Deep learning has radically changed the field of AI. Now neural networks are revolutionizing the way that we discover new drugs. We talk to Francesca Grisoni, assistant professor at the department of Biomedical Engineering, who is at the forefront of research into automating drug design. “In the end, we want to augment human intelligence for the development of novel medicines.”
Let’s start at the beginning. As we all know, the development of drugs is a very laborious process. The estimated number of molecules that could be used as starting points for drug discovery is estimated to range in the order of 1060–10100, which is way more than there are stars in the observable universe. So, finding a drug molecule that has the properties you desire 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 might inhibit or promote the activity of that protein. This effect, known as bioactivity, can be leveraged to cure or prevent diseases,” explains Grisoni.
“The central idea of drug design is that you want to identify molecules that are highly active toward the intended targets, and not toward other targets that might cause unwanted side effects. Because of the large number of possible molecules that one could consider, navigating this ‘chemical space’ efficiently is extremely hard.”
And it doesn’t end there. Once you find a potential candidate drug molecule (or compound), you still have to produce the molecule in a lab (a process known as chemical synthesis), and test it in increasingly complex experiments. This is not only very time-consuming (many molecules turn out to be duds), but also very expensive. It’s one of the reasons why big pharma so jealously guards their hard-won patents.
Computer assistance in drug discovery
It won’t come as a surprise then that some 30 years ago medicinal chemists and biologists turned to computers to help them speed up the process of drug discovery and development. “Computers can help in selecting the molecules that are most likely to be effective for the intended purpose and synthesizable,” explains Grisoni.
“One way to tackle the problem is what we call virtual screening, where you use computational methods to choose what candidates to test from a library of molecules that you know can be synthesized. These libraries are much smaller that the whole chemical universe (usually between 103 and 106 molecules), so they are easier to navigate. In some cases, however, one might want to explore different regions of the chemical space, that are not included in such virtual screening libraries.”
Here is where ‘de novo’ design, where you design molecules ‘from scratch,’ can come to the rescue. “De novo design has the added advantage that you can generate molecules focused on the goals you want to achieve, including hopefully some that nobody else has thought of yet.”
From rule-based design to deep learning
But how do we build such molecules from scratch? Traditionally, this has been done by following a specific set of rules. “Think of how grammar works in language. If you just put a bunch of words together, you won’t get a sentence that makes sense. So, you need rules. Similarly, 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 the computer linguists who tried unsuccessfully in the 80s and 90s of the last century to automate translation by coming up with a rulebook, the rule-based approach in medicine design may run against its limits.
“In specific cases, you soon realize that the rules become either too restrictive or too complex,” says Grisoni. It is here that machine learning, and more specifically deep learning, comes to the rescue. It has allowed computational 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 needed to have the desired bio-activity on a given target?).
Making molecules ‘smile’
To achieve all of this, the researchers use deep learning models borrowed from natural language processing, where AI has also revolutionized the world of automatic translation and speech recognition (bringing us such unmissable apps as Google Translate and Siri). To be able to use NLP in drug design, the structure of the molecules first had to be expressed as a string of words.
Luckily, such a language has been available since the eighties: SMILES (see image). “By adding one character at a time to complete the SMILES string, the NLP model is able to automatically generate new molecules. This process is not random. The new characters are chosen based on what the model has learnt on previously available data,” explains Grisoni. “Compare it to Google Search, which automatically auto-completes your search entry based on earlier queries.”
However, in contrast to Google Search or Translate, Grisoni and her colleagues face a tricky problem that is specific to the world of drug design: the lack of big training data for the computer algorithms before they can be used to generate new molecules. “Large datasets for deep learning in drug design are quite scarce. In some cases, you may have only a handful of compounds which are known to work on a given target,” she explains.
Making scarce data work
It was one of the challenges which she had to tackle for a research paper on generative AI, which she wrote together with former colleagues at ETH Zürich and was published recently in the journal Science Advances. In the paper, the researchers for the first combine a ‘rule-free’ deep learning approach to generate bioactive molecules with on-chip synthesis, a form of miniaturized automated synthesis that further minimizes the amount of manual labor needed.
To get around the small-data problem the researchers used a method called transfer learning.
“The basic idea is that you leverage data that is somehow related to your problem, but for which many more examples might be available, even if it’s not exactly the data you need. Think of when someone writes a scientific article for the first time. Maybe they read 50 similar documents before, and then they are already able to start writing theirs. But, of course, they didn’t start from scratch, they have been learning how to read and write their whole lives.”
“Similarly, we pre-train our deep learning models on tens of thousands of molecules that have general properties that are interesting for our goal. Once the models have learnt sufficient information, we ‘refine’ them on a more specific set, focused on what we want to achieve, such as being active in a certain target protein. And this has shown to work in multiple occasions, with sets as small as five molecules in the second phase of training! In the Science Advances paper, we used 40.”
Making better decisions faster
In the end, Grisoni and her colleagues managed to identify 12 novel bioactive compounds for so-called liver X receptors, which have emerged as promising drug targets due to their regulatory role in lipid metabolism and inflammation.
Of course, as is always the case for 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 the on-chip experiments. Grisoni also points out that the structural diversity of the novel bioactive compounds is still rather limited and might need further expansion.
Still, the Italian-born researcher is quite happy with the results. “Our study pioneers the integration of ‘chemical language’ AI models for molecule design with automated synthesis in a miniaturized system. We are facing unprecedented opportunities driven by new AI technology and interdisciplinary collaboration in the field of molecular design and synthesis, drug discovery, and beyond. In the future, approaches like ours will support medicinal chemists to make ‘better decisions faster.'”
Originally Published by PhysOrg