How AI Designs New Protein Sequences for Specific Shapes
This patent describes a method where a computer-powered neural network fills in missing parts of a protein's amino acid sequence, ensuring the final protein has a desired 3D shape or "fold structure."
Patent Number
US 12655176
Status
Active
Filing Date
December 4, 2020
Grant Date
June 16, 2026
Expiration
December 4, 2040
Claims
22
Assignee
University of Toronto
Inventors
Alexey V. STROKACH, Philip M. KIM, David B. ROMERO, Carlos CORBI VERGE, Alberto Perez Riba
Citations
1 forward · 11 backward
What it covers
The patent describes a computer-based method for designing protein sequences. It starts by taking an incomplete list of amino acids (called "node elements") for a protein and a map of how these amino acids should physically interact to form a specific 3D shape (an "edge index" defining a "desired fold structure"). A neural network then processes this information. It first figures out improved details about the interactions between amino acids ("enhanced edge attribute values") based on the known amino acids. Then, it uses these improved interaction details to determine better amino acid characteristics ("enhanced amino acid values") for each position. Finally, for any amino acid positions that were initially missing, the network generates the specific amino acid that should go there, completing the protein sequence to match the desired shape. For example, if you wanted a protein with a specific binding pocket, you could provide the known amino acids around that pocket and the desired distances between them, and the network would fill in the rest of the sequence.
What it doesn't cover
- —Does not cover generating protein sequences without defining a desired physical interaction or "fold structure" beforehand.
- —Does not cover methods that predict a protein's 3D structure from an existing amino acid sequence.
- —Does not cover protein design approaches that do not use a computer-implemented neural network.
- —Does not cover generating protein sequences without an iterative refinement process of edge and node attributes.
- —Does not cover designing proteins where the "edge attribute value set" does not represent a characteristic of physical interaction, such as a Euclidean distance.
The clever bit
The novelty lies in the iterative refinement process where the neural network simultaneously enhances both the understanding of physical interactions (edge attributes) and the characteristics of individual amino acids (node attributes) based on a desired 3D structure. This allows it to fill in missing amino acids while ensuring the generated sequence is consistent with the target fold.
Why it matters
Proteins are the workhorses of biology, performing nearly all cellular functions. Being able to design new protein sequences that reliably fold into specific, desired 3D shapes is crucial for developing new drugs, enzymes, and biomaterials. This technology could accelerate the discovery of novel therapeutics or industrial catalysts by allowing researchers to computationally explore a vast space of potential protein designs.
Real-world examples
- 1.Designing novel enzymes for industrial processes
- 2.Creating new therapeutic proteins with specific binding sites
- 3.Developing custom antibodies
- 4.Engineering proteins for biosensors
- 5.Designing stable protein scaffolds for drug delivery
Generated by PatentBrief · Not legal advice · patentbrief.org
US 12655176 · 2026