Patch scale generative molecule design with neural rotamer density estimation

Abstract

A method may include identifying a residue of interest within a protein molecule. The protein molecule may be generated by a molecule design computation model trained at a full length scale of protein molecules. An input patch may be generated to include the residue of interest and additional amino acid residues. A patch design computation model may be applied to determine, based on the input patch, an identity and conformation of the residue of interest. The patch design computation model may be trained at a patch scale to determine the identity and conformation of the residue of interest based on an atomic context of the residue of interest. The identity and the conformation of the residue of interest in the protein molecule may be validated based at least on an output of the patch design computation model.

Description

Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1PATCH SCALE GENERATIVE MOLECULE DESIGN WITH NEURAL ROTAMER DENSITY ESTIMATIONCROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to Greek Provisional Application No. 20240100601, entitled “NEURAL ROTAMER DENSITY ESTIMATION FOR GENERATIVE MOLECULE DESIGN” and filed on September 2, 2024, the disclosure of which is incorporated herein by reference in its entirety.TECHNICAL FIELD

[0002] The subj ect matter described herein relates generally to molecule design and more specifically to machine learning enabled inverse rotamer density estimation for determining the composition and conformation of a portion of a molecule.INTRODUCTION

[0002] A molecule is a group of two more atoms held together by chemical bonds. Molecules form the smallest identifiable unit into which a pure substance can be divided while still retaining the composition and chemical properties of that substance. Various properties of a molecule, including its ability to function as a therapeutic, may be contingent upon the composition (e g., chemical makeup) and, in the case of biologies, the conformation (or three-dimensional structure) of the molecule.

[0003] Small molecules, which are low-weight compounds having a molecular weight between approximately 100 Daltons and 1000 Daltons, are one example of a molecule with applications as therapeutics. A small molecule drug can modulate biochemical processes to diagnose, treat, and prevent a gamut of illnesses. Owing to a number of compelling advantages,NAI-5002497275vl 1Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 small molecule drugs have been a cornerstone in modern pharmacology. For example, small molecule drugs are capable of penetrating cell membranes to reach intracellular targets. Moreover, small molecule drugs are adaptable to a wide variety of therapeutic applications. For instance, a small molecule drug may be formulated as pills and capsules, intravenous or subcutaneous injectables, inhalational medicines, or suppositories. The development of the small molecule drug may further extend to tailoring various pharmacokinetic properties including liberation, absorption, distribution, metabolism, potency, efficacy, phenotypic effects, and excretion.

[0004] Another example of molecules with applications as therapeutics are large molecules, which are also known as biopharmaceuticals, biologicals, or biologies. Large molecules can range between approximately 3000 Daltons and 150,000 Daltons in molecular weight. Moreover, large molecule drugs are often derivatives of natural human proteins. Proteins are genetically encoded macromolecules whose diversity in size and chemical composition give rise to a gamut of functionalities. For example, by regulating biological systems, proteins facilitate many essential cellular functions including, for example, enzymatic reactions, molecular transport, regulation and execution of a number of biological pathways, cell growth, proliferation, nutrient uptake, morphology, motility, intercellular communication, and / or the like. A protein molecule may include one or more polypeptides, each of which including a sequence of amino acid residues linked together by peptide bonds (e.g., covalent peptide bonds). The amino acid residues that are encoded directly by the genetic code are called standard or canonical amino acid residues. Of the different types of amino acid residues, such as the twenty canonical amino acid residues, each type of amino acid residue may include the same backbone atoms (e.g., an amino group (NH2), an alpha carbon (Ca), and a carboxylic group (COOH)) coupled with a different combination of sidechain atoms forming the sidechain or radical (R) group of the amino acid residue.NAI-5002497275vl 2Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0005] The primary structure of a protein molecule refers to the sequence of amino acid residues in each of the polypeptide chains forming the protein molecule. The backbone atoms in adjacent amino acid residues that participate in the peptide bonds (e.g., covalent peptide bonds) therebetween form a repeating sequence of atoms known as the polypeptide backbone (or backbone) of the protein molecule. The local folded structures (e.g., a helixes, 0 pleated sheet, and / or the like) that form within an individual polypeptide chain due to interactions between the backbone atoms (e.g., amino hydrogen atoms, carboxyl oxygen atoms, and / or the like) are referred to as the secondary structure of the protein molecule. Further interactions (e.g., non-covalent bonds such as hydrogen bonding, ionic bonding, dipole-dipole interactions, and van der Waals forces) between the side chains (or R-groups) of the amino acid residues in the protein molecule may cause individual polypeptide chains to fold, thus forming the tertiary structure of the protein molecule. The tertiary structure of the protein molecule is also known as the conformation (or three-dimensional structure) of the protein molecule. In protein molecules having multiple polypeptide chains, the protein molecule may also exhibit a quaternary structure, which is formed when the polypeptide chains are packed and held together by hydrogen bonds and van der Waals forces (e.g., between nonpolar side chains).

[0006] The functions of a protein molecule may be contingent upon the sequence of amino acids in the polypeptide chains forming the protein molecule as well as the three- dimensional structure adopted by the polypeptide chains. For example, the primary structure of the protein molecule may determine the three-dimensional structure assumed by the protein molecule through the folding of the constituent polypeptide chains. In some cases, the binding affinity of the protein molecule towards a target molecule, such as a viral or tumor antigen, may depend on whether the polypeptide chains in the protein molecule are able to assume aNAI-5002497275vl 3Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 conformation (or three-dimensional structure) that complements that of the target molecule and is sufficiently stable to support a binding interaction between the two molecules. As such, one notable objective of computational molecule design is to construct one or more protein sequences (e g., antibodies and / or the like) that exhibit certain properties of interest, including the ability to adopt a particular conformation (or three-dimensional structure).SUMMARY

[0005] Systems, methods, and articles of manufacture, including computer program products, are provided for patch scale generative molecule design with inverse rotamer density estimation. In some example embodiments, a patch design computation model may be trained to determine the sequence and conformation of a partial protein molecule. That is, in some cases, the patch design computation model may be trained to determine the identity (or type) and conformation (e.g., sidechain conformation) of some but not all of the amino acid residues in a protein molecule and / or a target molecule (e.g., another protein molecule, a small molecule, and / or the like). For example, in some cases, the patch design computation model may be trained to operate on one or more patches in a protein molecule. In the context of molecule design, the term “patch” may refer a set of amino acid residues in the protein molecule that are within the neighborhood of a residue of interest, such as an amino acid residue in a binding interface between the protein molecule and a target molecule. In some cases, the patch may include one or more amino acid residues that are within a threshold distance of the residue of interest. Alternatively and / or additionally, the patch may include a threshold quantity of amino acid residues closest in space to the residue of interest. The patch design computation model may be trained to operate on individual patches within a protein molecule in order to leverage the greater availability of data atNAI-5002497275vl 4Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 the patch level. For instance, while data specifying the identity (or type) and conformation of amino acid residues across the entire length of protein molecules (e.g., antibodies and / or the like) may be limited, more data may be available for specific portions of the protein molecule (e.g., paratope, variable region (Fv), antigen binding fragment (Fab), variable region, complementarity determining region (CDR), and / or the like). Once trained, the patch design computation model may determine the identity (or type) and conformation (e.g., sidechain conformation) of amino acid residues within a patch of a protein molecule. For each amino acid residue included in the patch, the patch design computation model may output a probability distribution across the possible sidechain rotamers forming the amino acid residue. In some cases, the sidechain conformation of the amino acid residue may be determined by sampling from this probability distribution.

[0006] In some example embodiments, instead of or in addition to partial protein molecules, the patch design computation model may also be trained to determine the chemical composition and conformation of a portion of a small molecule (e.g., drug compound). For example, in some cases, the patch design computation model may be trained to operate on one or more “patches,” which in the context of a small molecule may include a subset of the atoms forming the small molecule. In some cases, each patch may include a functional group of interest, which in this context may refer to a functional group participating in an interaction (e.g., binding, modulation, and / or the like) with a target molecule (e.g., a target protein molecule). In some cases, each patch may further include one or more adjacent functional groups (or portion thereof) attached to a scaffold (or core structure) of the small molecule. For instance, in some cases, the patch design computation model may be applied to determine, at a patch level, the identity (or the patch level identity) of the functional group of interest (e.g., alcohol, amine, carbonyl, halide, aromatic ring,NAI-5002497275vl 5Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 and / or the like). In some cases, training the patch design computation model to operate on individual patches within a small molecule leverages the greater availability of data at the patch level such that the resulting patch design computation model achieves better performance than molecule design computation models operating at the full molecule scale. Once trained, the patch design computation model may be applied to determine the identity of one or more functional groups of interest in the small molecule. For instance, in some cases, the patch design computation model may output a probability distribution (e.g., categorical distribution) across the possible functional groups and the identity of each functional group of interest may be determined by sampling from this probability distribution.

[0007] In one aspect, there is provided a system for patch scale generative molecule design with inverse rotamer density estimation. The system may include at least one data processor and at least one memory. The at least one memory may store instructions that result in operations when executed by the at least one data processor. The operations may include: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity and a conformation determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the input patch, a patch level residue identity and a patch level conformation of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity and the patch level conformation of the residue of interest; and validating, based at least on the output of the patch design computation model, theNAI-5002497275vl 6Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 residue identity and the conformation of the residue of interest determined by the molecule design computation model.

