Computer implemented method

Abstract

A computer implemented method for identifying amino acid sequences predicted to form a scaffold-bound peptide ligand that binds to a target, the method comprising: generating an amino acid sequence describing a peptide, the sequence satisfying one or more predefined sequence constraints such that the peptide binds to one of one or more predefined scaffolds in at least two locations, to form a scaffold-bound peptide ligand; determining at least one predicted interaction characteristic relating to an interaction between the peptide and one or more predefined targets; determining a favourability of the at least one predicted interaction characteristic; providing output data identifying one or more amino acid sequences predicted to form a scaffold-bound peptide ligand that binds to a target, based on the determined favourability.

Description

[0001] COMPUTER IMPLEMENTED METHOD

[0002] TECHNICAL FIELD

[0003] The present invention relates to computer implemented methods of identifying amino acid sequences likely to form scaffold-bound peptide ligands that bind to a target. Such methods may be used for identifying novel therapeutics in the form of scaffold-bound peptide ligands.

[0004] BACKGROUND ART

[0005] Cyclic peptides are able to bind with high affinity and target specificity to protein targets and hence are an attractive molecule class for the development of therapeutics. In fact, several cyclic peptides are already successfully used in the clinic, as for example the antibacterial peptide vancomycin, the immunosuppressant drug cyclosporine or the anticancer drug octreotide (Driggers et al. (2008), Nat Rev Drug Discov 7 (7), 608-24). Good binding properties result from a relatively large interaction surface formed between the peptide and the target as well as the reduced conformational flexibility of the cyclic structures. Typically, macrocycles bind to surfaces of several hundred square angstrom, as for example the cyclic peptide CXCR4 antagonist CVX15 (400 A2; Wu et al. (2007), Science 330, 1066-71), a cyclic peptide with the Arg-Gly-Asp motif binding to integrin aVb3 (355 A2) (Xiong et al. (2002), Science 296 (5565), 151-5) or the cyclic peptide inhibitor upain-1 binding to urokinase-type plasminogen activator (603 A2; Zhao et al. (2007), J Struct Biol 160 (1), 1-10).

[0006] Due to their cyclic configuration, peptide macrocycles are less flexible than linear peptides, leading to a smaller loss of entropy upon binding to targets and resulting in a higher binding affinity. The reduced flexibility also leads to locking target-specific conformations, increasing binding specificity compared to linear peptides. This effect has been exemplified by a potent and selective inhibitor of matrix metalloproteinase 8, MMP-8) which lost its selectivity over other MMPs when its ring was opened (Cherney et al. (1998), J Med Chem 41 (11), 1749-51). The favourable binding properties achieved through macrocyclization are even more pronounced in multicyclic peptides having more than one peptide ring as for example in vancomycin, nisin and actinomycin. Different research teams have previously tethered polypeptides with cysteine residues to a synthetic molecular structure (Kemp and McNamara (1985), J. Org. Chem; Timmerman et al. (2005), ChemBioChem). Meloen and co-workers had used tris(bromomethyl)benzene and related molecules for rapid and quantitative cyclization of multiple peptide loops onto synthetic scaffolds for structural mimicry of protein surfaces (Timmerman et al. (2005), ChemBioChem). Methods for the generation of candidate drug compounds wherein said compounds are generated by linking cysteine containing polypeptides to a molecular scaffold as for example tris(bromomethyl)benzene are disclosed in WO 2004 / 077062 and WO 2006 / 078161.

[0007] Phage display-based combinatorial approaches have been developed to generate and screen large libraries of bicyclic peptides to targets of interest (Heinis et al. (2009), Nat Chem Biol 5 (7), 502-7 and W02009 / 098450). Briefly, combinatorial libraries of linear peptides containing three cysteine residues and two regions of six random amino acids (Cys-(Xaa)6- Cys-(Xaa)6-Cys) were displayed on phage and cyclised by covalently linking the cysteine side chains to a small molecule (tris-(bromomethyl)benzene).

[0008] However, phage display-based screening is time and resource intensive. Therefore, there is a need for a computational approach to identifying peptides that bind to a scaffold and have favourable characteristics.

[0009] SUMMARY OF THE INVENTION

[0010] According to an aspect of the disclosure there is provided A computer implemented method for identifying amino acid sequences predicted to form a scaffold-bound peptide ligand that binds to a target, the method comprising: generating an amino acid sequence describing a peptide, the sequence satisfying one or more predefined sequence constraints such that the peptide binds to one of one or more predefined scaffolds in at least two locations, to form a scaffold-bound peptide ligand; determining at least one predicted interaction characteristic relating to an interaction between the peptide and one or more predefined targets; determining a favourability of the at least one predicted interaction characteristic; providing output data identifying one or more amino acid sequences predicted to form a scaffold-bound peptide ligand that binds to a target, based on the determined favourability.

[0011] Optionally, the step of generating an amino acid sequence comprises modification of a previous amino acid sequence. Optionally, the modification comprises adding at least one residue, removing at least one residue, or changing at least one residue of a previous amino acid sequence. Optionally, the modification is made at a random position in the previous sequence.

[0012] Optionally, the modification comprises the insertion of an amino acid selected based on a probability distribution. Optionally, the probability distribution ensures that at least one of the one or more sequence constraints are satisfied. Optionally, the probability distribution is generated by: generating an amino acid sequence describing a peptide; determining at least one predicted interaction characteristic relating to an interaction between the peptide and one or more predefined targets; determining a favourability of the at least one predicted interaction characteristic; generating a probability distribution for all possible amino acid sequences and all possible positions, based on the generated amino acid sequences and the determined favourability.

[0013] Optionally, the step of generating an amino acid sequence, and subsequent steps, are repeated for a plurality of amino acid sequences.

[0014] Optionally, the output data comprises the amino acid sequence.

