Method for predicting HLA-binding peptides using protein structural features
The present invention discloses a method for predicting peptides that are capable of binding to HLA molecules that incorporate the crystal structure of HLA molecules. An improved HLA-specific peptide docking workflow is used to simulate the occupancy of a peptide on the binding pocket of an HLA molecule, and three models are trained to predict the binding of the peptide to HLA molecules. The results show that these models predict HLA-allele specific binding peptides with extremely high accuracy.
1. A method of identifying one or more selected candidate peptides capable of binding a class I major histocompatibility complex (MHC) molecule of a single human leukocyte antigen (HLA) allele, the method comprising:
a. generating at least 100 models simulating occupancy for each of one or more candidate peptides on an HLA binding pocket, wherein the HLA binding pocket is in (i) a crystal structure of the MHC molecule of the single HLA allele or (ii) a crystal structure of a similar MHC molecule;
b. extracting structural features indicative of occupancy from the at least 100 models of step (a); and
c. providing the structural features extracted in step (b) to a machine learning algorithm, wherein the machine learning algorithm has been trained using a prior dataset comprising:
peptide sequence features of one or more binding peptides on the HLA binding pocket,
peptide sequence features of one or more non-binding peptides on the HLA binding pocket,
structural features of one or more binding peptides on the HLA binding pocket, and
structural features of one or more non-binding peptides on the HLA binding pocket,
whereby the machine learning algorithm outputs selected candidate peptides for binding the MHC molecule of the single HLA allele
thereby identifying one or more selected candidate peptides capable of binding the MHC molecule of the single HLA allele.
2. The method of claim 1 ,
wherein the machine learning algorithm is trained with one or more peptides bound to MHC molecules in cell lines mono-allelic for the HLA allele; and/or
wherein simulating occupancy is performed using one or more protein-peptide docking models and/or molecular dynamics simulation; and/or
wherein the machine learning algorithm comprises one or more peptide encoding schemes.
3. The method of claim 2 , wherein the one or more peptide encoding schemes comprise one-hot/dummy encoding, blosum62 encoding, or peptide: MHC binding energy covariance (PMBEC) encoding.
4. The method of claim 1 , wherein each selected candidate peptide is predicted to bind one or more MHC molecules of one or more HLA alleles; and/or
wherein the selected candidate peptides capable of binding the MHC molecule of the single HLA allele are 8, 9, 10, or 11 amino acids in length; and/or
wherein the selected candidate peptides capable of binding the MHC molecule of the single HLA allele are any combination of peptides selected from the group consisting of 8, 9, 10, and 11 amino acids in length.
5. The method of claim 1 , wherein each selected candidate peptide is predicted to bind one or more MHC molecules of one or more HLA alleles, and wherein the prediction is evaluated using metrics comprising area under a curve for a receiver operating characteristic curve and positive predictive value.
6. The method of claim 1 , wherein the candidate peptides are selected from a subject and the MHC molecule of the single HLA allele is expressed in the subject.
7. The method of claim 6 , wherein the subject is suffering from a disease or condition.
8. The method of claim 7 , wherein the disease or condition is selected from the group consisting of cancer, an infection, an autoimmune disease, and a transplant.
9. The method of claim 8 , wherein the peptides are identified by nucleic acid sequencing of a sample obtained from the subject, wherein the sample comprises tumor cells, infected cells, cells targeted by an autoimmune response, or cells to be transplanted.
10. The method of claim 8 , wherein the disease is cancer and the peptides are neoantigens and/or novel unannotated open reading frames (nuORFs).