[0008] In another aspect, there is a computer-implemented method for patch scale generative molecule design with inverse rotamer density estimation. The method may include: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity and a conformation determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the input patch, a patch level residue identity and a patch level conformation of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity and the patch level conformation of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity and the conformation of the residue of interest determined by the molecule design computation model.

[0009] In another aspect, there is provided a computer program product for patch scale generative molecule design with inverse rotamer density estimation. The computer program product may include a non-transitory computer readable medium storing instructions that result in operations when executed by at least one data processor. The operations may include: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity and a conformation determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch designNAI-5002497275vl 7Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 computation model to determine, based at least on the input patch, a patch level residue identity and a patch level conformation of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity and the patch level conformation of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity and the conformation of the residue of interest determined by the molecule design computation model.

[0010] In some variations, one or more features disclosed herein including the following features can optionally be included in any feasible combination.

[0011] In some variations, the output of the patch design computation model includes a probability distribution across a plurality of possible types of amino acid residues for the residue of interest.

[0012] In some variations, the output of the patch design computation model further includes, for at least one possible type of amino acid residue, a probability distribution across one or more possible rotamers of that type of amino acid residue.

[0013] In some variations, the output of the patch design computation model further includes, for at least one possible type of amino acid residue, a probability distribution across one or more possible rotamers of that type of amino acid residue.

[0014] In some variations, the output of the patch design computation model further includes, for each torsion angle present in each type of amino acid residue, a probability distribution across a plurality of different values.NAI-5002497275vl 8Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0015] In some variations, the probability distribution comprises a categorical distribution that includes, for each torsion angle present in a corresponding type of amino acid residue, a probability of the torsion angle being a first value and a probability of the torsion angle being a second value.

[0016] In some variations, the probability distribution comprises a continuous distribution.

[0017] In some variations, the output of the patch design computation model includes a probability of the residue of interest being a first type of amino acid residue and a probability of the residue of interest being a second type of amino acid residue.

[0018] In some variations, the output of the patch design computation model further includes a probability of the residue of interest being a rotamer of the first type of amino acid residue and a probability of the residue of interest being another rotamer of the first type of amino acid residue.

[0019] In some variations, the residue of interest is located at an interface between the protein molecule and a target molecule.

[0020] In some variations, the method further includes: generating the input patch by at least identifying, within the plurality of amino acid residues forming the protein molecule and / or a plurality of amino acid residues forming the target molecule, the one or more additional amino acid residues for inclusion in the input patch.

[0021] In some variations, the method further includes: selecting, within a plurality of amino acid residues including the plurality of amino acid residues forming the protein moleculeNAI-5002497275vl 9Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 and excluding a plurality of amino acid residues forming the target molecule, the one or more additional amino acid residues for inclusion in the input patch.

[0022] In some variations, the target molecule is another protein molecule or a small molecule.

[0023] In some variations, the protein molecule is an antibody and the target molecule is an antigen.

[0024] In some variations, the method further includes: generating an embedding of the input patch to include a description of the atomic context of the one or more additional amino acid residues; and applying the patch design computation model to determine, based at least on the embedding of the input patch, the patch level identity and the patch level conformation of the residue of interest.

[0025] In some variations, the description of the atomic context of the one or more additional amino acid residues include an encoding of an identity of each additional amino acid residue.

[0026] In some variations, the description of the atomic context of the one or more additional amino acid residues include a positional encoding of a sequential position of each additional amino acid residue in a corresponding sequence of amino acid residues.

[0027] In some variations, the description of the atomic context of the one or more additional amino acid residues include one or more three-dimensional coordinates of a position of at least one backbone atom in each additional amino acid residue.

[0028] In some variations, at least one backbone atom includes an alpha carbon (Ca) atom.NAI-5002497275vl 10Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0029] In some variations, the method further includes: extracting, from a complex formed by a binding interaction between a sample protein molecule and a sample target molecule, a sample patch including one or more amino acid residues; generating a training sample in which an identity and a conformation of one or more amino acid residues in the sample patch are masked; and training, based at least on the training sample, the patch design computation model to recover an identity and a conformation of the one or more masked amino acid residues.

[0030] In some variations, the method further includes: applying the molecule design computation model to generate the protein molecule, where the molecule design computation model is trained at a scale of protein molecules, and where the molecule design computation model identifies a first type of amino acid residue as the identity of the residue of interest.

[0031] In some variations, the validating the identity and the conformation of the residue of interest includes determining, based at least on the output of the patch design computation model, a probability of the residue of interest being the first type of amino acid residue.

[0032] In some variations, the validating the identity and the conformation of the residue of interest further includes in response to the probability of the residue of interest being the first type of amino acid residue satisfying one or more thresholds, determining, based at least on the output of the patch design computation model, a probability of the residue of interest being a rotamer of the first type of amino acid residue identified by the molecule design computation model, and in response to the probability of the residue of interest being the rotamer of the first type of amino acid residue failing to satisfy one or more thresholds, replacing, based at least on the output of the patch design computation model, the rotamer of the first type of amino acid residue with a different rotamer of the first type of amino acid residue.NAI-5002497275vl 11Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0033] In some variations, the configuration of the protein molecule comprises a conformation of one or more amino acid residues forming the protein molecule.

[0034] In some variations, the probability of the configuration of the protein molecule is further determined by at least applying the patch design computation model to determine a probability of a different residue of interest in the protein molecule being a rotamer of a second type of amino acid residue identified by the molecule design computation model.

[0035] In some variations, the validating the identity and conformation of the residue of interest further includes determining, based at least on the output of the patch design computation model, a probability of an angular value of one or more torsion angles present in the first type of amino acid residue identified by the molecule design computation model.

[0036] In some variations, the validating the identity and conformation of the residue of interest further includes in response to the probability of the one or more torsion angles present in the first type of amino acid residue identified by the molecule design computation model failing to satisfy one or more thresholds, adjusting, based at least on the output of the patch design computation model, a value of the one or more torsion angles determined by the molecule design computation model.

[0037] In some variations, the method further includes: in response to the probability of the residue of interest being the first type of amino acid residue failing to satisfy one or more thresholds, replacing, based at least on the output of the patch design computation model, the first type of amino acid residue with a second type of amino acid residue, and identifying, based at least on the output of the patch design computation model, a rotamer of the second type of amino acid residue as the residue of interest.NAI-5002497275vl 12Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0038] In some variations, the method further includes: determining, based at least on the output of the patch design computation model, a probability of a configuration of the protein molecule generated by the molecule design computation model.

[0039] In some variations, the probability of the configuration of the protein molecule is determined by at least determining, based at least on the output of the patch design computation model, a probability of the residue of interest being the rotamer of the first type of amino acid residue.

[0040] In some variations, the method further includes: identifying, as the one or more additional amino acid residues in the patch, one or more amino acid residues located within a threshold distance of the residue of interest.

[0041] In some variations, the method further includes: identifying, as the one or more additional amino acid residues in the patch, a threshold quantity of amino acid residues that are closest to the residue of interest.

[0042] In some variations, the atomic context around the residue of interest includes the one or more additional amino acid residues.

[0043] In some variations, the atomic context around the residue of interest includes one or more of a fluid, a ligand, a cofactor, or a crystallographic adjuvant.

[0044] In another aspect, there is provided a system for validating an output of a molecule design computation model using an output of a patch design computation model. The system may include at least one data processor and at least one memory. The at least one memory may store instructions that result in operations when executed by the at least one data processor. The operations may include: selecting, for validation, a residue of interest from a plurality of aminoNAI-5002497275vl 13Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 acid residues forming a protein molecule, where the residue of interest is associated with a residue identity determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the input patch, a patch level residue identity of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity of the residue of interest determined by the molecule design computation model.

[0045] In another aspect, there is a computer-implemented method for validating an output of a molecule design computation model using an output of a patch design computation model. The method may include: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the input patch, a patch level residue identity of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity of the residue of interest determined by the molecule design computation model.

[0046] In another aspect, there is provided a computer program product for validating an output of a molecule design computation model using an output of a patch design computationNAI-5002497275vl 14Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 model. The computer program product may include a non-transitory computer readable medium storing instructions that result in operations when executed by at least one data processor. The operations may include: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the input patch, a patch level residue identity of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity of the residue of interest determined by the molecule design computation model.

[0047] In some variations, one or more features disclosed herein including the following features can optionally be included in any feasible combination.

[0048] In some variations, the patch design computation model has been further trained to generate, based at least on the atomic context around the residue of interest, the output to be indicative of a patch level conformation of the residue of interest.

[0049] In some variations, the method further includes: applying the patch design computation model to determine, based at least on the input patch, a patch level conformation of the residue of interest; and validating, based at least on the output of the patch design computation model, a conformation of the residue of interest determined by the molecule design computation model.NAI-5002497275vl 15Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0050] In some variations, the output of the patch design computation model includes a probability distribution across a plurality of possible types of amino acid residues.

[0051] In some variations, the output of the patch design computation model further includes, for each type of amino acid residue of the plurality of possible types of amino acid residues, a probability distribution across a plurality of possible rotamers.