[0015] Optionally, the one or more sequence constraints comprises one or more constraints that increase a likelihood that the peptide will bind to one of one or more predefined scaffolds.

[0016] Optionally, the one or more sequence constraints comprises one or more constraints that increase a likelihood that the peptide will form at least one loop structure when bound with one of the one or more predefined scaffolds.

[0017] Optionally, the one or more sequence constraints comprises a constraint on a total length of the sequence. Optionally, the one or more sequence constraints comprises a constraint that the sequence consists of from 6 to 30 amino acid residues. Optionally, the one or more sequence constraints comprises a constraint that the sequence comprises at least two cysteine residues.

[0018] Optionally, the one or more sequence constraints comprises a constraint that the sequence comprises three cysteine residues.

[0019] Optionally, the one or more sequence constraints comprises a constraint on the relative positions of at least two cysteine residues.

[0020] Optionally, the one or more sequence constraints comprises a constraint that at least one pair of successive cysteine residues are separated by no more than a first predefined number of amino acid residues.

[0021] Optionally, the one or more sequence constraints comprises a constraint that at least one pair of successive cysteine residues are separated by no fewer than a second predefined number of amino acid residues.

[0022] Optionally, the sequence comprises three cysteine residues, including two pairs of successive cysteine residues, and both pairs satisfy at least one of the constraints defined above regarding the pairs of cysteine residues.

[0023] Optionally, the one or more constraints comprises a constraint on the amino acid residue at one or more terminal positions in the sequence.

[0024] Optionally, the one or more constraints comprises a constraint that the amino acid residue at one or more terminal positions in the sequence is an alanine residue.

[0025] Optionally, the one or more interaction characteristics comprises a binding score based on the likelihood of the peptide binding to the target.

[0026] Optionally, the one or more interaction characteristics comprises a folding score based on the likelihood of the peptide being well-folded Optionally, the one or more interaction characteristics comprises a score for one or more residues of the target, based on the likelihood of each of these residue binding to the peptide.

[0027] Optionally, the one or more interaction characteristics comprises a score for one or more residues of the peptide, based on the likelihood of each of these residues binding to the target.

[0028] Optionally, the one or more interaction characteristics is determined based on a predicted three-dimensional structure of the peptide and / or the target.

[0029] Optionally the method further comprises a step of determining one or more structural characteristics of the peptide and a step of determining a favourability of the one or more structural characteristics.

[0030] Optionally, the one or more structural characteristics relate to a likelihood that the peptide will bind with one of one or more predefined molecular scaffolds.

[0031] Optionally, the one or more structural characteristics comprises a distance between two cysteine residues, in the predicted three-dimensional structure of the peptide.

[0032] Optionally, the one or more structural characteristics comprises a distance between respective sulphur atoms of two cysteine residues in the predicted three-dimensional structure of the peptide.

[0033] Optionally, the one or more structural characteristics comprises an angle formed between a first cysteine residue and two further cysteine residues, in the predicted three-dimensional structure of the peptide.

[0034] Optionally, the one or more structural characteristics comprises a dihedral angle formed between a first cysteine residue and a second cystine residue, in the predicted three- dimensional structure of the peptide. Optionally, the predicted three-dimensional structure of the peptide is generated by a machine learning algorithm based on input data comprising the generated amino acid sequence.

[0035] Optionally, the method further comprises a step of determining one or more functional characteristics of the peptide and a step of determining a favourability of the one or more functional characteristics. Optionally, the one or more functional characteristics comprises one more of: membrane permeability of peptide and blood-brain barrier permeability of the peptide.

[0036] Optionally, the favourability is determined based on a comparison between the determined characteristics and predefined corresponding desired or optimal parameters. Optionally, the favourability is determined by calculating a favourability score. Optionally, the favourability score is calculated using an objective function comprising terms corresponding to each of the determined characteristics.

[0037] Optionally, the scaffold-bound peptide ligand is bicyclic, comprising two loop sequences, and is covalently bound to the scaffold at three locations.

[0038] Optionally, the one or more scaffolds comprise one or more of TATA, TATB, TCTZ, TBMB, TTZ, TBAB, TSTA, TCAZ, TCAN, and TCCU.

[0039] Optionally, the peptide forms a peptide ligand, and the peptide in the peptide ligand comprises at least three cysteine residues, separated by at least two loop sequences, and the scaffold forms covalent bonds with the cysteine residues of the peptide such that at least two peptide loops are formed on the scaffold.

[0040] Optionally, the target is a protein, a protein on a cell, a tumour antigen, a viral antigen, bacterial antigen.

[0041] According to a second aspect of the disclosure there is provided a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the first aspect. According to a third aspect of the disclosure there is provided a data processing system comprising means for carrying out the method of the first aspect.

[0042] According to a fourth aspect of the disclosure there is provided a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the method the first aspect.

[0043] According to a fifth aspect of the disclosure there is provided a method of identifying a therapeutic agent or diagnostic agent targeting a target, the therapeutic comprising a scaffold-bound peptide ligand, the method comprising using the computer implemented method of the first aspect to identify a peptide forming the scaffold-bound peptide ligand.

[0044] According to a sixth aspect of the disclosure there is provided a method of making a peptide or scaffold-bound peptide ligand, comprising using the computer implemented method of the first aspect to identify a peptide forming the scaffold-bound peptide ligand or the scaffold-bound peptide ligand.

[0045] According to a seventh aspect of the disclosure there is provided a scaffold-bound peptide ligand made by the method of the sixth aspect.

