Leipzig. Medicine is becoming more molecular, especially in rheumatology. Understanding the molecular pathological basics is the key to new therapeutic approaches. Developing a suitable active substance that docks specifically to the desired target structure is time-consuming and costly. Unless researchers outsource this work: to powerful artificial neural networks that they have previously trained with clinical and genetic patient data and structures of known biologically relevant proteins and drugs.
Professor Jens Meiler, head of the Institute for Drug Development, which was newly founded at the University of Leipzig in 2020, assumes that this will fundamentally change how doctors approach diseases. “We’ll be sitting here in five years and talking about completely different topics,” he predicted at the opening event of the German Rheumatology Congress 2023. Keywords: precision and personalized medicine. Both begin in oncology and are also useful for the often very heterogeneous clinical pictures in rheumatology.
Tens of billions of active ingredients already pre-designed
The prerequisites are there: “We have almost infinite amounts of data, and we have computer clusters that are large enough that we can build very, very large networks.”
If one limits the molecular size realistically for drugs, there are theoretically still 1060 possible structures for therapeutic agents, according to the chemist and structural biologist. “That’s more than atoms in the universe.” Too many to commission chemists in the laboratory to synthesize them. And not necessary either. Because there are companies with virtual libraries with around 40 billion potential active ingredients. Meiler: “They didn’t synthesize them, but they tell us: ‘We have the synthesis paths, and if you say you need the molecule, we’ll synthesize it in four weeks'” – regardless of whether it’s a small molecule, vaccine or antibody.
Protein folding determines function
The three-dimensional structure is decisive for the biological effect of proteins, both endogenous and externally supplied. “It’s not the primary structure of the protein, it’s not the order of the amino acids,” explained Meiler. “Only if they arrange themselves correctly in three-dimensional space can they bind to the molecule, which could possibly be the drug.”
And this is where neural networks come into play. AlphaFold can calculate protein conformation based on amino acid sequence with high accuracy. In no time. Even if the protein was not previously known to the network. It took Meiler and his team of ten eight years to determine the three-dimensional fold of Caveolin-1. It is a membrane-bound protein on the surface of various cells. Its reduced expression in lung tissue, for example, plays a role in idiopathic pulmonary fibrosis. AlphaFold promptly generated a model that was identical to Meiler’s results – without knowing them. “There is nothing comparable in the protein database,” emphasized Meiler. “We determined the structure after AlphaFold was trained. AlphaFold couldn’t have known what that thing looked like.”
revolutionary development
Since the source code was released in 2021, the number of scientific articles on it has exploded. “We have here a revolution in structural biology that will lead to us knowing the structure of every protein, the structure of every mutant that causes disease. And in the next ten years we will be able to predict molecules that specifically address these mutants,” predicted Meiler. With artificial intelligence, scientists could use the 40 billion molecules in the databases to screen out the molecule that not only binds to the target protein in the body, but also to a specific conformation of the protein with the mutation in the patient in question.