[0052] In some variations, the output of the patch design computation model further includes, for each torsion angle present in the residue of interest, a probability distribution across a plurality of possible angular values.

[0053] In some variations, the probability distribution includes a probability of a first torsion angle present in the residue of interest having one angular value and a probability of the first torsion angle having a different angular value.

[0054] In some variations, the probability distribution further includes a probability of a second torsion present in the residue of interest having the one angular value and a probability of the second torsion angle having the different angular value.

[0055] In another aspect, there is provided a system for validating an output of a molecule design computation model using an output of a patch design computation model. The system may include at least one data processor and at least one memory. The at least one memory may store instructions that result in operations when executed by the at least one data processor. The operations may include: selecting, for validation, a functional group of interest from a small molecule generated by the molecule design computation model; generating an input patch to include the functional group of interest; applying the patch design computation model to determine, based at least on the input patch, a patch level identity of the functional group ofNAI-5002497275vl 16Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the functional group of interest, an output indicative of the patch level residue identity of the functional group of interest; and validating, based at least on the output of the patch design computation model, an identity of the functional group of interest determined by the molecule design computation model.

[0056] In another aspect, there is a computer-implemented method for validating an output of a molecule design computation model using an output of a patch design computation model. The method may include: selecting, for validation, a functional group of interest from a small molecule generated by the molecule design computation model; generating an input patch to include the functional group of interest; applying the patch design computation model to determine, based at least on the input patch, a patch level identity of the functional group of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the functional group of interest, an output indicative of the patch level residue identity of the functional group of interest; and validating, based at least on the output of the patch design computation model, an identity of the functional group of interest determined by the molecule design computation model.

[0057] In another aspect, there is provided a computer program product for validating an output of a molecule design computation model using an output of a patch design computation model. The computer program product may include a non-transitory computer readable medium storing instructions that result in operations when executed by at least one data processor. The operations may include: selecting, for validation, a functional group of interest from a small molecule generated by the molecule design computation model; generating an input patch to include the functional group of interest; applying the patch design computation model toNAI-5002497275vl 17Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 determine, based at least on the input patch, a patch level identity of the functional group of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the functional group of interest, an output indicative of the patch level residue identity of the functional group of interest; and validating, based at least on the output of the patch design computation model, an identity of the functional group of interest determined by the molecule design computation model.

[0058] In some variations, one or more features disclosed herein including the following features can optionally be included in any feasible combination.

[0059] In some variations, the output of the patch design computation model includes a probability distribution across a plurality of possible functional groups.

[0060] In some variations, the plurality of possible functional groups include one or more of an alcohol, an amine, a carbonyl, a halide, or an aromatic ring.

[0061] In some variations, the validating of the identity of the functional group of interest determined by the molecule design computation model includes in response to the probability of the functional group of interest a first type of functional group determined by the molecule design computation model failing to satisfy one or more thresholds, replacing, based at least on the output of the patch design computation model, the first type functional group determined by the molecule design computation model with a second type of functional group having a threshold probability of being the functional group of interest.

[0062] In some variations, the patch further includes at least a portion of one or more adjacent functional groups in the small molecule.NAI-5002497275vl 18Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0063] In some variations, the functional group of interest is located at an interface between the small molecule and a protein molecule.

[0064] In some variations, the patch further includes at least a portion of one or more amino acid residues from the protein molecule, and the one or more amino acid residues are located at an interface between the small molecule and the protein molecule.

[0065] In some variations, the patch design computation model determines the patch level identity of the functional group of interest based at least on an embedding of the patch that includes a description of the atomic context surrounding the functional group of interest.

[0066] Implementations of the current subject matter can include, but are not limited to, methods consistent with the descriptions provided herein as well as articles that comprise a tangibly embodied machine-readable medium operable to cause one or more machines (e.g., computers, etc.) to result in operations implementing one or more of the described features. Similarly, computer systems are also described that may include one or more processors and one or more memories coupled to the one or more processors. A memory, which can include a non- transitory computer-readable or machine-readable storage medium, may include, encode, store, or the like one or more programs that cause one or more processors to perform one or more of the operations described herein. Computer implemented methods consistent with one or more implementations of the current subject matter can be implemented by one or more data processors residing in a single computing system or multiple computing systems. Such multiple computing systems can be connected and can exchange data and / or commands or other instructions or the like via one or more connections, including, for example, to a connection over a network (e.g. the Internet, a wireless wide area network, a local area network, a wide area network, a wired network, or the like), via a direct connection between one or more of the multiple computing systems, etc.NAI-5002497275vl 19Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0067] The details of one or more variations of the subject matter described herein are set forth in the accompanying drawings and the description below. Other features and advantages of the subject matter described herein will be apparent from the description and drawings, and from the claims. While certain features of the currently disclosed subject matter are described for illustrative purposes in relation to the computational design of protein molecules, such as antibodies, it should be readily understood that such features are not intended to be limiting. The claims that follow this disclosure are intended to define the scope of the protected subject matter.DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings, which are incorporated in and constitute a part of this specification, show certain aspects of the subject matter disclosed herein and, together with the description, help explain some of the principles associated with the disclosed implementations. In the drawings,

[0069] FIG. 1 depicts a system diagram illustrating an example of a molecule design system, in accordance with some example embodiments;

[0070] FIG. 2 depicts a flowchart illustrating an example of a process for computational protein design, in accordance with some example embodiments;

[0071] FIG. 3 depicts a flowchart illustrating another example of a process for computational protein design, in accordance with some example embodiments;

[0072] FIG. 4 depicts a flowchart illustrating another example of a process for computational protein design, in accordance with some example embodiments;NAI-5002497275vl 20Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0073] FIG. 5A depicts a flowchart illustrating an example of a process for computational small molecule design, in accordance with some example embodiments;

[0074] FIG. 5B depicts a flowchart illustrating another example of a process for computational small molecule design, in accordance with some example embodiments;

[0075] FIG. 6A depicts a schematic diagram illustrating an example of a process for computational protein design, in accordance with some example embodiments;

[0076] FIG. 6B depicts a schematic diagram illustrating an example of an amino acid residue, in accordance with some example embodiments; and

[0077] FIG. 7 depicts a block diagram illustrating an example of a computing system, in accordance with some example embodiments.

[0078] When practical, similar reference numbers denote similar structures, features, or elements.DETAILED DESCRIPTION

[0079] Drug design (or drug discovery) refers to the process of discovering molecules, including small molecule drugs and biopharmaceuticals, with applications as therapeutics. For example, a drug molecule may modulate a biomolecular target (e.g., a protein, a nucleic acid, and / or the like) involved in the mechanisms of a disease, thus engendering a therapeutic benefit to the patients treated with the drug molecule. Accordingly, to be viable as a therapeutic, a molecule may be designed to exhibit one or more drug-like properties, such as binding affinity, specificity, biological activity, developability, and / or the like. In the case of molecular drug design, a lead molecule known to exhibit some activity against a biomolecular target undergoes modifications toNAI-5002497275vl 21Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 further enhance efficacy, safety, and / or pharmacokinetics. Contrastingly, with de novo drug design, a molecule capable of modulating a biomolecular target may be designed from scratch, absent any lead molecules with clinically useful pharmacological or biological activities.

[0080] The combinatorial design spaces of possible candidate molecules for small and large molecules are prohibitively large for conventional wet-lab based drug discovery methodologies. For example, the molecular space (or chemical space) of possible small molecules is estimated to contain IO60possible chemical compounds and scales exponentially with molecule size (e.g., the number of constituent atoms). Meanwhile, the size of the combinatorial space is magnitudes larger for larger molecules and biologies. For instance, for a protein molecule containing an N quantity of amino acid residues, approximately 20Npossible protein sequences exist if each of the N quantity of amino acid residues is one of the twenty canonical amino acid residues. The exorbitant number of possible candidate molecules that require screening for druglike properties renders wet-lab assays neither feasible nor affordable. Although computational tools reduce reliance on expensive wet-lab assays, the data driven nature of these solutions is ill suited for drug discovery’s typical low-data regimes. Conventional computational tools in low data regimes perform especially poorly for de novo drug design where a molecule with one or more properties of interest, such as the ability to modulate a biomolecular target, is generated without the benefit of any lead molecules with clinically useful pharmacological or biological activities.

[0081] As noted, computational tools may be applied in drug discovery in order to reduce reliance on wet-lab assays, which are too costly to support an extensive exploration of the combinatorial design space of possible candidate molecules. For example, computational molecule design aims to generate protein sequences that exhibit a variety of properties of interest.NAI-5002497275vl 22Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1In the context of large molecule drug discovery (LMDD), the properties of a protein sequence may determine its viability as a protein-based therapeutic such as antibodies, enzymes, growth factors, hormones, interferons, interleukins, thrombolytics, and / or the like. Accordingly, as a proteinbased therapeutic, a protein sequence may be computationally engineered for affinity and targetability, in vivo stability, pharmacokinetics, cell permeability, and non-immunogenicity. Computational molecule design is a challenging and resource intensive task at least because protein molecules exhibit countless variations in sequence and conformation (or three-dimensional structure) but only a small proportion of these variants will have any therapeutic value. At the outset, of the 20Lpossible protein sequences formed by an L-quantity of amino acid residues selected from the twenty canonical amino acid residues, few will have the drug-like properties (e.g., affinity, specificity, biological activity, and developability) required for a protein-based therapeutic. Moreover, many critical properties of a protein molecule, such as its binding affinity towards a target molecule, are also contingent on the three-dimensional structure formed by the folding of the underlying sequence of amino acid residue. Thus, in addition to navigating numerous variations in possible protein sequences, computational molecule design is further complicated by the need to accurately predict, from amongst a myriad of possibilities, the three- dimensional structure of individual protein sequences. For instance, even if each of the L-quantity of amino acid residue in a protein sequence is confined to an A-quantity of discrete geometric states (e.g., rotamers), predicting the three-dimensional structure of the protein sequence may still entail exploring up to NLpossibilities.