[0046] BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Further features of the disclosure will be described below, by way of non-limiting examples and with reference to the accompanying drawings, in which:

[0048] Fig. 1 shows an example method or identifying amino acid sequences;

[0049] Fig. 2 shows an example method of generating a probability matrix for generating amino acid sequences satisfying sequence constraints;

[0050] Fig. 3 shows an example probability matrix;

[0051] Fig. 4 shows an example method of generating additional amino acid sequences;

[0052] Fig. 5 shows an example scaffold-bound peptide ligand; and

[0053] Fig. 6 shows A) a peptide identified by the method, and B) binding affinities for 100 peptides identified by the method.

[0054] DETAILED DESCRIPTION Peptide Ligands

[0055] Fig. 5 schematically shows an example scaffold-bound peptide ligand (also referred to herein as a peptide ligand) according the disclosure. The scaffold-bound peptide ligand comprises a peptide bound to a scaffold molecule.

[0056] A peptide ligand, as referred to herein, refers to a peptide, peptidic or peptidomimetic covalently bound to a molecular scaffold. Typically, such peptides, peptidics or peptidomimetics comprise a peptide having natural or non-natural amino acids, two or more reactive groups (i.e. cysteine residues) which are capable of forming covalent bonds to the scaffold, and a sequence subtended between said reactive groups which is referred to as the loop sequence, since it may form a loop when the peptide, peptidic or peptidomimetic is bound to the scaffold. The peptides, peptidics or peptidomimetics may comprise at least three cysteine residues, and form at least two loops on the scaffold, thus forming bicyclic peptides.

[0057] As shown in Fig. 5, the scaffold may be a trivalent molecule covalently bound, via bonds Li - L3, to cysteine reactive groups, and comprising loop sequences, Loop A and Loop B, subtended between said reactive groups.

[0058] Certain bicyclic peptides have a number of advantageous properties which enable them to be considered as suitable drug-like molecules for injection, inhalation, nasal, ocular, oral or topical administration. Such advantageous properties include:

[0059] - Species cross-reactivity. This is a typical requirement for preclinical pharmacodynamics and pharmacokinetic evaluation;

[0060] - Protease stability. Bicyclic peptide ligands should in most circumstances demonstrate stability to plasma proteases, epithelial ("membrane-anchored") proteases, gastric and intestinal proteases, lung surface proteases, intracellular proteases and the like. Protease stability should be maintained between different species such that a bicyclic peptide lead candidate can be developed in animal models as well as administered with confidence to humans; - Desirable solubility profile. This is a function of the proportion of charged and hydrophilic versus hydrophobic residues and intra / inter-molecular H-bonding, which is important for formulation and absorption purposes;

[0061] - An optimal plasma half-life in the circulation. Depending upon the clinical indication and treatment regimen, it may be required to develop a bicyclic peptide with short or prolonged in vivo exposure times for the management of either chronic or acute disease states. The optimal exposure time will be governed by the requirement for sustained exposure (for maximal therapeutic efficiency) versus the requirement for short exposure times to minimise toxicological effects arising from sustained exposure to the agent.

[0062] Molecular Scaffold

[0063] A scaffold molecule may be any molecule configured to bind covalently to and orient a peptide. A scaffold molecule may be a non-aromatic molecular scaffold. The term “nonaromatic molecular scaffold” may refer to any molecular scaffold which does not contain an aromatic (i.e. unsaturated) carbocyclic or heterocyclic ring system.

[0064] Suitable examples of non-aromatic molecular scaffolds are described in Heinis et al (2014) Angewandte Chemie, International Edition 53(6) 1602-1606.

[0065] As noted in the foregoing documents, the molecular scaffold may be a small molecule, such as a small organic molecule. The molecular scaffold may comprise reactive groups that are capable of reacting with functional group(s) of the peptide to form covalent bonds.

[0066] The molecular scaffold may comprise chemical groups which form the linkage with a peptide, such as amines, thiols, alcohols, ketones, aldehydes, nitriles, carboxylic acids, esters, alkenes, alkynes, azides, anhydrides, succinimides, maleimides, alkyl halides and acyl halides. One example of a suitable scaffold is l,3,5-Triacryloylhexahydro-l,3,5-triazine (TATA) (Angewandte Chemie, International Edition (2014), 53(6), 1602-1606). Another example of a suitable scaffold is l,3,5-tris(bromomethyl)benzene (TBMB) (W02016 / 067035 Al).

[0067] Other example scaffolds include TATB (l,T,l"-(l,3,5-triazinane-l,3,5-triyl)tris(2- bromoethan-l-one)), TCTZ (1,3,5-trichloromethyltriazine) (Kale, S.S., Villequey, C., Kong, XD. et al. Cyclization of peptides with two chemical bridges affords large scaffold diversities. Nature Chem 10, 715-723 TTZ (2,4,6-Tris(Halomethyl)-l,3,5-Triazine), TBAB (2,4,6-Trioxo-l,3,5-triazinane-l,3,5- triyl)-trisethane-2,l-diyltriacrylate), TSTA (l,3,5-tri(ethenesulfonyl)-l,3,5-triazinane) (W02020 / 084305), TCAZ (l,4,7-tris(chloroacetyl)octahydro-l,4,7-triazonane), TCAN (N,N',N"-(nitrilotris(ethane-2, 1 -diyl))tris(2 -chloroacetamide)), TCCU ( 1 , T, 1 "-( 1 H,4H- 3a,6a-(methanoiminomethano)pyrrolo[3,4-c]pyrrole-2,5,8(3H,6H)-triyl)tris(2-chloroethan- 1-one)) (WO2018 / 197893).

[0068] Example Computer Implemented Method

[0069] The methods described herein may be executed by a computer, or other data processing system. The methods may be embodied by a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any one of the preceding claims. The methods may also be embodied by a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the methods. The methods may also be embodied by a data processing system comprising means for carrying out the method.

[0070] Fig. 1 is a flow chart illustrating a first example computer implemented method according the disclosure, for identifying amino acid sequences likely to form scaffold-bound peptide ligands that bind to a target.

