The construction of novel protein structures remains a problem in protein engineering for biomedical and chemical applications. 

The field of biology is a beautiful and intricate web. DNA is the central regulator of every biological process that keeps humans alive. DNA is the master weaver that encodes proteins. But the human body, like a musical instrument, can quickly lose its pitch. After all, we must deal with nature's unpredictable and unrelenting forces, such as cancer, viruses, and other diseases. 

3D frames

A diffusion model over rigid bodies in 3D (referred to as frames) has successfully developed unique, functional protein backbones that have not been observed in nature in this field of research. Imagine if we could quickly build treatments and vaccinations against emerging diseases. Imagine we had gene-editing technology that could automatically generate proteins to fix cancer-causing DNA mistakes. Drug development, diagnostics, and other industrial applications rely on identifying proteins that can tightly attach to targets or accelerate chemical reactions, but this process is time-consuming and costly.

The researchers developed "FrameDiff," a computational tool for building new protein structures to expand protein engineering skills beyond what nature has generated. The machine learning approach produces "frames" aligned with the inherent features of protein structures, allowing it to manufacture novel proteins independently of prior designs and create novel protein structures.

FrameDiff structure

Proteins are large macromolecules consisting of numerous atoms linked together by chemical interactions. Like the spine, the "backbone" refers to the protein's most crucial atoms in determining its 3D structure. The backbone comprises identical triplets of particles in bonding and elemental composition. Using concepts from differential geometry and probability, researchers realised this pattern might be used to develop machine learning systems. The frames serve this purpose. These sets of three can be modelled in mathematics as rigid bodies with a position and rotation in three dimensions; these bodies are called "frames" and are common in physics. 

Diffusion models

AlphaFold2 is a deep learning system developed by DeepMind in 2021 that can predict protein three-dimensional structures from their sequences. The two most essential processes in developing synthetic proteins are creation and prediction. Generating novel protein structures and sequences is called "generation," while "prediction" refers to determining a sequence's predicted three-dimensional structure. AlphaFold2's usage of frames to model proteins is not coincidental. Diffusion models, a generative AI technique prevalent in picture generation with programmes like Midjourney, inspired SE(3) diffusion and FrameDiff, which expand on frames by including them in the model. 

The top models from both ends were compatible because of the common frameworks and concepts used in protein structure development and prediction. The University of Washington's Institute for Protein Design is already employing SE(3) diffusion to design and verify new proteins in the lab. They took RosettaFold2, a protein structure prediction tool similar to AlphaFold2, and merged it with SE(3) diffusion to create "RFdiffusion." The production of particular protein binders for rapid vaccine design, engineering of symmetric proteins for gene transport, and robust motif scaffolding for precise enzyme design are all made easier with this new tool for protein designers. 

Conclusion

FrameDiff's future endeavours include enhancing generality to situations that combine numerous criteria for biologics, such as pharmaceuticals. Another extension would be to apply the models to all biological modalities, such as DNA and tiny molecules. The team believes that by expanding FrameDiff's training on larger datasets and improving its optimisation method, it will be able to build foundational structures with design capabilities comparable to RFdiffusion while retaining FrameDiff's natural simplicity.

Image source: Unsplash