[0082] Due to the sheer number of possible protein sequences and conformations, a naive design approach that relies on a brute force examination of every possible sequence and conformational variation is computationally intractable for protein sequences of meaningful lengthNAI-5002497275vl 23Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1(e g., 110 amino acid residues in the variable domain of an antibody). However, indiscriminate efforts to reduce computational burden, for example, by assessing a random selection of possible protein sequences, may result in protein sequences with better properties than those in the existing repertoire being inadvertently overlooked. Moreover, given the nexus between protein sequence and conformation, segregating sequence design and structural predictions may also yield suboptimal outcomes. Thus, the protein sequences that are generated by a language-based model without concomitant structural predictions are less likely to exhibit properties of interest, such as binding affinity and stability, at least because the language-based model lacks an awareness of the structural characteristics that contribute to these properties. Conventional efforts to integrate protein sequence and structure design are thwarted by the scarcity of known protein structures. For example, the extent to which a mutation in a protein sequence, such as the insertion, deletion, or substitution of one or more amino acid residues, influences the binding affinity of the corresponding protein molecule is known for merely a few thousand such mutations, far too few for training a molecule design computation model to operate with adequate performance at the full length scale of protein molecules. While capable of predicting the backbone conformation of a protein molecule, physics based molecule design computation models perform poorly when applied to recover the corresponding sidechain conformation. For at least this reason, the performance of molecule design computation models trained to operate at the length scale of protein molecules remains subpar. For instance, a molecule design computation model may be trained to jointly determine the identity (or type) and conformation (e g., sidechain conformation) of every amino acid residue in a protein molecule. However, due to the difficulty and expense associated with crystallizing protein molecules, too few known protein structures are available to train the molecule design computation model to adequately capture the nuanced relationshipsNAI-5002497275vl 24Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 between the identity (or type) and conformation of amino acid residues across the entire length of protein sequences. As such, the molecule design computation model may be incapable of generating entire protein molecules with a sufficiently high probability of exhibiting one or more properties of interest, including the ability to adopt a particular conformation (e.g., a conformation that complements that of a target molecule).

[0083] It should be appreciated that the design of small molecules is also constrained by similar limitations arising from a low data regime. As noted, the molecular space (or chemical space) of possible small molecules is estimated to contain IO60possible chemical compounds and scales exponentially with molecule size (e.g., the number of constituent atoms). The functional groups forming a small molecule may affect its chemical properties, such as solubility, acidity (or alkalinity), reactivity, and / or the like. As such, the constituent functional groups of a small molecule may determine how that small molecule interacts with other molecules. However, merely a small fraction of the possible small molecules in the molecular space (or chemical space) have undergone laboratory testing to assess various drug-like properties like affinity, specificity, biological activity, and developability. By extension, the extent to which a change in the identity of one or more functional groups affects the drug-like properties of the resulting small molecule is indeterminate for many such changes. Consequently, an insufficient quantity of small molecules with known drug-like properties are available to train a molecule design computation model to operate at the full molecule scale. The performance of a molecule design computation model trained with too few small molecules with known drug-like properties is generally poor, especially in its inability to generate small molecules outside of the distribution of those used to train the molecule design computation model.NAI-5002497275vl 25Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0084] Various embodiments of the present disclosure address the paucity of data at the full molecule scale. In the context of molecule design, the lack of data at the full-length scale of protein molecules may be by providing for a patch design computation model that is trained to operate on a patch scale instead of the length scale of protein molecules. As used herein, the term “patch” may refer to a set of amino acid residues in a protein molecule and / or a target molecule. In this context, the target molecule may be a different molecule, such as another protein molecule (e.g., a viral antigen, a tumor antigen, and / or the like), a small molecule (e.g., a drug compound), and / or the like. In some cases, the target molecule may be a different molecule (e.g., another protein molecule) that is a binding target of interest for the protein molecule such that the protein molecule and the target molecule are able to participate in a binding interaction to form a stable complex. In some cases, this set of amino acid residues may include those amino acid residues in the protein molecule and / or the target molecule that are located within a threshold distance (e.g., radius and / or the like) of a residue of interest. Alternatively and / or additionally, the set of amino acid residues may include a threshold quantity of amino acid residues that are closest in space to the residue of interest. In some cases, the residue of interest may be an amino acid residue in the protein molecule that is located in a binding interface between the protein molecule and the target molecule. Furthermore, in some cases, the residue of interest may be at the center of the patch that also includes one or more other amino acid residues in the protein molecule and / or the target molecule located. In instances where the target molecule is a small molecule, the patch may include a set of amino acid residues from the protein molecule and at least a portion of the atoms forming the small molecule.

[0085] Accordingly, a patch design computation model operating at the patch scale may be trained to determine the identity (or type) and conformation (e.g., sidechain conformation) ofNAI-5002497275vl 26Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 some but not all of the amino acid residues in a protein molecule. For instance, in some cases, the patch design computation model may operate on a patch in order to determine the identity (or type) and conformation (e.g., sidechain conformation) of a residue of interest included in the patch. As described in more details below, in some cases, the patch design computation model may determine the identity (or type) and conformation (e.g., sidechain conformation) of a residue of interest in a patch by operating on an embedding of the patch that includes a description of each amino acid residue included in the patch including, for example, identity, sequential position, spatial position, and / or the like. In some cases, the description of an amino acid residue may be at an atomic resolution, meaning that the description of the amino acid residue may include the identity and spatial position of one or more constituent atoms. In doing so, the patch design computation model may leverage the greater availability of data at the patch level. In some cases, for example, the patch design computation model may be trained to operate on patches corresponding to specific regions within a protein molecule, such as a protein molecule generated by a molecule design computation model that is trained to operate on the full length scale of protein molecules. More data may be available for regions within a protein molecule, such as the paratope, the variable region (Fv), the antigen binding fragment (Fab), or the complementarity determining region (CDR) of an antibody. As such, the patch design computation model may be trained to better capture the relationships between the identity and conformation of amino acid residues at the patch level.

[0086] In some example embodiments, the patch design computation model may be trained based on a training dataset that includes sample patches from protein molecules exhibiting one or more properties of interest. For example, in the case of antibody design, the training dataset may include sample patches from the paratope, the variable region (Fv), the antigen bindingNAI-5002497275vl 27Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 fragment (Fab), or the complementarity determining region (CDR) of antibodies that exhibit properties of interest such as expression, binding affinity towards a target molecule, binding specificity towards the target molecule, lack of nonspecificity, stability, non-immunogenicity, human-ness, absence of self-association (or non-aggregation), lack of chemical liability motifs, and / or the like. By exposing the patch design computation model to sample patches from protein molecules exhibiting one or more properties of interest, the patch design computation model may be trained to capture the physics governing protein structure, stability, and association at the patch level. In some cases, one or more amino acid residues in each sample patch may be masked. For instance, in some cases, the identity (or type) and conformation (e.g., sidechain conformation) of one or more amino acid residues in a sample patch, such as the amino acid residue located at a center of the sample patch, may be changed to a special mask token. The patch design computation model may be trained to recover the identity (or type) and conformation (e.g., sidechain conformation) of one or more masked amino acid residues in each sample patch.

[0087] In some example embodiments, the identity or type of an amino acid residue in a patch may be specified as a probability distribution across the set of possible amino acid residues, such as the twenty canonical amino acid residues. For example, in some cases, the probability distribution may be a categorical distribution that includes a probability of a first type of amino acid residue, a probability of a second type of amino acid residue, and / or the like. As used herein, the term “identity” or “type” may be used interchangeably to refer to the different kinds of amino acid residues that can be found in a protein molecule. As noted, each type of amino acid residue is formed by the same combination of backbone atoms (e.g., an amino group (NH2), an alpha carbon (Ca), and a carboxylic group (COOH)) coupled with a different sidechain or radical (R) group that includes a different combination of sidechain atoms. Accordingly, in some cases, twoNAI-5002497275vl 28Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 different types of amino acid residues may have two different sidechains (or R groups). For instance, the sidechain of glycine is a single hydrogen (-H) atom while the sidechain of alanine is a methyl group (-CH3).