[0071] In a first step, SI, an amino acid sequence describing a peptide is generated. The generated sequence satisfies one or more predefined sequence constraints such that the peptide has a propensity to form a scaffold-bound peptide ligand when bound to one of one or more predefined scaffold molecules. In a second step, S2, at least one predicted interaction characteristic of the peptide is determined. The interaction characteristic relates to an interaction between the peptide and one or more predefined targets.

[0072] In a third step, S3, a favourability of the at least one predicted interaction characteristic of the peptide is determined.

[0073] In a fourth step, S4, output data is provided by the method based on the favourability determined in step S3. The output data identifies one or more amino acid sequences likely to form a scaffold-bound peptide ligand with favourable characteristics.

[0074] As shown, in Fig. 1, the method may be configured to repeat, e.g. iterate through, the first to third, steps SI, S2, S3. In some examples, step S4 may be performed a plurality of times, e.g. with each iteration or only iterations for which the determined favourability is relatively high (e.g. above a predefined threshold). Accordingly, output data may be provided for a plurality of amino acid sequences. In some examples, step S4 may be performed once at the end of a predefined number of iterations.

[0075] Interaction Characteristics

[0076] Favourability of the at least one predicted interaction characteristic may be determined based on a comparison between the at least one predicted interaction characteristic and one or more predefined corresponding interaction requirements. The predefined corresponding interaction requirements may generally be desired (or optimal) parameters for the interaction characteristics.

[0077] In a preferred example, favourability may be determined by calculating a favourability score. The calculation may be made based on an objective function, such as a loss function. The objective function may comprise terms that mathematically compare the at least one predicted interaction characteristic and the one or more predefined corresponding interaction requirements. In a preferred example, a combined favourability for a plurality of predicted interaction characteristic is determined using a single objective function that calculates a single favourability score for each peptide, based on a plurality of predicted interaction characteristics.

[0078] In other examples, comparison with a predefined interaction requirement may be performed separately for each interaction characteristic and these may optionally be combined to provide a single favourability score.

[0079] The method may iterate through steps SI to S3 to optimise the objective function, e.g. minimise a loss function. In an example method, the method may use multivariate gradient descent, for example. In an example method, a predefined number or loops may be executed before exiting.

[0080] In a preferred example, a predicted interaction characteristic may relate to the likelihood of the peptide binding to a target. For example, one predicted interaction characteristic may be a binding score, based on the likelihood of the peptide binding to a target. Favourability may be determined based on how the binding score compares to a desired (or optimal) binding score.

[0081] In a preferred example, a predicted interaction characteristic may relate to the stability of the peptide. For example, one predicted interaction characteristic may be a folding score based on the likelihood of the peptide being well-folded. Favourability may be determined based on how the folding score compares to a desired (or optimal) folding score.

[0082] Another example interaction characteristic may relate to the parts of the target at which the peptide is predicted to bind to the target. For example, a predicted interaction characteristic may comprise a score for specific residues of the target (or all residues), based on the likelihood of each of these residue binding to the peptide. Favourability may be determined based on the scores for these specific residues (or specific residues out of all the residues). The specific residues may correspond to residues located at a part of the target preferred for binding to the peptide, with high favourability corresponding to a high likelihood of binding at these residues. It may be advantageous for the peptide to bind at a specific location on the target. Another example interaction characteristic may relate to the likelihood of one or more specific residues of the peptide binding to the target. For example, a predicted interaction characteristic may comprise a score for specific residues of the peptide (or all residues), based on the likelihood of each of these residues binding to the target. Favourability may be determined based on the scores for these specific residues (or specific residues out of all the residues). The specific residues may correspond to residues, such as lysine residues, with high favourability corresponding to a low likelihood of binding at these residues. Lysine residues may provide cross-linking within the peptide, so it may be advantageous for these residues not to interact with the target in order to provide cross-linking.

[0083] At least some predicted interaction characteristics, including those described above, may be determined based on a predicted three-dimensional structure of the peptide and / or the target. The three-dimensional structure of the peptide may be determined by a machine learning algorithm, based on input data comprising the generated amino acid sequence. The three-dimensional structure of the target may be determined by a machine learning algorithm, based on input data comprising the amino acid sequence of a target protein. In a preferred example, the machine learning algorithm, in either case, may be the AlphaFold™ algorithm. In an example method, input data may be provided comprising data describing a predefined target. This data may comprise an amino acid sequence of the target and / or the three-dimensional structure of the target.

[0084] In a preferred example in which predicted interaction characteristics are determined using the AlphaFold™ algorithm, the predicted interaction characteristics may comprise a Predicted Aligned Error (PAE) score, such as an i_PAE score. A PAE score is an AlphaFold™ metric to assess the confidence in the relative positions of different parts of the three-dimensional models e.g., different domains. A specific example method uses the metric as a measure of confidence in the relative positioning of the target and peptide by extracting the interface PAE (i_PAE) confidence score, which relates to the relative positioning of the peptide to the target at their binding interface. For example, if a peptide has a central helix that interacts with the target while the flanking regions do not, then the i_PAE score would correspond to the confidence in the relative positioning of the central helix and the region of the target close to it, the non-interacting parts are not including in the final i_PAE score. It may be advantageous to maximize this score as it gives preference to an accurate model of the binding interface over the flanking non-binding regions. The i_PAE score relates to the likelihood of the peptide binding to the target.

