Recursive transformers for AI-based protein-protein interaction and drug design

1. A method, comprising:

a) receiving, at a processor, representations of a plurality of protein-protein complexes;

b) using the representations of the plurality of protein-protein complexes to train a neural network to obtain a representation of a protein-protein complex, given a representation of a constituent target complex of that protein-protein complex:

ii) wherein the constituent target complex is a protein or protein-protein complex,

iii) wherein the neural network is configured to proceed recursively such that:

(1) for each iteration of the recursion, the neural network is configured to generate and output a representation of a candidate protein, if any, in complex with the constituent target complex,

(2) for each iteration of the recursion, a representation of the complex of the generated candidate protein (the output of the iteration) and the constituent target complex (the input of the iteration) is passed back into the neural network as input for the next iteration of the recursion;

c) using the trained neural network to obtain a representation of a candidate protein-protein complex, given a representation of a constituent target complex;

d) synthesizing a constituent protein of the candidate protein-protein complex.

2. The method of claim 1 , wherein the neural network is a transformer.

3. The method of claim 2 , wherein during each iteration of the recursion, the representation of the generated candidate protein includes sequence, structure, and docking information.

4. The method of claim 3 , wherein the transformer is multicapitate (has two or more heads) including a sequence head that generates the sequence and a structure head that generates the structure and docking representation of each output protein.

5. The method of claim 4 , wherein the transformer architecture is encoder-decoder.

6. The method of claim 5 , wherein embeddings of the constituent target complex structure and sequence are input into the encoder, and wherein the context array output of the final output layer of the encoder is input into the cross attention layers of the decoder.

7. The method of claim 6 , wherein the sequence and structure representation of each constituent protein of the inferred protein-protein complex is generated via autoregression.

8. The method of claim 7 , wherein the structure parameters are specified via voxels of a three dimensional grid in which the representative atom(s) of each amino acid residue is located.

9. The method of claim 8 , wherein the structure head's loss function uses a penalty to exclude already occupied voxels from the distribution of probable voxels for the generated residue during any given iteration of the autoregression.

10. A method, as in the method of claim 9 , for obtaining sequence, structure, and docking site representations of the constituent proteins of a protein-protein complex given a constituent target complex, and for synthesizing a constituent protein of the protein-protein complex, wherein the method is also for obtaining and synthesizing an effective ligand drug of a given target receptor, the method further comprising:

a) obtaining for each of a plurality of candidate ligands, the given target receptor's sequence and structure representation when in complex with the candidate ligand;

b) using the representation of each of the plurality of candidate conformations of the target receptor respectively as input into the transformer neural network; to obtain a respective candidate protein-protein complex associated with each candidate ligand;

c) assessing, for each of the candidate protein-protein complexes, the biological activity and associated effectiveness;

d) selecting the most effective candidate ligand based on the assessment;

e) synthesizing the ligand.

11. A method, comprising:

a) receiving, at a processor, representations of a plurality of antigen-antibody-protein complexes, wherein the plurality of antigen-antibody-protein complexes includes at least one of:

i) a plurality of complexes of antigens and antibodies,

ii) a plurality of complexes of antigen-antibody complexes and proteins,

iii) a plurality of complexes of antigen-antibody complexes and protein complexes;

b) using the representations of the plurality of antigen-antibody-protein complexes to train a neural network to obtain a representation of an antigen-antibody-protein complex, given a representation of a constituent target complex of that antigen-antibody-protein complex:

ii) wherein the constituent target complex is an antigen or antigen-antibody-protein complex,

iii) wherein an antigen-antibody-protein complex is any of the following:

(1) a complex of an antigen and an antibody (i.e. an antigen-antibody complex),

(2) a complex of an antigen-antibody complex and a protein,

(3) a complex of an antigen-antibody complex and a protein complex,

iv) wherein the neural network is configured to proceed recursively such that:

(1) for each iteration of the recursion, the neural network is configured to generate and output a representation of a candidate protein, if any, in complex with the constituent target complex,

(2) for each iteration of the recursion, a representation of the complex of the generated candidate protein (the output of the iteration) and the constituent target complex (the input of the iteration) is passed back into the neural network as input for the next iteration of the recursion;

c) using the trained neural network to obtain a representation of a candidate antigen-antibody-protein complex, given a representation of a constituent target complex:

d) synthesizing a constituent protein of the generated antigen-antibody-protein complex representation.

12. The method of claim 11 , wherein the neural network is a transformer.

13. The method of claim 12 , wherein the transformer is multicapitate (has multiple heads) including a structure head which generates the structure and docking site representation, and a sequence head which generates the sequence of the output protein at each iteration.

14. A method, as in the method of claim 13 , for generating an output antigen-antibody-protein complex given a constituent target complex and for synthesizing a constituent protein of the antigen-antibody-protein complex, wherein the method is also for obtaining and synthesizing an effective antibody given a peptide antigen, the method further comprising:

a) receiving, at a processor, representations of a plurality of candidate antigens;

b) generating a candidate antigen-antibody-protein complex for each candidate antigen; wherein for each generated candidate antigen-antibody-protein complex, a constituent antibody is synthesized as output;

c) assessing the interaction efficacy and properties of the antigen and synthesized antibody of each of the plurality of candidate antigen-antibody-protein complexes;

d) selecting, based on the interaction efficacies and properties, the most effective candidate antibody of the plurality of represented candidate antibodies.

15. The method of claim 14 , wherein the antigen is an endogenous protein and the antibody is a synthetic biologic drug.

16. The method of claim 14 , wherein the antigen is a component or product of a pathogen and the antibody is a monoclonal antibody (mAb) drug.

17. The method of claim 14 , wherein the antigen is a component or product of a cancerous cell and the antibody is the antibody part of an Antibody Drug Conjugate (ADC).

18. A method, comprising:

a) receiving, at a processor, representations of a plurality of protein-protein complexes;

b) using the representations of the plurality of protein-protein complexes to train a neural network to obtain a representation of a protein-protein complex, given a representation of a constituent target complex of that protein-protein complex:

i) wherein the constituent target complex is a protein or protein-protein complex,

ii) wherein the neural network is configured to proceed recursively such that:

(1) for each iteration of the recursion, the neural network is configured to generate and output a representation of a candidate protein, if any, in complex with the constituent target complex,

(2) for each iteration of the recursion, a representation of the complex of the generated candidate protein (the output of the iteration) and the constituent target complex (the input of the iteration) is passed back into the neural network as input for the next iteration of the recursion;

c) receiving, at a processor, a plurality of representations of a target protein, wherein each representation includes a representation of the structure of the target protein when in complex with one of a plurality of small molecule drug (SMD) candidate ligands;

d) using the trained neural network to obtain a representation of a candidate protein-protein complex for each of the plurality of protein-SMD ligand complexes; wherein each candidate protein-protein complex output arises from using one of the plurality of protein-SMD complexes as input into the trained neural network;

e) assessing the properties and efficacy of each the plurality of candidate protein-protein complexes and their constituent proteins;

f) selecting, based on the assessment, the most effective small molecule drug;

g) manufacturing the small molecule drug ligand.

19. The method of claim 18 , wherein the neural network is a multicapitate transformer (a transformer with two or more heads), including a sequence head and a structure head, such that at each amino acid residue-generating iteration, the sequence head outputs the next residue in the sequence and the structure head outputs that residue's associated structure parameters.

20. The method of claim 18 , wherein the selected candidate protein is a receptor.