11. The method of claim 1 , wherein the one or more binding peptides are selected from the group consisting of: A*01:01 (SEQ ID Nos: 44-1120); A*02:01 (SEQ ID Nos: 1121-4202); A*02:02 (SEQ ID Nos: 4203-7373); A*02:03 (SEQ ID Nos: 7374-9953); A*02:04 (SEQ ID Nos: 9954-11940); A*02:05 (SEQ ID Nos: 11941-14981); A*02:06 (SEQ ID Nos: 14982-17191); A*02:07 (SEQ ID Nos: 117192-20710); A*02:11 (SEQ ID Nos: 20711-22696); A*03:01 (SEQ ID Nos: 22697-24233); A*11:01 (SEQ ID Nos: 24234-27505); A*11:02 (SEQ ID Nos: 27506-29812); A*23:01 (SEQ ID Nos: 29813-32133); A*24:02 (SEQ ID Nos: 32134-34347); A*24:07 (SEQ ID Nos: 34348-35681); A*25:01 (SEQ ID Nos: 35682-36682); A*26:01 (SEQ ID Nos: 36683-37957); A*29:02 (SEQ ID Nos: 37958-38921); A*30:01 (SEQ ID Nos: 38922-40029); A*30:02 (SEQ ID Nos: 40030-42114); A*31:01 (SEQ ID Nos: 42115-42919); A*32:01 (SEQ ID Nos: 42920-44874); A*33:01 (SEQ ID Nos: 44875-46761); A*33:03 (SEQ ID Nos: 46762-49053); A*34:01 (SEQ ID Nos: 49054-50948); A*34:02 (SEQ ID Nos: 50949-53677); A*36:01 (SEQ ID Nos: 53678-55165); A*66:01 (SEQ ID Nos: 55166-56901); A*68:01 (SEQ ID Nos: 56902-58374); A*68:02 (SEQ ID Nos: 58375-59804); A*74:01 (SEQ ID Nos: 59805-61821); B*07:02 (SEQ ID Nos: 61822-63473); B*07:04 (SEQ ID Nos: 63474-64885); B*08:01 (SEQ ID Nos: 64886-65609); B*13:01 (SEQ ID Nos: 65610-69419); B*13:02 (SEQ ID Nos: 69420-71587); B*14:02 (SEQ ID Nos: 71588-72970); B*15:01 (SEQ ID Nos: 72971-76378); B*15:02 (SEQ ID Nos: 76379-77762); B*15:03 (SEQ ID Nos: 77763-80458); B*15:10 (SEQ ID Nos: 80459-81940); B*15:17 (SEQ ID Nos: 81941-83632); B*18:01 (SEQ ID Nos: 83633-85593); B*27.05 (SEQ ID Nos: 85594-87076); B*35:01 (SEQ ID Nos: 87077-87772); B*35:03 (SEQ ID Nos: 87773-89157); B*35:07 (SEQ ID Nos: 89158-90977); B*37:01 (SEQ ID Nos: 90978-92452); B*38:01 (SEQ ID Nos: 92453-94858); B*38:02 (SEQ ID Nos: 94859-97742); B*40:01 (SEQ ID Nos: 97743-100731); B*40:02 (SEQ ID Nos: 100732-104409); B*40:06 (SEQ ID Nos: 104410-106653); B*42:01 (SEQ ID Nos: 106612-019885); B*44:02 (SEQ ID Nos: 109886-110903); B*44:03 (SEQ ID Nos: 110904-111749); B*45:01 (SEQ ID Nos: 111750-113153); B*46:01 (SEQ ID Nos: 113154-114113); B*49:01 (SEQ ID Nos: 114114-117833); B*50:01 (SEQ ID Nos: 117834-118468); B*51:01 (SEQ ID Nos: 118469-119991); B*52:01 (SEQ ID Nos: 119992-121525); B*53:01 (SEQ ID Nos: 121526-123560); 54:01 (SEQ ID Nos: 123561-124684); B*55:01 (SEQ ID Nos: 124685-126136); B*55:02 (SEQ ID Nos: 126137-127557); B*56:01 (SEQ ID Nos: 127558-129239); B*57:01 (SEQ ID Nos: 129240-130274); B*57:03 (SEQ ID Nos: 130275-132636); B*58:01 (SEQ ID Nos: 132637-134577); B*58:02 (SEQ ID Nos: 134578-135530); C*01:02 (SEQ ID Nos: 135531-136878); C*02:02 (SEQ ID Nos: 136879-137802); C*03:02 (SEQ ID Nos: 137803-138984); C*03:03 (SEQ ID Nos: 138985-141074); C*03:04 (SEQ ID Nos: 141075-143394); C*04:01 (SEQ ID Nos: 143395-145236); C*04:03 (SEQ ID Nos: 145237-146269); C*05:01 (SEQ ID Nos: 146270-147708); C*06:02 (SEQ ID Nos: 147709-149028); C*07:01 (SEQ ID Nos: 149029-149822); C*07:02 (SEQ ID Nos: 149823-150900); C*07:04 (SEQ ID Nos: 150901-151615); C*08:01 (SEQ ID Nos: 151616-153388); C*08:02 (SEQ ID Nos: 153389-156499); C*12:02 (SEQ ID Nos: 156500-157889); C*12:03 (SEQ ID Nos: 157890-160043); C*14:02 (SEQ ID Nos: 160044-161408); C*14:03 (SEQ ID Nos: 161409-164186); C*15:02 (SEQ ID Nos: 164187-167475); C*16:01 (SEQ ID Nos: 167476-170317); C*17:01 (SEQ ID Nos: 170318-171281); G*01:01 (SEQ ID Nos: 171282-172073); G*01:03 (SEQ ID Nos: 172074-172742) and G*01:04 (SEQ ID Nos: 172743-173477).