[0085] In a preferred example in which predicted interaction characteristics are determined using the AlphaFold™ algorithm, the predicted interaction characteristics may comprise a pLDDT score. The pLDDT may relate to the likelihood of the peptide forming a well- folded structure. The pLDDT score is an AlphaFold™ per-residue estimate confidence score (0-100) corresponding to the IDDT-Ca metric. It is a superimposition-free measure to assess the local quality of the model based on the agreement between the model and the available PDB data. 1DDT, Mariani T et al, 2013, (https: / / www.ncbi.nlm.nih.gov / pmc / articles / PMC3799472 / ) measures how well the environment in a reference structure is reproduced in a protein model. AlphaFold’s pLDDT uses the alpha-carbon atoms 1DDT score for the pLDDT score.

[0086] In a preferred example in which predicted interaction characteristics are determined using the AlphaFold™ algorithm, the predicted interaction characteristics may comprise a distogram. A distogram is a matrix of distances between all the CP atoms of the target and binder, and represents a contact map of the model. The distogram may be used to determined predicted interaction characteristics relating to where the peptide binds or does not bind to the target. For example, a relatively small distance between a target residue and a peptide residue may correspond with a relatively high likelihood of binding, and vice versa.

[0087] In an example method, input data may be provided comprising data describing the one or more interaction requirements. For example, these may be in the form of parameters for terms of the objective function described above. Alternatively, these may be provided in the form of the objective function itself, or any combination of a structure of the objective function and parameters of the objective function.

[0088] Sequence Generation

[0089] The step SI of generating an amino acid sequence, when this is repeated, may comprise modification of a previous amino acid sequence. In a preferred example, the modification may comprise changing (mutating) at least one residue in the sequence. In some examples, the modification may alternatively, or additionally, comprise adding at least one residue or removing at least one residue. An initial amino acid sequence may be provided as input data or may be generated as a random sequence.

[0090] Modification of the amino acid sequence may be performed at a randomly, pseudo- randomly, or probabilistically selected position in the sequence. Modification may comprise inserting (e.g. by changing or adding) a randomly, pseudo-randomly or probabilistically selected residue. In a preferred example, the modification may be made at a random position in the sequence and the modification may change the residue at the randomly selected position to a probabilistically selected residue.

[0091] The one or more sequence constraints may comprise one or more constraints that increase a likelihood that the peptide will bind to one of one or more predefined scaffolds. For example, a sequence constraint may require that the sequence comprises at least one cysteine residue. A cysteine residue may bind with a scaffold molecule, for example.

[0092] Alternatively, or additionally, the sequence constraints may comprise a constraint on a total length of the sequence. For example, the total length may be constrained to be in a range of from 6 to 30 amino acid residues, e.g. 13 residues long.

[0093] In a preferred example, the one or more sequence constraints may comprise one or more constraints that increase a likelihood that the peptide will bind to one of the one or more predefined scaffolds at at least two locations. For example, a sequence constraint may require that the sequence comprises at least two cysteine residues. In some preferred examples, a sequence constraint may require that the sequence comprises three cysteine residues, for example to form scaffold-bound peptide ligands having a bicyclic structure.

[0094] Generation of a sequence that satisfies one or more predefined sequence constraints may be achieved in a number of ways. In a preferred example, the method may be configured to only generate sequences that satisfy the sequence constraints, e.g. sequences having a predefined length and / or having cysteine residues at predefined locations. Alternatively, sequences that do not satisfy the sequence constraints may be generated, but then discarded, so as not progress to step S2. In one example method, modification of the sequence may only occur at predefined positions that are not the positions of one or more required cysteine residues. In another example method, all residues may be modified, but the positions of one or more required cysteine residues may have a relatively high probability that the cysteine residues remain at those positions.

[0095] Probability Distribution

[0096] In a preferred example, sequences may be generated probabilistically based on a predefined probability distribution of residues being located at specific positions. The probability distribution may define a probability for each possible amino acid residue at each possible position in the sequence. For sequence constraints requiring cysteine residues at specific positions, the probability of a cysteine residue at these positions may be 1. The probability distribution may be a position-specific scoring matrix (PSSM), for example. The PSSM may have rows corresponding to each of the possible amino acid residues and columns corresponding to each of the possible positions in the sequence. Modifications (mutations) may be made to the sequence weighted based on the probability distribution.

[0097] In a specific example method, the step S 1 of generating an amino acid sequence may comprise a first step of generating a probability distribution of residues being located at specific positions and a second step of generating an amino acid sequence based on the probability distribution.

[0098] Fig. 2 shows a preferred example method of generating a probability distribution. A first step S5 may comprise generating an amino acid sequence. The method may not constrain the residues that may be located at specific locations. In other words, all possible positions may have an equal probability of being modified to all residues. However, the method may constrain the length of the sequence to correspond with any length required by the general predefined sequence constraints.

[0099] A second step S6 may comprise determining at least one predicted interaction characteristic relating to an interaction between the peptide and the one or more predefined targets. The predicted interaction characteristic may be the same as those described above in relation to the method shown in Fig. 1. A third step S7 may comprise determining a favourability of the at least one predicted interaction characteristic, e.g. in the same way as described above in relation to the method shown in Fig. 1. As shows, the method may iterate through steps SI to S3, e.g. in the same way as described above in relation to the method shown in Fig. 1.

[0100] A fourth step S8 may comprise generating a probability matrix based on each sequence generated and a corresponding favourability, e.g. favourability score. Sequences with a higher score may have a higher weighting for generating the probability distribution.

[0101] Accordingly, if sequences with a relatively high favourability score have a relatively high probability of having a particular residue at a particular location, then the probability of that particular residue at that particular location will be relatively high in the probability distribution. Similarly, if sequences with a relatively high favourability score have a relatively low probability of having a particular residue at a particular location, then the probability of that particular residue at that particular location will be relatively low in the probability distribution. Conversely, sequences with a relatively low favourability score may have relatively little influence on the probability distribution.