[0088] In some example embodiment, given a particular type of amino acid residue, the conformation of the amino acid residue may be defined as a probability distribution across a set of possible rotamers associated with that type of amino acid residue. In some cases, the sidechain (or R-group) of an amino acid residue may assume different conformations, each of which including a different spatial arrangement of the constituent sidechain atoms. As such, in the context of molecule design, the conformation of an amino acid residue may be described in terms of the conformation of its sidechain (or R-group). The different sidechain conformations of an amino acid residue may be described in terms of rotamers. For example, different rotamers of an amino acid residue may exhibit sidechain conformations. It should be appreciated that the difference between two rotamers may vary depending on how rotamers are defined. In some cases, different definitions of rotamers may impose different constraints on how and to what extent the constituent sidechain atoms can move in three-dimensional space. These constraints may be referred to as degrees-of-freedom (DoF). For instance, in some cases, rotamers are rotational isomers, which are a form of isomers in which an identical composition of atoms vary structurally due to rotations around the interconnecting bonds between the atoms. Accordingly, in some cases, different rotamers of an amino acid residue may exhibit different combinations of torsion angles (or chi angles) between the constituent sidechain atoms. However, it may also be possible for different rotamers of an amino acid residue to exhibit variations in the distances (e.g., bond lengths) between the constituent atoms. In this instance, different rotamers of an amino acid residue may also exhibit different combinations of distances (e.g., bond lengths) between the constituent sidechain atoms.NAI-5002497275vl 29Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1In some cases, each rotamer may correspond to a specific sidechain conformation as defined by a discrete combination of torsion angles and, in some cases, bond lengths. However, it should be appreciated that, as used herein, the term rotamer may also refer to the sidechain conformation of an amino acid residue generally. That is, instead of the possible sidechain conformations of an amino acid residue being defined as a set of discrete rotamers, each of which having a fixed combination of torsion angles, the sidechain conformation of the amino acid residue may vary across the continuum of possible torsion angles. Rotamer in that case refers to the sidechain conformation of the amino acid residue generally rather than a specific combination of torsion angles that is one of several possible combinations of torsion angles.

[0089] In some example embodiments, the conformation of the sidechain of an amino acid residue may be defined based on the constituent torsion angles instead of as a single rotamer with a fixed combination of torsion angles and, in some cases, bond lengths. Defining sidechain conformation based on individual torsion angles may increase the granularity of the representation of the amino acid residue, which in turn increases the expressiveness of the patch design computation model (or its ability to capture the complexity and variability of sidechain conformations). In some cases, each torsion angle in the amino acid residue may be represented as a continuous value, such as a probability distribution across a continuous variable (e.g., a probability density function (PDF) and / or the like). Alternatively, instead of a continuous value, each torsion angle in the amino acid residue may be represented as a probability distribution (e.g., a categorical distribution) across a set of discrete angular values (e.g., degrees, radians, and / or the like), for example, between —180° and +180°. For example, in some cases, each torsion angle may be represented as a probability distribution across the values — 180°, —90°, +90°, and + 180°. In some cases, the set of discrete angular values associated with each torsion angle may includeNAI-5002497275vl 30Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 more possible values to increase the granularity of the representation or fewer possible values to decrease the granularity of the representation. In some cases, the sidechain conformation of the amino acid residue may be determined by at least sampling from the probability distributions of each constituent torsion angle. For instance, in some cases, the sidechain conformation of the amino acid residue may include, for each torsion angle, a value identified by the patch design computation model as having a threshold probability (or a highest probability) of being the value of the torsion angle.

[0090] In some example embodiments, the patch design computation model may operate on an input patch that includes a set of the amino acid residues from a protein molecule and / or a target molecule. For example, in some cases, the set of amino acid residues may include at least one residue of interest that is located at an interface between the protein molecule and the target molecule. In some cases, the set of amino acid residues may also include one or more amino acid residues located within a threshold distance of at least one residue of interest at the interface between the protein molecule and the target molecule. Alternatively and / or additionally, the set of amino acid residues may also include a threshold quantity of amino acid residues that are located closest to the at least one residue of interest. In some cases, the at least one residue of interest may be located at the center of the input patch while the other amino acid residues in the patch may be a part of the protein molecule and / or the target molecule. In instances where the target molecule is a small molecule, it should be appreciated that the patch may include one or more of the atoms forming the target molecule. As noted, in some cases, the patch design computation model may be trained based on sample patches from protein molecules exhibiting one or more properties of interest, such as a binding affinity towards the same target molecule or a different target molecule, in order to capture the physics governing protein structure, stability, and association at the patchNAI-5002497275vl 31Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 level. Accordingly, in some cases, the patch design computation model may operate on the input patch by determining, at a patch level, the identity and / or conformation of residue of interest such that the identity and conformation of the residue of interest conform to the patch level physics learned by the patch design computation model. For instance, in some cases, the patch design computation model may determine, based on a description of the amino acid residues included in an embedding of the patch, the identity (or type) of the residue of interest that is at the center of the input patch. Furthermore, in some cases, the patch design computation model may determine, based on the description of the amino acid residue included in the embedding of the patch, the sidechain conformation of the residue of interest. Because the patch design computation model leverages the relative abundance of sample patches by being trained to operate at the patch level, the identity and conformation of the residue of interest determined by the patch design computation model may exhibit a greater likelihood of being a part of a protein molecule exhibiting the one or more properties of interest than the identity and conformation of a corresponding amino acid residue in a protein molecule generated by a molecule design computation model trained to operate at the full length scale of protein molecules.

[0091] In some example embodiments, the patch design computation model may determine one or more probability distributions specifying the identity (or type) and conformation (e.g., sidechain conformation) of each amino acid residue in the output patch. In some cases, the patch design computation model may output, for the residue of interest in the input patch, a probability distribution (e.g., categorical distribution) across the different possible types of amino acid residues (e.g., 20 canonical amino acid residues). For example, in some cases, the output of the patch design computation model may include a probability that the residue of interest is a first type of amino acid residue, a probability that the residue of interest is a second type of amino acidNAI-5002497275vl 32Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 residue, and / or the like. Furthermore, for each possible type of amino acid residue that may occupy a position in the output patch, the patch design computation model may further determine a probability distribution (e.g., categorical distribution) across the different possible rotamers or, alternatively, a probability distribution (e.g., a continuous distribution, a categorical distribution, and / or the like) of the value of each constituent torsion angle. As noted, in some cases, each type of amino acid residue may be associated with a different set of rotamers, each of which having a fixed combination of torsion angles and, in some cases, bond lengths. For instance, while alanine and glycine have a single possible rotamer each, glutamine, serine, and threonine may have up to 55 different rotamers. Accordingly, in some cases, the output of the patch design computation model may include, for each possible type of amino acid residue, a probability that the sidechain conformation of the residue of interest is one rotamer, a probability that the sidechain conformation of the residue of interest is a different rotamer, and / or the like. Alternatively, for a more granular representation, each amino acid residue in the output patch may be represented as a probability distribution (e.g., a continuous distribution, a categorical distribution, and / or the like) across the possible values of each constituent torsion angle.

[0092] In some example embodiments, the output of the patch design computation model may be used to guide the design of entire protein molecules. For example, in some cases, a molecule design computation model may be trained and applied to determine the sequence and conformation of an entire protein molecule. In some cases, the molecule design computation model may be applied to generate the protein molecule de novo, meaning that the protein molecule is generated from scratch instead of by modifying another molecule (e.g., lead molecule). The dearth of data to train the molecule design computation model to operate at adequate performance at the full molecule scale may be especially acute for de novo design, which is why proteinNAI-5002497275vl 33Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 molecules generated de novo at the full molecule scale tend to lack the necessary properties, such as binding affinity, specificity, biological activity developability, and / or the like. Contrastingly , due to the greater availability of data at the patch level than at the full length scale of protein molecules, the patch design computation model may better capture the physics that govern protein structure, stability, and association at the patch level than a protein molecule design computation operating on the length scale of protein molecules. As such, in some cases, the identity and conformation of one or more amino acid residues in a protein molecule generated by the molecule design computation model may be validated and, in some cases, corrected based on the output of the patch design computation model. For instance, while the molecule design computation model operating on the entire sequence of a protein molecule may determine that the type and sidechain conformation of a particular amino acid residue in the protein molecule is a rotamer of a first type of amino acid residue, the output of the patch design computation model operating on a patch of the protein molecule that includes that amino acid residue (e.g., as a residue of interest at the center of the patch) may indicate that the amino acid residue is more likely to be a different rotamer of the first type of amino acid residue or a rotamer of a second type of amino acid residue. In instances where there is a discrepancy between the output of the molecule design computation model and that of the patch design computation model, the identity (or type) and conformation (e.g., sidechain conformation) of one or more amino acid residues in the protein molecule generated by the molecule design computation model may be modified in accordance with the output of the patch design computation model. In the foregoing example, the protein molecule generated by the molecule design computation model may be modified in accordance with the output of the patch design computation model including by replacing the rotamer identified by the molecule design computation model with another rotamer identified by the patch design computation model.NAI-5002497275vl 34Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0093] In some example embodiments, the patch design computation model may be trained to determine the chemical composition and conformation of a portion of a small molecule, such as a patch within the small molecule, instead of or in addition to partial protein molecules. In the context of small molecule design, a “patch” may include a subset of atoms forming a small molecule. For example, in some cases, the patch design computation model may operate on a patch that includes a functional group of interest and one or more functional groups (or portions thereof) that are within a threshold distance of the functional group of interest. In some cases, the small molecule may be generated by a molecule design computation model operating at the full molecule scale. In some cases, the small molecule may be generated by the molecule design computation model modifying a lead molecule exhibiting one or more properties of interest (e.g., drug-like properties). Alternatively, the molecule design computation model may generate the small molecule de novo (e.g., from scratch and without any lead molecules). In either instances, the molecule design computation model may be operating in a low data regime with an insufficient quantity of data at the full molecule scale to achieve adequate performance. Accordingly, in some cases, the patch design computation model may be applied to determine the identity of the functional group of interest at a patch level to leverage the greater availability of data at the patch level. For instance, in some cases, the patch design computation model may output, for the patch level identity of the functional group of interest, a probability distribution (e.g., categorical distribution) across the possible functional groups such as, for example, alcohol, amine, carbonyl, halide, aromatic ring, and / or the like.