12. The method of claim 1 , wherein the structural features are selected from the group consisting of energies of attraction, energies of repulsion, energies of solvation, energies of side chain and backbone hydrogen bonds, energies of side chain and backbone conformations, Lennard-Jones attractive potential between atoms in different residue, Lennard-Jones repulsive potential between atoms in different residues, Lazaridis-Karplus solvation energy, Lennard-Jones repulsive potential between atoms in the same residue, coulombic electrostatic potential with a distance-dependent dielectric, proline ring closure energy and energy of psi angle of preceding residue, backbone-backbone hbonds close in primary sequence, backbone-backbone hbonds distant in primary sequences, sidechain-backbone hydrogen bond energy, sidechain-sidechain hydrogen bond energy, disulfide geometry potential, Ramachandran preferences, omega dihedral in the backbone, internal energy of sidechain rotamers as derived from Dunbrack's statistics, probability of amino acid at phi/psi, and reference energy for each amino acid.
13. The method of claim 1 , wherein the prior dataset used to train the machine learning algorithm further comprises data selected from the group consisting of:
amino acid physical properties, peptide physical properties, expression level of a source protein of each peptide, protein stability, protein translation rate, protein degradation rate, translational efficiencies from ribosomal profiling, protein cleavability, protein localization, motifs of host proteins that facilitate TAP transport, whether host protein is subject to autophagy, motifs that favor ribosomal stalling, protein features that favor nonsense-mediated mRNA decay (NMD), peptide cleavability, dummy peptide encoding, PCA peptide encoding, and Kidera peptide-level features.
14. The method of claim 1 , wherein the machine learning algorithm comprises a neural network model.
15. The method of claim 14 , wherein the neural network model comprises a single hidden layer neural network model and/or a convolutional neural network model.
16. A method of identifying one or more selected candidate peptides capable of binding a class I major histocompatibility complex (MHC) molecule of a single human leukocyte antigen (HLA) allele, the method comprising:
a. generating at least 100 models simulating occupancy for each of one or more candidate peptides on an HLA binding pocket, wherein the HLA binding pocket is in (i) a crystal structure of the MHC molecule of the single HLA allele or (ii) a crystal structure of a similar MHC molecule;
b. extracting structural features indicative of occupancy from the at least 100 models of step (a); and
c. providing the structural features extracted in step (b) to a machine learning algorithm, wherein the machine learning algorithm has been trained using a prior dataset comprising (i) peptide sequence features of one or more binding peptides on the HLA binding pocket, peptide sequence features of one or more non-binding peptides on the HLA binding pocket, structural features of one or more binding peptides on the HLA binding pocket, and structural features of one or more non-binding peptides on the HLA binding pocket and (ii) non-structural features of the one or more binding peptides and non-structural features of the one or more non-binding peptides,
whereby the machine learning algorithm outputs selected candidate peptides for binding the MHC molecule of the single HLA allele,
thereby identifying one or more selected candidate peptides capable of binding the MHC molecule of the single HLA allele.
17. The method of claim 16 , wherein the prior dataset used to train the machine learning algorithm further comprises data selected from the group consisting of: amino acid physical properties, peptide physical properties, expression level of the source protein of a peptide, protein stability, protein translation rate, protein degradation rate, translational efficiencies from ribosomal profiling, protein cleavability, protein localization, motifs of host proteins that facilitate TAP transport, whether host protein is subject to autophagy, motifs that favor ribosomal stalling, protein features that favor nonsense-mediated mRNA decay (NMD), peptide cleavability, dummy peptide encoding, PCA peptide encoding, and Kidera peptide-level features.
18. A method of identifying one or more selected candidate peptides capable of binding a class I major histocompatibility complex (MHC) molecule of a single human leukocyte antigen (HLA) allele, the method comprising:
a. generating at least 100 models simulating occupancy for each of one or more candidate peptides on an HLA binding pocket, wherein the HLA binding pocket is in (i) a crystal structure of the MHC molecule of the single HLA allele or (ii) a crystal structure of a similar MHC molecule, wherein the output models are generated by a peptide docking tool and comprise one or more protein-peptide docking models;
b. extracting structural features indicative of occupancy from the at least 100 models of step (a); and
c. providing the structural features extracted in step (b) to a machine learning algorithm, wherein the machine learning algorithm has been trained using a prior dataset comprising (i) peptide sequence features of one or more binding peptides on the HLA binding pocket, peptide sequence features of one or more non-binding peptides on the HLA binding pocket, structural features of one or more binding peptides on the HLA binding pocket, and structural features of one or more non-binding peptides on the HLA binding pocket and (ii) non-structural features of the one or more binding peptides and non-structural features of the one or more non-binding peptides,
whereby the machine learning algorithm outputs selected candidate peptides for binding the MHC molecule of the single HLA allele,
thereby identifying one or more selected candidate peptides capable of binding the MHC molecule of the single HLA allele.
19. The method of claim 18 , herein the peptide docking tool is selected from the group consisting of FlexPepDock, DockTope, pDOCK and HADDOCK.