Using AI to Design and Predict Protein Structures
This patent describes a method for predicting or designing proteins by feeding detailed 3D structural information around a specific protein building block into a specially designed deep learning network.
Patent Number
US 12603151
Status
Active
Filing Date
July 23, 2020
Grant Date
April 14, 2026
Expiration
July 23, 2040
Claims
11
Assignee
Florida State University Research Foundation
Inventors
Wei Wu, Xiuwen Liu, Jinfeng Zhang, Yuan Zhang
Citations
0 forward · 56 backward
What it covers
This patent outlines a method for either predicting what a specific protein building block (called a 'target residue') should be, or for designing a protein with a new function. It starts by focusing a small, gridded 3D box, about 10 to 30 angstroms on each side, around a specific atom (like the Cα atom) of the target residue within a protein, as described in claim 1. This box collects detailed local structural information, including data about covalent bonds, bond angles, and dihedral angles. This information is then fed into a unique nine-layer 3D deep convolutional neural network (CNN), also detailed in claim 1, which processes the data to determine the target residue. Finally, based on this determination, the target residue can be modified to give the protein a new function, or a new protein can be designed to include this residue to achieve a specific function or help it fold correctly.
What it doesn't cover
- —Does not cover protein prediction or design methods that do not use a gridded box centered on a selected atom of a target residue.
- —Does not cover methods that use a convolutional neural network without the specific architecture described in claim 1, including the three parallel 3D convolution layers for bonds, angles, and dihedrals.
- —Does not cover protein design methods that do not involve modifying a target residue or designing a protein to achieve a new function or facilitate folding.
- —Does not cover protein prediction or design using machine learning models other than a convolutional neural network, such as recurrent neural networks or transformer models.
- —Does not cover methods where the gridded box is outside the specified size range of about 10 to 30 angstroms in length.
The clever bit
The truly clever part is the highly specialized architecture of the convolutional neural network. It includes three parallel 3D convolution layers, each specifically tuned to recognize fundamental protein structural features: covalent bond lengths, bond angles, and dihedral angles. This allows the network to deeply understand the local geometry of a protein in a way that generic neural networks might miss.
Why it matters
Understanding and designing proteins is crucial for developing new medicines, enzymes for industrial processes, and advanced materials. This patent offers a computational approach to tackle the complex problem of protein structure and function. By using a specialized deep learning model, it aims to make the process of engineering proteins more efficient and predictable, potentially accelerating drug discovery and biotechnology innovations.
Real-world examples
- 1.Computational drug discovery platforms
- 2.Enzyme engineering for industrial applications
- 3.De novo protein design software
- 4.Predicting protein-ligand binding sites
- 5.Developing new protein-based therapeutics
Generated by PatentBrief · Not legal advice · patentbrief.org
US 12603151 · 2026