[0094] FIG. 1 depicts a system diagram illustrating an example of a molecule design system 100, in accordance with some example embodiments. Referring to FIG. 1, the molecule design system 100 may include a molecule design engine 110, a client device 120, an analysisNAI-5002497275vl 35Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 engine 130, and one or more laboratory equipment 140. As shown in FIG. 1, the molecule design engine 110, the client device 120, the analysis engine 130, and the one or more laboratory equipment 140 may be communicatively coupled via a network 150. In some cases, the client device 120 may be a processor-based device including, for example, a workstation, a desktop computer, a laptop computer, a smartphone, a tablet computer, a wearable apparatus, and / or the like. The one or more laboratory equipment 140 may include any wet lab and dry lab equipment capable of performing in vitro measurements and / or in vivo characterizations. Examples of the one or more laboratory equipment 140 may include sequencers, mass spectrometers, centrifuges, and / or the like. The network 150 may be a wired network and / or a wireless network including, for example, a local area network (LAN), a virtual local area network (VLAN), a wide area network (WAN), a public land mobile network (PLMN), the Internet, and / or the like.

[0095] In some example embodiments, the molecule design engine 110 may include a molecule design computation model 113, a patch design computation model 115, and a design controller 117. In some cases, the molecule design computation model 113 may be trained to generate entire protein molecules, which may include determining the identity and / or conformation of every amino acid residue present in the full sequence of a protein molecule. By contrast, the patch design computation model 115 may be trained to operate at a patch level. As noted, in this context, the term “patch” may refer to a set of the amino acid residues present in a protein molecule. In some cases, a patch may include a set of amino acid residues located within a threshold distance (e.g., radius and / or the like) of a residue of interest, such as an amino acid residue located in a binding interface between the protein molecule and a target molecule. In some cases, the residue of interest may be located at a center of the patch while the other amino acid residues in the patch are those located within a threshold distance (e.g., radius and / or the like) ofNAI-5002497275vl 36Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 the residue of interest in the protein molecule. Alternatively and / or additionally, the patch may include, at its center, the residue of interest along with a threshold quantity of amino acid residues that are located closest to the residue of interest.

[0096] By operating at the patch level instead of the full length scale of protein molecules, the patch design computation model 115 may leverage the greater availability of data at the patch level. For example, more data be available for certain regions within a protein molecule, such as the paratope, the variable region (Fv), the antigen binding fragment (Fab), or the complementarity determining region (CDR) of an antibody, than for entire protein molecules. The relative abundance of data at the patch level means that the patch design computation model 115 may be trained to better capture the many nuanced relationships that exist between the identity and conformation of amino acid residues at the patch level. Accordingly, the identity (or type) and conformation (e.g., sidechain conformation) of one or more amino acid residues in the patches determined by the patch design computation model 115 are more likely to be consistent with those present in protein molecules that exhibit one or more properties of interest (e.g., drug-like properties). Contrastingly, the paucity of data at the length-scale of protein molecules means that the protein molecules generated by the molecule design computation model 113 may be less likely to exhibit the one or more properties of interest. As such, in some cases, the design controller 117 may validate and, in some cases, correct the identity and conformation of one or more amino acid residues in a protein molecule generated by the molecule design computation model 113 based on the identity (or type) and conformation (e.g., sidechain conformation) of the corresponding amino acid residues in the patches determined by the patch design computation model 115.

[0097] FIG. 2A depicts a flowchart illustrating an example of a process 200 for computational molecule design, in accordance with some example embodiments. Referring toNAI-5002497275vl 37Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1FIGS. 1-2, the process 200 may be performed by the molecule design engine 110 to generate one or more protein molecules for synthesis and testing (e.g., in vitro measurements, in vivo characterization, and / or the like) by the one or more laboratory equipment 140. In some cases, the one or more protein molecules generated by the molecule design engine 110 may undergo in silico analysis at the analysis engine 130 before a set of protein molecules whose properties (e.g., druglike properties) satisfy one or more thresholds are selected for synthesis and testing (e.g., in vitro measurements, in vivo characterization, and / or the like) by the one or more laboratory equipment 140. Due to the high cost of wet lab assessment, the process 200 may be performed to increase the likelihood that a protein molecule generated by the molecule design computation model 113 exhibits one or more properties of interest (e.g., drug-like properties). As described in more details below, the likelihood of that the protein molecule generated by the molecule design computation model 113 exhibits the one or more properties of interest may be increased (or maximized) by validating and, in some cases, correcting the identity (or type) and conformation (e.g., sidechain conformation) of one or more amino acid residues in the protein molecule based on the output of the patch design computation model 115.

[0098] At 202, a molecule design computation model may be applied to generate a protein molecule by at least determining an identity and conformation of a plurality of amino acid residues in the protein molecule. In some example embodiments, a molecule design computation model trained to operate at the full length scale of protein molecules may be applied to generate a protein molecule. In some cases, the molecule design computation model may generate the protein molecule by modifying another protein molecule (e.g., a lead molecule). Alternatively, the molecule design computation model may be applied to generate the protein molecule de novo (e.g., from scratch and without a lead molecule). In some cases, the molecule design computation modelNAI-5002497275vl 38Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 may generate the protein molecule by determining the identity and conformation of each amino acid residue in the protein molecule. For example, in some cases, the protein molecule may be an output protein molecule that is generated by the molecule design computation model denoising an input molecule by updating the identity and / or conformation of one or more amino acid residues in the input protein molecule. In some cases, the molecule design computation model may be trained based on sample protein molecules exhibiting one or more properties of interest (e.g., druglike properties). As such, in some cases, the identity (or type) and conformation (e.g., sidechain conformation) of the amino acid residues in the protein molecule generated by the molecule design computation model should be consistent with those present in protein molecules that exhibit the one or more properties of interest. However, due to the paucity of data specifying the identity and conformation of amino acid residues across the entire length of protein molecules, the likelihood that the protein molecule with the identity and conformation of amino acid residues determined by the molecule design computation model actually exhibits the one or more properties of interest (e.g., drug-like properties) may not be sufficiently high. In particular, where the protein molecule is generated to be a candidate molecule for synthesis and testing (e.g., in vitro measurements, in vivo characterization, and / or the like), the subpar performance of the molecule design computation model in generating the protein molecule may increase the risk that precious wet lab resources are squandered on a candidate molecule with low viability as a protein-based therapeutic. Contrastingly, as described in more details below, generating at least a portion of a protein molecule at a patch scale or, alternatively, performing patch scale validation of the protein molecule may alleviate at least some of the uncertainty that the protein molecule generated at the full-length scale fails to exhibit the one or more properties of interest (e.g., drug-like properties).NAI-5002497275vl 39Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0099] At 204, a patch design computation model may be applied to determine an identity and conformation of one or more amino acid residues in a patch that includes a set of amino acid residues from the protein molecule and a target molecule. In some example embodiments, a patch design computation model trained to operate at the patch level may be applied to determine the identity (or type) and conformation (e.g., sidechain conformation) of one or more amino acid residues in a patch that includes amino acid residues from the protein molecule generated in operation 202 and / or a target molecule (e.g., another protein molecule, a small molecule, and / or the like). For example, in some cases, the patch design computation model may be applied to operate on a patch from a complex formed by a binding interaction between the protein molecule and the target molecule. In some cases, the patch may include a set of amino acid residues from the protein molecule and / or the target molecule that are located within a threshold distance (e.g., radius and / or the like) of a residue of interest. Alternatively and / or additionally, the patch may include a threshold quantity of amino acid residues from the protein molecule and / or the target molecule that are located closest to the residue of interest. In some cases, the residue of interest may be an amino acid residue in the protein molecule generated by the molecule design computation model. For instance, in some cases, the aforementioned residue of interest may be an amino acid residue that is located in a binding interface between the protein molecule and the target molecule. Moreover, in some cases, the residue of interest may be at a center of the patch while the other amino acid residues in the patch are those amino acid residues in the protein molecule and / or the target molecule. In instances where the target molecule is a small molecule and not another target molecule, the patch may include one or more atoms in the small molecule.NAI-5002497275vl 40Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0100] It should be appreciated that the set of amino acid residues that are included in the patch operated upon by the patch design computation model may vary depending on the residue of interest. For example, for a first residue of interest in the protein molecule, the patch design computation model may operate on a first patch that includes a first set of amino acid residues. Alternatively, for a second residue of interest in the protein molecule, the patch design computation model may operate on a second patch that includes a second set of amino acid residues.

