Anyone who wants to delight their granny with a poem on her 90th birthday doesn’t need to be a poet nowadays: A short prompt in ChatGPT is all it takes, and within a few seconds the AI spits out a long list of words that rhyme with the birthday girl’s name. It can even produce a sonnet to go with it if you like.
Researchers at the University of Bonn have implemented a similar model in their study – known as a chemical language model. This does not, however, produce rhymes. Instead, the AI displays the structural formulas of chemical compounds that may have a particularly desirable property: They are able to bind to two different target proteins. In the organism, this means, for example, they can inhibit two enzymes at once.
"2In pharmaceutical research, these types of active compounds are highly desirable due to their polypharmacology," explains Prof. Dr. Jürgen Bajorath. The computational chemistry expert heads the AI in Life Sciences area at the Lamarr Institute for Machine Learning and Artificial Intelligence and the Life Science Informatics program at b-it (Bonn-Aachen International Center for Information Technology) at Uni Bonn. "Because compounds with desirable multi-target activity influence several intracellular processes and signaling pathways at the same time, they are often particularly effective - such as in the fight against cancer." In principle, this effect can also be achieved by co-administration of different drugs. However, there is a risk of unwanted drug-drug interactions and different compounds are also often broken down at different rates in the body, making it difficult to administer them together.
Finding a molecule that specifically influences the effect of a single target protein is no easy task. Designing compounds that have a predefined double effect is even more complicated. Chemical language models may help here in the future. ChatGPT is trained with billions of pages of written text and learns to formulate sentences itself. Chemical language models work in a similar way, but only have comparably very small amounts of data available for learning. However, in principle, they are also fed with texts, such as what are known as SMILES strings, which show organic molecules and their structure as a sequence of letters and symbols. "We have now trained our chemical language model with pairs of strings," says Sanjana Srinivasan from Bajorath’s research group. "One of the strings described a molecule that we know only acts against one target protein. The other represented a compound that, in addition to this protein, also influences a second target protein."
The model was fed with more than 70,000 of these pairs. This allowed it to acquire an implicit knowledge of how the normal active compounds differed from those with the double effect. "When we then fed it with a compound against a target protein, it suggested molecules on this basis that would act not only against this protein but also against another," explains Bajorath.
The training compounds with the double effect often target proteins that are similar and thus perform a similar function in the body. In pharmaceutical research, however, people are also looking for active ingredients that influence completely different classes of enzymes or receptors. To prepare the AI for this task, fine-tuning took place after the general learning phase. The researchers used several dozen special training pairs to teach the algorithm which different classes of proteins the suggested compounds should target. This is a bit like instructing ChatGPT not to create a sonnet this time, but instead a limerick.
After the fine-tuning, the model actually spat out molecules that have already been shown to act against the desired combinations of target proteins. "This shows that the process works," says Bajorath. In his opinion, however, the strength of the approach is not that new compounds exceeding the effect of available pharmaceuticals can immediately be found. "It is more interesting, from my point of view, that the AI often suggests chemical structures that most chemists would not even think of right away," he explains. "To a certain extent, it triggers ‘out of the box’ ideas and comes up with original solutions that can lead to new design hypotheses and approaches."
Srinivasan, Sanjana et al.
Generation of dual-target compounds using a transformer chemical language model.
Cell Reports Physical Science, 2024. doi: 10.1016/j.xcrp.2024.102255