Machine Learning Aids Rapid Design of Protein Therapeutics
Researchers at Ecole Polytechnique Fédérale de Lausanne (EPFL) in Switzerland have created a machine learning approach to scan millions of protein fragments and assess their structure and binding properties. Based on the surface chemistry and geometry of a protein, the developed software can determine a ‘fingerprint’ for each protein and predict how they might bind to various protein fragments. The researchers have now used their approach to design new protein ‘binders’ that have been created specifically to bind to proteins of therapeutic interest, such as the SARS-CoV-2 spike protein. The technique could allow researchers to create an array of therapeutic proteins very quickly, which could be particularly useful in cases where time is of the essence, such as future pandemics.
Numerous factors affect how and whether proteins will bind to each other, making it difficult to predict using just the human brain. However, identifying protein fragments that can interact with therapeutic protein targets in the body could yield enormous clinical prizes in treating various diseases. Thankfully, computers are well adapted to tasks that involve mind-numbing amounts of detailed data and complex permutations of the same. This latest approach involves deep learning, which can assess the subtle interplay between the different factors affecting binding.
“A puzzle piece is two-dimensional, but with protein surfaces, we are looking at multiple dimensions: chemical composition, such as positive versus negative charge interactions; shape complementarity, curvature, etc.,” said Anthony Marchand, a researcher involved in the study. “The idea that everything in nature that binds is complementary – for example, a positive charge binds with a negative charge – has been a long-standing idea in the field, which we captured in our computational framework.”
So far, the researchers have used their system to create a series of protein ‘binders’ that can attach to therapeutic targets, including the SARS-CoV-2 spike protein. This involved using their deep learning approach to create protein ‘fingerprints’ and then scanning through a database of protein fragments to find those that are predicted to bind well with the fingerprint. They then tested the potential of the fragments with the best predicted binding activity to actually bind their targets, first through a digital simulation and finally in the lab after the fragment had been synthesized.
“The fact that we’re able to design novel, site-specific protein binders in just a couple of months makes this method very interesting for therapeutics. It’s not just a tool: it’s a pipeline,” said Marchand. “Further advances in machine learning methods will help improve our method, but our work today already provides a strategy for developing innovative therapies to benefit patients through the rapid design of protein-based therapeutics – straight from the computer.”
Study in Nature: De novo design of protein interactions with learned surface fingerprints