How machine learning can help you build new proteins

The natural protein universe is enormous, but designing beyond new proteins that are still not observed in nature can bring new functions and solve problems in medicine and materials science. The past few years mark the golden age of de novo protein design. Machine learning methods have provided an unprecedented level of modeling accuracy. This progress allows researchers to design protein structures with specific functional properties that have never been observed before. This is particularly interesting in sustainability issues such as biotechnology applications, therapeutic drug development, and plastic decomposition.

One of the key features of functional proteins, biomolecules with complex structures, are inherent structural flexibility. However, current designs lack this important feature.

For a team of researchers at the Heidelberg Theory Institute (HITS) and the Max Planck Institute for Polymer Research (MPIP), this was the starting point for deliberating whether proteins could be designed with custom flexibility from scratch. They presented the results of their work at the International Conference on Machine Learning (ICML) held in Vancouver, Canada.

Flow Matching: Model of de novo protein

“We wanted to build a model where we learn how to produce proteins so that their structure is flexible at a given location,” says first author vSevolod Viliuga (MPIP).

To that end, the team introduced a framework for generating flexible protein structures. This framework is based on both protein backbone flexibility and neural networks trained to predict generative models of protein structure.

“Natural proteins are extremely good at achieving tasks because they have the flexibility where they need them,” says co-author Leif Seute (Hits). “We can now design new proteins that mimic this important property.”

Hit group leader Jan Stühmer adds, “It's an extension of the geometric algebraic flow matching model we developed last year: GAFL.”

GAFL is three times faster than comparable models, not only achieves high specification, but also resembles natural proteins in many ways.

Ultimately, the team showed that the model could produce proteins with the desired flexibility pattern, even in patterns that are unusual for natural proteins. “This work is a step forward in designing new proteins for applications that require flexibility, such as enzyme catalysts,” said Frake Gräter (MPIP), one of the team leaders.