[0102] In a specific example, the first probability distribution may be generated using the ColabDesign algorithm (https: / / github.com / sokrypton / ColabDesign). The ColabDesign algorithm generates random sequences of a predefined length, uses the AlphaFold algorithm to determine interaction characteristics and characteristics of the peptide, calculates a loss function and iterates to optimise the loss function, in order to generate a PSSM. A modified version of the ColabDesign algorithm may be used to perform the steps of the method of Fig. 1.

[0103] A fifth step S9 may comprise generating a second probability matrix by modifying the first probability matrix. For example, this may comprise modifying the probability of at least one specific residue at at least one specific location to have a predetermined value. The predetermined value may be a relatively high value, such as 1 , or a relatively low value such as 0. The predetermined probabilities may be used to fix the specific residues at specific positions. For example, the positions of one or more cysteine residues may be fixed. Fig. 3 shows an example probability matrix. As shown, the probability of cysteine resdiues at three specific positions is set to 1 , and the probability of cysteine residues at alternative positions and the probability of alternative residues at the three specific positions are set to 0.

[0104] In an alternative example method, Steps S8 and S9 may be replaced by a single step of generating a probability matrix in which the probabilities for some combinations of positions and residue are based on the generated sequences and associated score and the probabilities for other combinations of positions and residue are predetermined.

[0105] In an example method, the one or more sequence constraints may comprise a constraint on the amino acid residues at one or more terminal positions in the sequence. For example, the one or more constraints comprises a constraint that the amino acid residue at one or more terminal positions in the sequence is an alanine residue. In a preferred example, alanine residues may be fixed at both terminal positions.

[0106] The above described methods may be used, not only to fix the positions of cysteine residues and terminal alanine residues, but any other residues that may be required at specific positions.

[0107] In an example method, input data may be provided describing the one or more sequence constraints. For example, data describing one or more predetermined residues at predetermined positions may be provided as input data. For example, this data may be provided in the form of a matrix, e.g. identifying specific residues and their fixed specific positions. A first probability distribution may be modified, as described above, based on this input data.

[0108] The one or more sequence constraints may be based on the scaffold molecule or molecules to which the generated peptides are configured to bind. For example, the scaffold molecule may determine the specific residues required and / or their specific positions. For example, the scaffold molecule may determine the required number of cystine residues. In some cases, the scaffold molecule may determine the required distance between cystine residues. Accordingly, in some examples, the one or more sequence constraints may comprise a constraint on the relative positions of at least two cysteine residues. For example, the one or more sequence constraints may comprise a constraint that at least one pair of successive cysteine residues are separated by no more than a first predefined number of amino acid residues and / or no fewer than a second predefined number of amino acid residues. When the sequence comprises three cysteine residues, including two pairs of successive cysteine residues, both pairs may satisfy at least one of the constraints above. Accordingly, in some examples, predetermined positions of one or more cysteine residues may satisfy one or more of the above constraints.

[0109] In some cases, the specific positions of one or more required specific residues may be variable, but otherwise constrained by their relative positions. In such cases, a plurality of different probability distributions may be generated and sequences generated using each of the probability distributions, as described above. The different probability distributions may be generated based on different sequence constraints. For example, the different probability distributions may fix cysteine residues (or other residues) at different positions in the sequence.

[0110] In an example, the method may generate sequences based on each of the probability distributions in turn. For example, the method may iterate through the steps of the method, SI to S4, for each different probability distribution.

[0111] Structural characteristics

[0112] In some examples, in addition to determining one or more interaction characteristics, the method may further comprise determining one or more additional characteristics of the generated peptide. Further, favourability of any additional characteristics may be determined and the output data based on the favourability of the additional characteristics.

[0113] The favourability of additional characteristics may be performed in the same way as described above in relation to the interaction characteristics. For example, favourability may be determined based on a comparison between the additional characteristic and one or more predefined corresponding additional requirements, as described above; the predefined corresponding additional requirements may generally be desired (or optimal) parameters for the additional characteristics, as described above; and, favourability may be determined by calculating a favourability score, as described above.

[0114] The determination of additional characteristics may be performed as part of step S2 described above, for example. The determination of favourability of additional characteristics may be performed as part of step S3 described above, for example.

[0115] One example type of additional characteristics may be structural characteristics. The one or more structural characteristics may relate to a likelihood that the peptide will bind with one of one or more predefined molecular scaffolds. For example, one or more structural characteristics may relate to a likelihood that the peptide described by the sequence will bind to a scaffold at two locations.

[0116] One example structural characteristic may be a distance between two cysteine residues, in the predicted three-dimensional structure of the peptide. The distance may be between respective sulphur atoms of the cysteine residues. Favourability may be determined based on how the distance compares to a desired (or optimal) distance or distance range. For example, a desired (or optimal) distance range may be a range from 2 to 20 Angstrom. For particular scaffold molecules, such as TATA, a narrower range may be more desirable, such as a range of from 2.8 to 12.5 Angstrom. For particular scaffold molecules, such as TATB, a yet narrower range may be more desirable, such as a range of from 3.2 to 10 Angstrom. For particular scaffold molecules, such as TTZ, a yet narrower range may be more desirable, such as a range of from 4.5 to 8 Angstrom.

[0117] Another example structural characteristic may be an angle formed between a first cysteine residue and two further cysteine residues, in the predicted three-dimensional structure of the peptide. The angles may be between respective sulphur atoms of the cysteine residues. Favourability may be determined based on how the angle compares to a desired (or optimal) angle or range of angles. For example, a desired (or optimal) range of angles may be a range from 10 to 160 degrees. For particular scaffold molecules, such as TATA, a narrower range may be more desirable, such as a range of from 14 to 150 degrees. For particular scaffold molecules, such as TATB, a yet narrower range may be more desirable, such as a range of from 20 to 115 degrees. For particular scaffold molecules, such as TTZ, a yet narrower range may be more desirable, such as a range of from 25 to 85 degrees.

