Deep learning has radically changed the field of artificial intelligence. Today, neural networks are revolutionizing the way we discover new drugs. We chat with Francesca Grisoni, Assistant Professor in the Department of Biomedical Engineering, who is at the forefront of research into the automation of drug design. “Ultimately, we want to increase human intelligence for the development of new drugs.”

Let’s start with the beginning. As we all know, drug development is a very laborious process. The estimated number of molecules that could be used as starting points for drug discovery is estimated to be around 10⁶⁰-ten¹⁰⁰, which is much more than there are stars in the observable universe. So finding a drug molecule that has the properties you want is like finding a needle in a haystack.

“Among other things, most drug molecules are designed to modulate specific target proteins in our cells. By interacting with these targets, a drug can inhibit or promote the activity of this protein. This effect, known as bioactivity, can be used to cure or prevent disease, ”says Grisoni.

“The central idea in drug design is that you want to identify molecules that are highly active towards the intended targets, and not towards other targets that may cause unwanted side effects. Because of the large number of possible molecules that one might envision, navigating this “chemical space” efficiently is extremely difficult. “

And it does not stop there. Once you find a potential candidate drug molecule (or compound), you still need to produce the molecule in the lab (a process known as chemical synthesis) and test it in increasingly complex experiments. This is not only very time consuming (a lot of molecules turn out to be misfires), but also very expensive. This is one of the reasons why the big pharmaceutical companies so jealously guard their hard-earned patents.

IT assistance in drug discovery

Not surprisingly, then, thirty years ago, medicinal chemists and biologists turned to computers to help them speed up the process of drug discovery and development. “Computers can help select molecules that are most likely to be effective for the intended purpose and synthesize,” says Grisoni.

“One way to solve the problem is what we call virtual screening, where you use computational methods to choose candidates to test from a library of molecules that you know can be synthesized. These libraries are much smaller than the entire chemical universe (typically between 103 and 106 molecules), so they are easier to navigate. In some cases, however, one may want to explore different regions of chemical space, which are not included in such virtual screening libraries.

This is where “de novo” design, where you design molecules “from scratch”, can come to the rescue. “The de novo design has the added benefit that you can generate molecules focused on the goals you want to achieve, including, hopefully, some that no one else has thought of yet.”

From rules-based design to deep learning

But how do you build such molecules from scratch? Traditionally, this has been done by following a specific set of rules. “Think about how grammar works in language. If you just put a bunch of words together, you won’t end up with a meaningful sentence. So you need rules. Likewise, you can come up with rules on how to assemble atoms or molecular fragments together into a compound that not only makes chemical sense, but also has the desired biological properties. “

But, just like computer scientists who tried unsuccessfully in the 80s and 90s of the last century to automate translation by developing a rulebook, the rules-based approach in the design of medicine may go to the extreme. against its limits.

“In specific cases, you quickly realize that the rules are either too restrictive or too complex”, explains Grisoni. This is where machine learning, and more specifically deep learning, comes to the rescue. This allowed computer drug designers like Grisoni to automatically learn not only the “grammar” of known compounds (what elements are needed to make a valid molecule that can be synthesized?), But also its “semantics” (what elements are necessary to have the desired bioactivity on a given target?).

Make molecules “smile”

To achieve all this, researchers are using deep learning models borrowed from natural language processing, where AI has also revolutionized the world of machine translation and speech recognition (bringing us must-have applications like Google Translate and Siri). In order to be able to use NLP in drug design, the structure of molecules first had to be expressed as a string of words.

Fortunately, such a language has been available since the 80s: SMILES (see image). “By adding one character at a time to complete the SMILES chain, the PNL model is able to automatically generate new molecules. This process is not random. New characters are chosen based on what the model has learned about them. previously available data, ”says Grisoni. “Compare it to Google Search, which automatically completes your search entry based on previous queries.”

However, unlike Google Search or Translate, Grisoni and his colleagues face a delicate problem, specific to the world of drug design: the lack of large training data for computer algorithms before they can be used for. generate new molecules. “Large datasets for deep learning in drug design are quite rare. In some cases you may only have a handful of compounds known to work on a given target,” she explains. .

Make rare data work

This was one of the challenges she faced for a research paper on generative artificial intelligence, which she wrote with former colleagues at ETH Zürich and was recently published in the journal. Scientists progress. In the article, the researchers for the first time combine a ‘rule-less’ deep learning approach to generate bioactive molecules with on-chip synthesis, a form of miniaturized automated synthesis that further minimizes the amount of manual work required.

To get around the problem of small data, the researchers used a method called transfer learning.

“The basic idea is that you are mining data that is in some way related to your problem, but for which many other examples might be available, even if it is not exactly the data that you need. where someone is writing a scientific paper for the first time. Maybe they’ve read 50 similar papers before, then they are already able to start writing their own. But, of course, they didn’t start from scratch , they have learned to read and write all their lives. “

Likewise, we preform our deep learning models on tens of thousands of molecules that have general properties of interest to our goal. Once the models have learned enough information, we ‘refine’ them on a set. more specific, focused on what we want to achieve, like being active in a certain target protein. And this has been shown to be effective on several occasions, with sets as small as five molecules in the second phase of training! Science Advances article, we used 40. “

Make better decisions faster

Ultimately, Grisoni and his colleagues managed to identify 12 new compounds that are bioactive for liver X receptors, which have become promising drug targets due to their regulatory role in lipid metabolism and inflammation.

Of course, as is always the case with innovative research, there is still a long way to go. For example, the chemical space in the experiment was limited to 17 one-step reactions to ensure that the compounds were compatible with on-chip experiments. Grisoni also points out that the structural diversity of the new bioactive compounds is still quite limited and may require further expansion.

However, the researcher of Italian origin is quite satisfied with the results. “Our study is a pioneer in the integration of AI models in ‘chemical language’ for the design of molecules with automated synthesis in a miniaturized system. We are facing unprecedented opportunities thanks to new AI technology and interdisciplinary collaboration in the field of molecular design and synthesis, drug discovery and In the future, approaches like ours will help medical chemists to make “better decisions, faster”.