AI Predicts 3D Protein Shapes from Genetic Code

This patent describes how a computer uses self-attention neural networks to predict the complex three-dimensional shape of a protein based on its simple amino acid sequence, a process crucial for understanding biology and developing new medicines.

What it covers

This method, performed by computers, predicts a protein's 3D structure from its amino acid sequence. First, it takes the amino acid sequence and creates an initial 'embedding' for every possible pair of amino acids in that sequence (Claim 1). Optionally, this initial embedding can be improved by comparing the protein's sequence to similar proteins from other species, creating a 'multiple sequence alignment' (Claim 2). Next, these initial pair embeddings are fed into a 'pair embedding neural network' that uses multiple 'self-attention neural network layers' to refine them into 'final embeddings' (Claim 1). Each self-attention layer updates an embedding by looking at other related embeddings, sometimes focusing only on pairs in the same 'row' or 'column' of a 2D arrangement (Claims 3-6). Finally, the computer determines the protein's predicted 3D structure using these refined final embeddings (Claim 1). For example, if you input the sequence for insulin, the system would output its predicted folded shape.

What it doesn't cover

• —Does not cover methods that determine protein structure using physical experiments like X-ray crystallography or cryo-electron microscopy.
• —Does not cover protein structure prediction systems that do not use self-attention neural network layers for processing pair embeddings.
• —Does not cover methods that predict protein structure without first generating embeddings for pairs of amino acids.
• —Does not cover systems that predict protein function or interactions without determining the 3D structure.
• —Does not cover predictions based on inputs other than an amino acid sequence, such as a protein's physical properties or environment.

The clever bit

The innovation lies in using 'self-attention neural networks' specifically on 'pair embeddings' of amino acids. This allows the system to efficiently learn complex relationships between distant parts of a protein, which is crucial for accurately predicting how the entire chain folds into a 3D shape, especially with the alternating row-wise and column-wise attention layers.

Why it matters

Understanding a protein's 3D shape is fundamental to knowing how it works. This technology speeds up scientific discovery in fields like drug design, where knowing a protein's structure helps scientists create molecules that can interact with it. It also aids in understanding diseases caused by misfolded proteins and designing new enzymes for industrial applications.

Real-world examples

• 1.DeepMind's AlphaFold
• 2.AlphaFold 2
• 3.Protein structure prediction services
• 4.Computational drug discovery platforms