That same principle applies in the molecular world, Özçelik explains. In recent years, the rise of AI has led to major changes in drug research. “When you start a project to find a new medicine, you have a choice of around 10 to the power of 60 different molecules. That’s even more than the number of stars in the universe.” (For comparison: a million is 10 to the power of 6, ed.)

“We call this the “chemical space,” and it’s enormous. But thanks to mathematical models—and especially advances in deep learning—we can predict chemical properties much more efficiently and generate new molecules.”

String of beads

Over the past four years, Özçelik explored how deep learning can be used as effectively as possible in the search for new medicines. Understanding molecular language played a key role in this. 

“You can describe any molecule using language by breaking it down into its building blocks. A molecule is essentially nothing more than a ‘string’ of components. A protein in your body can be translated into a sequence of amino acids represented by letters, and a small medicinal molecule can be represented as a sequence of atoms.”

Step by step, Özçelik learned the language of molecules, almost at the same time as he was taking Dutch courses. There was little doubt about leaving Istanbul for Eindhoven—“this project brings together all my passions”—but he did set one condition for himself: he had to invest in learning the language of his new environment.

“In both cases, you see letters and words—but what is the real meaning of an entire sentence? Only when the semantics are clear can you improve deep learning models. We feed them molecular language so the system keeps getting better at making accurate predictions. You can compare it to the smart keyboard on your phone, which predicts your sentences based on initial letters and words.”

Chemical search

During his PhD project, Özçelik developed five deep learning models that can now be used by researchers worldwide. “Each model has its own application. We started by establishing practical guidelines to improve the prediction of chemical properties and built general tools to support this.”

Together with researchers from the TU/e research group Chemical Biology and Italian collaborators, he then explored several case studies to make chemical searches more efficient. “We developed a model that can accurately identify where binding sites are located to a clinically important protein family, which is crucial for molecular interaction.”

To optimize what the body does with a drug—such as improving absorption—Özçelik developed a predictive model for suitable binding partners. “For researchers in the lab, this meant a much more efficient pipeline, as they only needed to test a limited subset of molecules.”

From negative to positive

He is most excited about a completely new approach to using deep learning in the search for new medicines. “We started feeding the model data in the opposite way. In the pharmaceutical field, far more ‘failed’ than successful molecules are found, so there is an abundance of ‘negative’ data.”

What initially seemed like an absurd idea, turned out to be surprisingly successful. “Extensive lab testing also shows that we developed a highly accurate model, in which a positive molecule can emerge from negative data. That’s good news for medicinal chemists.”

Of course, you have to be careful not to focus solely on designing “just another” model, Özçelik emphasizes. “By taking molecular language as our starting point and thinking outside the box, we aim to make developments in the field more robust. A solid foundation that other researchers can build on. And if that results in state-of-the-art models, that’s a very welcome bonus.”