[0118] Another example structural characteristic may be a dihedral angle formed between pairs of cysteines. The dihedral angle may be defined as the angle between two planes defined by the quadruplet of positions Cal, SI, S2, Ca2; where Cal is the alpha carbon atom of the first cysteine residue, SI is the sulphur atom of the first cysteine residue, Ca2 is the alpha carbon atom of the second cysteine residue, S2 is the sulphur atom of the second cysteine residue. The dihedral angle may be calculated around the axis connecting positions SI and S2, as the angle between first and second planes defined by the positions Cal, SI, S2 and SI, S2, Ca2 respectively. Favourability may be determined based on how the angle compares to a desired (or optimal) angle or range of angles. For example, a desired (or optimal) range of angles may be a range of from -175 to 175 degrees. For particular scaffold molecules, such as TATB and / or TTZ, a narrower range may be more desirable, such as a range of from -170 to 170 degrees.

[0119] The predicted three-dimensional structure may be generated by a machine learning algorithm, based on input data comprising the generated amino acid sequence. The machine learning algorithm may be the AlphaFold™ algorithm, for example.

[0120] Functional characteristics

[0121] Another example type of additional characteristics may be functional characteristics. Examples of functional characteristics may include: a membrane permeability of the peptide and a blood-brain barrier permeability of the peptide. Corresponding functional requirements may include desired parameters for these characteristics. Favourability may be determined based on how each of these parameters compare to a desired (or optimal) parameter.

[0122] Functional characteristics may be determined based on the generated peptide sequence and / or a predicted three-dimensional structure of the peptide. This may be determined using known algorithms. For example, solubility, hydrophobicity, pKa, total charge, oxidation sensitivity, etc. (10.26434 / chemrxiv-2023-cwr53). Expansion of sequences

[0123] In an example method, sequences output by the methods described above, i.e. having favourable properties, may be diversified by identifying alternative sequences that fold to the same backbone structure. For example, this may be performed by a machine learning algorithm configured to generate a sequence with a high probability of forming the same backbone structure as an input peptide. The algorithm may take, as input, the structure of the peptide described by a sequence output by step S4. An example of such an algorithm is the ProteinMPNN algorithm (Dauparas et al (2022) Science 378(6615) 49-56 (doi: 10.1126 / science.add2187)).

[0124] Fig. 4 shows an example method in which output protein sequences are diversified as describe above. The additional sequences are generated in a first step S10. In a second step Si l, predicted interaction (and additional) characteristics may be determined. In a third step S12, favourability of the predicted interaction (and additional) characteristics may be determined. In a fourth step SI 3, the method may output additional sequences configured to form a scaffold-bound peptide ligand with favourable characteristics, based on the determination in step S13. Determination of peptide characteristics and favourability may be performed as described above in relation to the method of Fig. 1.

[0125] Results

[0126] Fig. 6, part A shows the structure of one peptide (above in blue) and associated target (below in green), based on a sequence identified by a method according to the disclosure. In particular, the method combined the method steps shown in Figs. 1, 2 and 4, using AlphaFold parameters for the loss function and ColabDesign. Lines (in red) between the peptide and the target show the hydrogen bonds between the peptide and the target. Fig. 6, part B, shows predicted binding affinities of 100 sequences identified by the algorithm showing all of them have a favourable affinity for the target.

[0127] Example uses

[0128] In a specific example use, methods described above may be used in a method of identifying a therapeutic agent or diagnostic agent targeting a target. The therapeutic agent or diagnostic agent may be a scaffold-bound peptide that binds to the target and the computer implemented methods described above may be used to identify a peptide forming the scaffold-bound peptide. In another specific example use, methods described above may be used in a method of making a peptide or scaffold-bound peptide ligand. A computer implemented method may be used to identify a peptide forming a scaffold-bound peptide ligand, prior to making the peptide or the scaffold-bound peptide ligand. It should be understood that variations of the above described examples are possible without departing from the spirit or scope of the invention.

Claims

CLAIMS1. A computer implemented method for identifying amino acid sequences predicted to form a scaffold-bound peptide ligand that binds to a target, the method comprising: generating an amino acid sequence describing a peptide, the sequence satisfying one or more predefined sequence constraints such that the peptide binds to one of one or more predefined scaffolds in at least two locations, to form a scaffold-bound peptide ligand; determining at least one predicted interaction characteristic relating to an interaction between the peptide and one or more predefined targets; determining a favourability of the at least one predicted interaction characteristic; providing output data identifying one or more amino acid sequences predicted to form a scaffold-bound peptide ligand that binds to a target, based on the determined favourability.

2. The method of claim 1, wherein the step of generating an amino acid sequence comprises modification of a previous amino acid sequence.

3. The method of claim 2, wherein the modification comprises adding at least one residue, removing at least one residue, or changing at least one residue of a previous amino acid sequence.

4. The method of claim 2 or 3, wherein the modification is made at a random position in the previous sequence.

5. The method of any one of claims 2 to 4, wherein the modification comprises the insertion of an amino acid selected based on a probability distribution.

6. The method of any preceding claim, wherein the probability distribution ensures that at least one of the one or more sequence constraints are satisfied.

7. The method of any one of claims 5 or 6, wherein the probability distribution is generated by: generating an amino acid sequence describing a peptide;determining at least one predicted interaction characteristic relating to an interaction between the peptide and one or more predefined targets; determining a favourability of the at least one predicted interaction characteristic; generating a probability distribution for all possible amino acid sequences and all possible positions, based on the generated amino acid sequences and the determined favourability.

8. The method of any preceding claim, wherein the step of generating an amino acid sequence, and subsequent steps, are repeated for a plurality of amino acid sequences.