[0101] In some example embodiments, the patch design computation model may be trained to operate at the patch level in order to leverage the greater abundance of data at the patch level than at the full length scale of protein molecules. In particular, the greater abundance of data at the patch level may enable the patch design computation model to better capture the many nuanced relationships that exist between the identity and conformation of amino acid residues at the patch level than the molecule design computation model is able to capture the relationships that exist between the identity and conformation of amino acid residues at the full length scale of protein molecules. Those relationships, which reflect the physics that govern protein structure, stability, and association, may be paramount to generating protein molecules that are more likely to exhibit one or more properties of interest (e.g., drug-like properties). However, the paucity of such data for entire protein molecules means that the molecule design computation model is unable to capture these relationships at the full length scale of protein molecules as well as the patch design computation model is able to do at the patch scale. As such, in some cases, the identity (or type) and conformation (e.g., sidechain conformation) of the amino acid residues determined by the patch design computation model may be more likely to be consistent with those present in protein molecules that exhibit the one or more properties of interest (e.g., drug-like properties).NAI-5002497275vl 41Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0102] In some example embodiments, the patch design computation model may determine the identity (or type) and conformation (e.g., sidechain conformation) of a residue of interest in a patch from the protein molecule generated by the molecule design computation model. As noted, in some cases, the patch from the protein molecule may include the residue of interest, which may be located at an interface between the protein molecule and a target molecule (e.g., another protein molecule, a small molecule, and / or the like). In some cases, the patch may also include one or more other amino acid residues that are located within a threshold distance (e.g., radius and / or the like) of the residue of interest in the protein molecule. Alternatively and / or additionally, the patch may include a threshold quantity of amino acid residues that are located closest to the residue of interest. In some cases, the patch design computation model may determine the identity (or type) and conformation (e.g., sidechain conformation) of the residue of interest based on a description of the amino acid residues in the patch. As noted, those other amino acid residues may be in the protein molecule and / or the target molecule in cases where the target molecule is another protein molecule. For example, as described in more details below, the patch design computation model may determine, based at least a description of the amino acid residues included in an embedding of the patch, a probability distribution across the possible types of amino acid residues for the residue of interest. Furthermore, the patch design computation model may also determine, for at least one possible type of amino acid residue, a probability distribution across the possible rotamers and / or the constituent torsion angles of that type of amino acid residue. In instances where the target molecule is a small molecule and not another protein molecule, the embedding of the patch may include a description of one or more amino acid residues from the protein molecule and, in some cases, one or more of the atoms or functional groups forming the small molecule.NAI-5002497275vl 42Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0103] At 206, the identity and conformation of one or more amino acid residues in the protein molecule generated by the molecule design computation model may be validated based at least on an output of the patch design computation model. In some example embodiments, the identity of one or more amino acid residues in the protein molecule generated by the molecule design computation model, such as that of the residue of interest, may be validated based on the probability distributions included in the output of the patch design computation model. As noted, the molecule design computation model may lack sufficient data to generate protein molecules at the full molecule scale with adequate performance. For example, in some cases, the protein molecules generated by the molecule design computation model may include errors in the identity and / or conformation of one or more constituent amino acid residues. Such issues may be especially acute when the molecule design computation model is applied to generate, at the full molecule scale, protein molecules de novo. Accordingly, in some cases, the output of the patch design computation model may serve to validate at least a portion of the protein molecule generated by the molecule design computation model. For instance, as described in more details below, an identity and / or conformation of one or more amino acid residues in the protein molecule generated by the molecule design computation model may be corrected based on the output of the patch design computation model.

[0104] In some example embodiments, the output of the patch design computation model may include a probability distribution across the possible types of amino acid residues for the residue of interest in the protein molecule generated by the molecule design computation model. In some cases, this probability distribution may indicate that the residue of interest may be a different type of amino acid residue than what is identified by the molecule design computation model. For instance, the residue of interest may be a first type of amino acid residue in the proteinNAI-5002497275vl 43Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 molecule generated by the molecule design computation model but the output of the patch design computation model may indicate that the probability of the residue of interest being the first type of amino acid residue fails to satisfy one or more thresholds. Instead, the output of the patch design computation model may indicate that the probability of the residue of interest being a second type of amino acid residue is higher than the probability of the residue of interest being the first type of amino acid residue. Accordingly, in some cases, the identity of the amino acid residue in the protein molecule generated by the molecule design computation model may be corrected, for example, by replacing the first type of amino acid residue identified by the molecule design computation model with the second type of amino acid residue indicated by the output of the patch design computation model.

[0105] In some example embodiments, the conformation (e.g., sidechain conformation) of the amino acid residues in a patch of the protein molecule generated by the molecule design computation model (e.g., in operation 202) may also be validated based on the probability distributions included in the output of the patch design computation model. In instances where the identity of the residue of interest is changed from the first type of amino acid residue identified by the molecule design computation model to the second type of amino acid residue identified by the patch design computation model, the conformation (e.g., sidechain conformation) of the residue of interest may also be changed to one of the rotamers of the second type of amino acid residue. As noted, in some cases, the patch design computation model may output, for a given type of amino acid residue, a probability distribution across the possible rotamers of that type of amino acid residue or the possible values of the torsion angles present in the sidechain of that type of amino acid residue. Accordingly, in some cases, in addition to identifying the second type of amino acid residue as the residue of interest, the output of the patch design computation modelNAI-5002497275vl 44Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 may further indicate that the sidechain conformation of the residue of interest is more likely to be one rotamer of the second type of amino acid residue than another rotamer of the second type of amino acid residue. When that is the case, the validation of the protein molecule may further include replacing, for example, a rotamer of the first type of amino acid residue identified by the molecule design computation model as the identity (or type) and sidechain conformation of the residue of interest with the rotamer of the second type of amino acid residue identified by the patch design computation model. Alternatively, the output of the patch design computation model may indicate the different possible values of each torsion angle present in the sidechain of the residue of interest as well as the corresponding probabilities (e.g., continuous or discrete probabilities). In that case, the validation of the protein molecule may include adjusting one or more torsion angles with or without replacing the type of amino acid reside identified by the molecule design computation model as the identity of the residue of interest.

[0106] In some cases, the validation may include determining, based at least on the output of the patch design computation model, a probability (or likelihood) of the identity and conformation of one or more amino acid residues in the protein molecule generated by the molecule design computation model. For example, in some cases, the molecule design computation model may determine the overall configuration of the protein molecule by determining the conformation of every individual amino acid residue in the protein molecule. In some cases, the molecule design computation model may predict the identity as well as the conformation of each amino acid residue. The probability (or likelihood) that the protein molecule assumes the conformation predicted by the molecule design computation model may be evaluated based at least on the output of the patch design computation model. For instance, in some cases, the probability (or likelihood) of the conformation predicted by the molecule design computationNAI-5002497275vl 45Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 model may be determined based on the probability of the identity and conformation of one or more constituent amino acid residues determined by the patch design computation model.

[0107] In some cases, it is also possible for the molecule design computation model to identify a correct type of amino acid residue for the residue of interest but an incorrect rotamer of the type of amino acid residue as the conformation of residue of interest. For example, in some cases, the sidechain conformation of the residue of interest in the protein molecule generated by the molecule design computation model may be a rotamer of an amino acid residue but the output of the patch design computation model may indicate that the sidechain conformation of the residue of interest is more likely to be another rotamer of the same type of amino acid residue. Accordingly in some cases, the sidechain conformation of the amino acid residue in the protein molecule generated by the molecule design computation model may be corrected, for example, by being replaced with the other rotamer of the amino acid residue indicated by the output of the patch design computation model. For instance, in some cases, the protein molecule generated by the molecule design computation model may be corrected by replacing the rotamer of the amino acid residue identified by the molecule design computation model as the sidechain conformation of the residue of interest with the other rotamer of the amino acid residue identified by the patch design computation model.

[0108] FIG. 3 depicts a flowchart illustrating another example of a process 300 for computational molecule design, in accordance with some example embodiments. Referring to FIGS. 1-3, the process 300 may be performed by the molecule design engine 110 to generate a patch that includes a set of amino acid residues in a protein molecule and / or a target molecule (e.g., another protein molecule, a small molecule, and / or the like). In some cases, the process 300 may implement operation 204 of the process 200 shown in FIG. 2. As described in more details below,NAI-5002497275vl 46Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 in some cases, the process 300 may include applying a patch molecule design computation model trained to operate on an embedding of the patch that includes a description of each constituent amino acid residue. For example, in some cases, the patch design computation model may determine, based at least on the description of the amino acid residues included in the embedding of the patch, the identity and conformation of a residue of interest, such as an amino acid residue from the protein molecule that is located in an interface between the protein molecule and the target molecule. In instances where the target molecule is a small molecule and not another protein molecule, the embedding of the patch may include a description of one or more amino acid residues from the protein molecule and one or more atoms from the target molecule.

