be bolstered by rule-based filters to further improve its performance.
We’re in an era where machine learning can design tens of thousands of molecules, explained the team. An assistant AI chemist, armed with intuition, could help narrow down candidates at the critical early stage of drug discovery, and in turn, accelerate the whole process.