9. The method of any preceding claim, wherein the output data comprises the amino acid sequence.

10. The method of any preceding claim, wherein the one or more sequence constraints comprises one or more constraints that increase a likelihood that the peptide will bind to one of one or more predefined scaffolds.

11. The method of any preceding claim, wherein the one or more sequence constraints comprises one or more constraints that increase a likelihood that the peptide will form at least one loop structure when bound with one of the one or more predefined scaffolds.

12. The method of any preceding claim, wherein the one or more sequence constraints comprises a constraint on a total length of the sequence.

13. The method of claim 12, wherein the one or more sequence constraints comprises a constraint that the sequence consists of from 6 to 30 amino acid residues.

14. The method of any preceding claim, wherein the one or more sequence constraints comprises a constraint that the sequence comprises at least two cysteine residues.

15. The method of any preceding claim, wherein the one or more sequence constraints comprises a constraint that the sequence comprises three cysteine residues.

16. The method of any preceding claim, wherein the one or more sequence constraints comprises a constraint on the relative positions of at least two cysteine residues.

17. The method of any preceding claim, wherein the one or more sequence constraints comprises a constraint that at least one pair of successive cysteine residues are separated by no more than a first predefined number of amino acid residues.

18. The method of any preceding claim, wherein the one or more sequence constraints comprises a constraint that at least one pair of successive cysteine residues are separated by no fewer than a second predefined number of amino acid residues.

19. The method of claim 15, wherein the sequence comprises three cysteine residues, including two pairs of successive cysteine residues, and both pairs satisfy at least one of the constraints defined in claim 17 and claim 18.

20. The method of any preceding claim, wherein the one or more constraints comprises a constraint on the amino acid residue at one or more terminal positions in the sequence.

21. The method of any preceding claim, wherein the one or more constraints comprises a constraint that the amino acid residue at one or more terminal positions in the sequence is an alanine residue.

22. The method of any preceding claim, wherein the one or more interaction characteristics comprises a binding score based on the likelihood of the peptide binding to the target.

23. The method of any preceding claim, wherein the one or more interaction characteristics comprises a folding score based on the likelihood of the peptide being well- folded24. The method of any preceding claim, wherein the one or more interaction characteristics comprises a score for one or more residues of the target, based on the likelihood of each of these residue binding to the peptide.

25. The method of any preceding claim, wherein the one or more interaction characteristics comprises a score for one or more residues of the peptide, based on the likelihood of each of these residues binding to the target.

26. The method of any preceding claim, wherein the one or more interaction characteristics is determined based on a predicted three-dimensional structure of the peptide and / or the target.

27. The method of any preceding claim, further comprising a step of determining one or more structural characteristics of the peptide and a step of determining a favourability of the one or more structural characteristics.

28. The method of claim 27, wherein the one or more structural characteristics relate to a likelihood that the peptide will bind with one of one or more predefined molecular scaffolds.

29. The method of claim 27 or 28, wherein the one or more structural characteristics comprises a distance between two cysteine residues, in the predicted three-dimensional structure of the peptide.

30. The method of any one of claims 27 to 29, wherein the one or more structural characteristics comprises a distance between respective sulphur atoms of two cysteine residues in the predicted three-dimensional structure of the peptide.

31. The method of any one of claims 27 to 30, wherein the one or more structural characteristics comprises an angle formed between a first cysteine residue and two further cysteine residues, in the predicted three-dimensional structure of the peptide.

32. The method of any one of claims 27 to 31 , wherein the one or more structural characteristics comprises a dihedral angle formed between a first cysteine residue and a second cystine residue, in the predicted three-dimensional structure of the peptide.

33. The method of any one of claims 27 to 32, wherein the predicted three-dimensional structure of the peptide is generated by a machine learning algorithm based on input data comprising the generated amino acid sequence.

34. The method of any preceding claim, further comprising a step of determining one or more functional characteristics of the peptide and a step of determining a favourability of the one or more functional characteristics.

35. The method of claim 34, wherein the one or more functional characteristics comprises one more of: membrane permeability of peptide and blood-brain barrier permeability of the peptide.

36. The method of any preceding claim, where the favourability is determined based on a comparison between the determined characteristics and predefined corresponding desired or optimal parameters.

37. The method of claim 36, wherein the favourability is determined by calculating a favourability score.

38. The method of claim 37, wherein the favourability score is calculated using an objective function comprising terms corresponding to each of the determined characteristics.

39. The method of any preceding claim, wherein the scaffold-bound peptide ligand is bicyclic, comprising two loop sequences, and is covalently bound to the scaffold at three locations.

40. The method of any preceding claim, wherein the one or more scaffolds comprise one or more of TATA, TATB, TCTZ, TBMB, TTZ, TBAB, TSTA, TCAZ, TCAN, and TCCU.

41. The method of any preceding claim, wherein the peptide forms a peptide ligand, and the peptide in the peptide ligand comprises at least three cysteine residues, separated by at least two loop sequences, and the scaffold forms covalent bonds with the cysteine residues of the peptide such that at least two peptide loops are formed on the scaffold.

42. The method of any preceding claim, wherein the target is a protein, a protein on a cell, a tumour antigen, a viral antigen, bacterial antigen.

43. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any one of the preceding claims.

44. A data processing system comprising means for carrying out the method of any one of claims 1-42.

45. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of any one of claims 1-42.

46. A method of identifying a therapeutic agent or diagnostic agent targeting a target, the therapeutic comprising a scaffold-bound peptide ligand, the method comprising using the computer implemented method of any one of claims 1-42 to identify a peptide forming the scaffold-bound peptide ligand.

47. A method of making a peptide or scaffold-bound peptide ligand, comprising using the computer implemented method of any one of claims 1-42 to identify a peptide forming the scaffold-bound peptide ligand or the scaffold-bound peptide ligand.

48. A scaffold-bound peptide ligand made by the method of claim 47.

Images