Scientists at the University of Bonn have successfully developed an artificial intelligence (AI) model capable of predicting chemical compounds that serve dual functions, a significant breakthrough in pharmaceutical research. The study, published in Cell Reports Physical Science, was led by Prof. Dr. Jürgen Bajorath, a computational chemistry expert at the Lamarr Institute for Machine Learning and Artificial Intelligence and the Life Science Informatics programme at b-it (Bonn-Aachen International Center for Information Technology).
The AI, conceptualised as a chemical language model, is akin to large language models like ChatGPT used for generating text. However, instead of constructing sentences, this model outputs the structural formulas of chemical compounds. Specifically, it identifies those that have the potential to bind to two distinct target proteins. This functionality makes such compounds highly effective as medications due to their ability to influence multiple intracellular processes and signalling pathways simultaneously, a property known in pharmaceuticals as polypharmacology.
In traditional drug administration, achieving a similar polypharmacological effect often necessitates the co-administration of multiple drugs. This approach carries the risk of adverse drug-drug interactions and poses challenges with regards to the differing metabolic rates of each compound.
Designing drugs with dual-target capabilities is particularly challenging. However, the introduction of chemical language models may revolutionise this aspect of drug development by learning from molecular structures expressed in SMILES strings—a notation that encodes chemical structures as sequences of letters and symbols.
The research team trained the model using 70,000 pairs of chemical strings. Each pair included a molecule known to target one protein and another influencing two proteins. This extensive training enabled the AI to discern the structural differences between mono-target and dual-target compounds, subsequently allowing it to predict molecular structures that could potentially achieve dual functionality.
Further fine-tuning of the model enhanced its predictive power. Using several dozen special training pairs, the researchers conditioned the AI to identify compounds that target different protein classes. The AI successfully predicted molecules previously confirmed to interact with desired protein combinations, validating its accuracy and efficacy.
While the discovery of new, superior compounds is the ultimate goal, Bajorath highlights the AI's capability to devise unconventional chemical structures, offering new hypotheses and strategies for drug design. This innovation marks a significant step in the direction of more sophisticated, targeted, and effective pharmaceutical therapies, though further research and development are needed to translate these AI-generated predictions into clinically viable drugs.
Source: Noah Wire Services