Granted Patent B2

US 12,738,342 · App. 17/886,742 · Granted Sep 15, 2026

Inventors: Joshua Laniado (Los Angeles, CA); Julien Jorda (Los Angeles, CA); Matthias Maria Alessandro Malago (Santa Monica, CA); Thibault Marie Duplay (Los Angeles, CA); Mohamed El Hibouri (Los Angeles, CA); Lisa Juliette Madeleine Barel (Los Angeles, CA)

Assignee: Pythia Labs, Inc.

G16B15/20G16B15/30G16B40/20

Abstract

Claims (44)

1 . A method for designing a custom biologic structure for binding to a target via an artificial intelligence (AI)-powered scaffold docker module, the method comprising:

(a) generating, by a processor of a computing device, a candidate scaffold model, wherein the candidate scaffold model is a representation of at least a portion of a candidate peptide backbone, wherein the candidate peptide backbone is a prospective backbone of the custom biologic structure being designed and wherein the candidate scaffold model (i) comprises representations of peptide backbone atoms identifying types and locations of peptide backbone atoms, but (ii) excluding side chains;

(b) generating, by the processor, for the candidate scaffold model, a plurality of prospective scaffold-target complex models, each representing at least a portion of a complex comprising the target and the candidate peptide backbone at a particular pose with respect to the target;

(c) for each of the plurality of prospective scaffold-target complex models, determining, by the processor, a scaffold pose score, wherein determining the scaffold pose score for each particular one of the one or more prospective scaffold-target complex models comprises:

generating, based on the particular scaffold-target complex model, a corresponding representation; and

using the corresponding representation as input to a machine learning model that determines, as output, the scaffold pose score for the particular scaffold-target complex model, said machine learning model having been trained using representations of native scaffold-target complex models, each representing a peptide backbone of a particular protein and/or peptide, as oriented at a particular pose with respect to a target, in an existing native protein-protein and/or protein-peptide example complex, and wherein the scaffold pose score is a likelihood that the particular scaffold-target complex model represents a native complex;

(d) selecting, by the processor, a subset of the plurality of prospective scaffold-target complex models using the determined scaffold pose scores; and

(e) providing the selected subset of prospective scaffold-target complex models for use in designing the custom biologic structure for binding to the target.

2 . The method of claim 1 , wherein step (e) comprises populating at least an interface region of one or more of the selected subset of prospective scaffold-target complex models with amino acid side chains to generate one or more ligand models for use in designing the custom biologic structure.

3 . The method of claim 1 , wherein the target is a molecule associated with a particular disease.

4 . The method of claim 1 , comprising identifying, by the processor, an interface sub-region of the particular prospective scaffold-target complex model, the interface sub-region comprising representations of atoms of the candidate peptide backbone and/or target located in proximity to an interface between the candidate peptide backbone and/or target.

5 . The method of claim 1 , wherein the candidate scaffold model includes, for at least a portion of amino acid sites of the candidate backbone, a first side chain atom as a placeholder of potential side chains to be added.

6 . The method of claim 5 , wherein for each of the portion of amino acid sites, the first side chain atom is a beta-carbon atom or a hydrogen atom.

7 . The method of claim 1 , wherein the machine learning model is or comprises a neural network.

8 . The method of claim 1 , wherein step (a) comprises generating the candidate scaffold model by:

accessing a structural model of a template biologic; and

stripping at least a portion of amino acid side chain atoms from the structural model of the template biologic to generate the candidate scaffold model.

9 . The method of claim 1 , wherein the candidate scaffold model is a computationally generated backbone representing a candidate backbone not necessarily occurring in nature.

10 . The method of claim 1 , wherein the machine learning model is or comprises a neural network having been trained using training data comprising:

(A) a plurality of native complex models, each native complex model representing at least a portion of a native complex; and

(B) a plurality of artificially generated variant complex models.

11 . A system for designing a custom biologic structure for binding to a target via an artificial intelligence (AI)-powered scaffold docker module, the system comprising:

a processor of a computing device; and

a memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to:

(a) generate a candidate scaffold model, wherein the candidate scaffold model is a representation of at least a portion of a candidate peptide backbone, wherein the candidate peptide backbone is a prospective backbone of the custom biologic structure being designed and wherein the candidate scaffold model (i) comprises representations of peptide backbone atoms identifying types and locations of peptide backbone atoms, but (ii) excluding side chains;

(b) generate, for the candidate scaffold model, a plurality of prospective scaffold-target complex models, each representing at least a portion of a complex comprising the target and the candidate peptide backbone at a particular pose with respect to the target;

(c) for each of the plurality of prospective scaffold-target complex models, determine a scaffold pose score, wherein determining the scaffold pose score for each particular one of the plurality of prospective scaffold-target complex models comprises:

generating, based on the particular scaffold-target complex model, a corresponding representation; and

using the corresponding representation as input to a machine learning model that determines, as output, the scaffold pose score for the particular scaffold-target complex model, said machine learning model having been trained using representations of native scaffold-target complex models, each representing a peptide backbone of a particular protein and/or peptide, as oriented at a particular pose with respect to a target, in an existing native protein-protein and/or protein-peptide example complex, and wherein the scaffold pose score is a likelihood that the particular scaffold-target complex model represents a native complex;

(d) select a subset of the plurality of prospective scaffold-target complex models using the determined scaffold pose scores; and

(e) provide the selected subset of prospective scaffold-target complex models for use in designing the custom biologic structure for binding to the target.

12 . The system of claim 11 , wherein the instructions, when executed by the processor, cause the processor to, in step (e), populate at least an interface region of one or more of the selected subset of prospective scaffold-target complex models with amino acid side chains to generate one or more ligand models for use in designing the custom biologic structure.

13 . The system of claim 11 , wherein the target is a molecule associated with a particular disease.

14 . The system of claim 11 , wherein the instructions cause the processor to identify an interface sub-region of the particular prospective scaffold-target complex model, the interface sub-region comprising representations of atoms of the candidate peptide backbone and/or target located in proximity to an interface between the candidate peptide backbone and/or target.

15 . The system of claim 11 , wherein the candidate scaffold model includes, for at least a portion of amino acid sites of the candidate backbone, a first side chain atom as a placeholder of potential side chains to be added.

16 . The system of claim 15 , wherein for each of the portion of amino acid sites, the first side chain atom is a beta-carbon atom or a hydrogen atom.

17 . The system of claim 11 , wherein the machine learning model is or comprises a neural network.

18 . The system of claim 11 , wherein, at step (a), the instructions cause the processor to generate the candidate scaffold model by:

accessing a structural model of a template biologic; and

stripping at least a portion of amino acid side chain atoms from the structural model of the template biologic to generate the candidate scaffold model.

19 . The system of claim 11 , wherein the candidate scaffold model is a computationally generated backbone representing a candidate backbone not necessarily occurring in nature.

20 . The system of claim 11 , wherein the machine learning model is or comprises a neural network having been trained using training data comprising:

(A) a plurality of native complex models, each native complex model representing at least a portion of a native complex; and

(B) a plurality of artificially generated variant complex models.

Assignments (2)

Continuity (3)

References Cited (148)