Machine learning methods improve the accuracy of inverse protein folding for drug design

The AI ​​approach, developed by researchers at the University of Sheffield and AstraZeneca, could facilitate the design of proteins needed for new treatments.

Their research published in the journal Nature Machine IntelligenceSheffield computer scientists have worked with AstraZeneca and the University of Southampton to develop a new machine learning framework that demonstrates the possibility of being more accurate with inverse protein folding than existing cutting edge methods.

Reverse protein folding is an important process for creating novel proteins. It is the process of identifying amino acid sequences, which are components of a protein, folded into the 3D protein structure of interest, allowing the protein to perform certain functions.

Protein engineering plays an important role in drug development by designing proteins that can bind to specific targets in the body. However, this process is difficult due to the complexity of protein folding and the difficulty in predicting how amino acid sequences interact to form functional structures.

Scientists have turned to machine learning to more accurately predict which amino acid sequences fold into stable functional protein structures. These models are trained on a large dataset of known protein sequences and structures to improve inverse folding prediction.

A new machine learning framework called MapDiff from the University of Sheffield, AstraZeneca and the University of Southampton surpasses the most cutting-edge AI in successfully making predictions in simulated tests. The results are promising grounds for further development of the technology, and if successful, can accelerate the design of key proteins needed to develop new vaccines and gene therapy, as well as other therapeutic modalities.

It also complements other recent advances, such as Alphafold, which predicts the 3D structure of a protein by inverting the approach by starting with protein folding and then obtaining potential amino acid sequences.

“This work represents an important advance in using AI to design proteins with desired structures. It addresses these fundamental challenges in biology,” says Haiping Lu, a professor of machine learning at the University of Sheffield and a corresponding author of the study.

Peizhen Bai, a senior machine learning scientist at Astrazeneca, developed AI as part of his PhD. At the Computer Science School at the University of Sheffield, she said, “During my PhD, I was motivated by the possibility of accelerating the biological discovery of AI. Our method, MapDiff, is proud to help us design protein sequences that are likely to fold into the 3D structure we want.