Protein sequence-to-structure learning: Is this the end(-to-end revolution)?

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Deep learning's impact on protein structure prediction has grown significantly since CASP13.
• CASP14 demonstrated deep learning's capability to achieve near-experimental accuracy in protein structure prediction.

Purpose of the Study:

• To provide an overview of novel deep learning approaches in protein structure prediction.
• To offer an expert opinion on recent advancements in the field, particularly those used in CASP14.

Main Methods:

• Geometric learning on graph, Voronoi tessellation, and point cloud representations.
• Pretrained protein language models utilizing attention mechanisms.
• Equivariant architectures that maintain 3D spatial symmetry.
• Integration of large meta-genome databases.
• Combination of diverse protein representations.
• Development of end-to-end differentiable models from sequence to 3D structure.

Main Results:

• Deep learning methods achieved unprecedented accuracy in protein structure prediction by CASP14.
• Novel approaches have pushed the boundaries of what is possible in predicting protein structures from sequences.
• The integration of various machine learning advancements has been key to this success.

Conclusions:

• Deep learning is now indispensable in protein structure prediction, reaching near-experimental accuracy.
• Emerging methods like geometric learning and protein language models represent the future of the field.
• Continued innovation in deep learning promises further acceleration in understanding protein structures.