[0109] At 302, a residue of interest may be selected for validation from a plurality of amino acid residues forming a protein molecule. In some example embodiments, the generation of an input patch for ingestion by a patch design computation model may include selecting one or more residues of interest from the amino acid residues in a protein molecule. For example, in some cases, a residue of interest may be an amino acid residue in the protein molecule that is located at an interface between the protein molecule and a target molecule such as, for example, another protein molecule, a small molecule, and / or the like. In instances where the interface is between an antibody and an antigen, the residue of interest may be an amino acid residue located in a V-H interface in an antigen-binding fragment (Fab) and antigen complex, an antigen-binding fragment (Fab) and antigen interface, a T-cell receptor (TCR) peptide interface in a T-cell receptor and peptide-bound major histocompatibility complex (TCR-pMHC), and / or the like. In some cases, the residue identity and the conformation of the one or more residues of interest may already be determined by a molecule design computation model trained to operate at the full length scale of protein molecules. As described in more details below, the one or more residues of interest mayNAI-5002497275vl 47Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 be selected for validation, which may include a validation of the residue identity and, in some cases, also the conformation determined by the molecule design computation model.

[0110] At 304, an input patch may be generated to include the residue of interest and one or more additional amino acid residues. In some cases, the generation of the input patch may further include identifying one or more additional amino acid residues from the protein molecule and / or the target molecule. In some cases, the one or more additional amino acid residues may include those that are located within a threshold distance of the one or more amino acid residues of interest. Alternatively and / or additionally, the one or more additonal amino acid residues may include a threshold quantity of amino acid residues that are closest in space to the one or more amino acid residues. The input patch may include the resulting set of amino acid residues, which may include some but not all of the amino acid residues in the protein molecule and the target molecule. It should be appreciated that this distance may refer to the distance between two amino acid residues in three-dimensional space and not the distance between the two amino acid residues in the linear sequence of amino acid residues. Thus, in cases where two amino acid residues are distantly located in the linear sequence of amino acid residues, for example, by being separated by a large quantity of intervening amino acid residues, the two amino acid residues may still be proximately located in three-dimensional space due to the folding the polypeptide chains in the protein molecule. Moreover, it should be appreciated that in instances where the target molecule is a small molecule and not another protein molecule, the input patch may include, in addition to a set of amino acid residues from the protein molecule that includes the one or more residues of interest, a threshold quantity of atoms from the small molecule that are located closest to the one or more residues of interest in the protein molecule or any atom that is located within the threshold distance of the one or more residues of interest in the protein molecule. As described in moreNAI-5002497275vl 48Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 details below, an embedding of the input patch may be generated to include a description of the set of amino acid residues from the protein molecule and / or the target molecule.[OHl] At 306, an embedding of the input patch may be generated. In some example embodiments, the embedding of the input patch may be generated to a description of the set of amino acid residues identified for inclusion in the input patch. As noted, in some cases, the set of amino acid residues identified for inclusion in the input patch may include the residue of interest in the protein molecule and one or more additional amino acid residues. Those additional amino acid residues may include a threshold quantity of amino acid residues that are closest to the residue of interest or any amino acid residue that is located within a threshold distance of the residue of interest. In some cases, an encoder may be applied to generate an embedding of the input patch. For example, in some cases, the encoder may be a machine learning model, such as a graph neural network (GNN) and / or the like, that has been trained to generate the embedding of the input patch to reflect the atomic context around the residue of interest which, as noted, may be at a center of the input patch. As such, in some cases, the encoder may be applied to generate an embedding of the individual atoms present in the input patch, which are then aggregated to form embeddings of the corresponding amino acid residues and / or the embedding of the input patch as a whole.

[0112] In some cases, the embedding of the input patch may include a description of the identity (or type) of one or more amino acid residues present in the set of amino acid residues. For example, in some cases, the embedding of the input patch may include, for each amino acid residue, a one-hot vector indicating the identity (or type) of the amino acid residue. In instances where each amino acid residue can be one of M possible amino acid residues (e.g., 20 canonical amino acid residues), each one-hot vector may include an M quantity of binary values (e.g., 0 or 1). Furthermore, in each one-hot vector, one of the aforementioned M binary values may be set toNAI-5002497275vl 49Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 a first value (e.g., 1) while the remaining M — 1 binary values may be set to a second value (e.g., 0) in order to indicate identity (or type) of amino acid residue. It should be appreciated that one- hot encoding is just one example technique for generating the embedding of the input patch. In some cases, instead of one-hot encoding, a language model may be applied to generate the embedding of the input patch including, in some cases, the embedding of the identity of each amino acid residue in the patch.

[0113] In some cases, the embedding of the input patch may also include a description of the sequential position and / or the spatial position of each amino acid residue in the set of amino acid residues. It should be appreciated that while the sequential position of an amino acid residue refers to its position relative to other amino acid residues in the linear sequence of amino acid residues in the protein molecule or the target molecule whereas the spatial position of the amino acid residue refers to its position relative to other amino acid residues in three-dimensional space. For example, in some cases, the spatial position of an amino acid residue may be described in terms of the three-dimensional coordinates (e.g., (x,y, z) coordinates) of one or more constituent atoms. In some cases, the spatial position of an amino acid residue may be encoded as a finegrained representation or a coarse-grained representation of the spatial position of an amino acid residue. In some cases, a fine-grained representation of the spatial position of amino acid residue may include the three-dimensional coordinates (e.g., (x,y, z) coordinates) of a larger proportion of the atoms in the amino acid residue than a coarse-grained representation of the spatial position of the amino acid residue. For instance, in some cases, a fine-grained representation of the spatial position of an amino acid residue may include the three-dimensional coordinates (e.g., (x, y, z) coordinates) of every backbone atom present in the amino acid residue whereas a coarse-grainedNAI-5002497275vl 50Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 representation of the spatial position of the same amino acid residue may include only the three- dimensional coordinates (e.g., (x,y, z) coordinates) of the alpha carbon (Ca) atom.

[0114] As noted, in some cases, the input patch may include amino acid residues from the protein molecule and / or the target molecule. In particular, in some cases, the input patch may include a residue of interest, which is an amino acid residue in the protein molecule that is located at an interface between the protein molecule and the target molecule. In instances where the target molecule is another protein molecule, it is possible for the embedding of the input patch to be generated to include, in addition to a description of the residue of interest in the protein molecule, a description of the atoms in one or more additional amino acid residues from the protein molecule as well as the target molecule. Where the target molecule is a small molecule and not another protein molecule, the description may be of atoms in the small molecule and not of additional amino acid residues. In some cases, as an alternative to an embedding of the input molecule that includes a description of atoms from the protein molecule and the target molecule, it is also possible for the embedding of the input patch to be generated to include a description of the atoms in one or more amino acid residues from the protein molecule but not of the atoms either from a small molecule that is the target molecule or from one or more amino acid residues in the target molecule. With this variation of the input patch embedding, the patch design computation model may be conditioned on the atomic context of the target molecule and not the atomic context of the binding interface as a whole.

[0115] At 308, a patch molecule design computation model may be applied to determine, based at least on the embedding of the input patch, a patch level identity and / or a patch level conformation of the residue of interest in the protein molecule. In some example embodiments, the patch design computation model may be a machine learning model that is trained to determine,NAI-5002497275vl 51Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 based at least on the atomic context of a residue of interest from a protein molecule, the identity and conformation of the residue of interest. For example, in some cases, the patch design computation model may ingest, as input, the embedding of the input patch and operate on the embedding of the input patch to generate an output that indicates the identity (or type) and conformation (e.g., sidechain conformation) of the residue of interest as one or more probability distributions (e.g., of the possible types of amino acid residue, rotamers, torsion angles, and / or the like). The embedding of the input patch may include a description of the atomic context of the residue of interest including, for example, the identity and conformation of the threshold quantity of amino acid residues that are closest to the residue of interest or that of any amino acid residue located within a threshold distance of the residue of interest. In instances where there are an M quantity of possible amino acid residues, the output of the patch design computation model may include a probability distribution in the form of a vector containing M values, each of which being a probability that the residue of interest is a corresponding one of the M quantity of possible amino acid residues. As noted, each type of amino acid residue may be associated with a different set of rotamers. In some cases, two or more amino acid residues may be associated with different quantities of rotamers. Accordingly, to further represent the conformation (e.g., sidechain conformation) of the residue of interest given that the identity of the residue of interest is a type of amino acid residue having an N quantity of possible rotamers, the output of the patch design computation model may include a probability distribution in the form of a vector containing an N quantity of values, each of which being a probability that the sidechain conformation of the residue of interest is a corresponding one of the N quantity of rotamers. Alternatively, instead of the N quantity of possible rotamers, the conformation (e.g., sidechain conformation) of the residue of interest given that its identity (or type of amino acid residue) may be defined based on the possibleNAI-5002497275vl 52Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 values of each torsion angle present in the residue of interest. As noted, in instances where there are discrepancies between the identity and / or conformation of the residue of interest in the protein molecule generated by the molecule design computation model, which had been trained to operate at the full length scale of protein molecules, and the probability distributions output by the patch design computation model, the identity and / or conformation of the residue of interest in the protein molecule generated by the molecule design computation model may be corrected in accordance with the output of the patch design computation model.

[0116] FIG. 4 depicts a flowchart illustrating another example of a process 400 for computational molecule design, in accordance with some example embodiments. Referring to FIGS. 1-4, the process 400 may be performed by the molecule design engine 110 to train a patch design computation model to determine, based at least on a patch that includes a set of amino acid residues in a protein molecule and / or a target molecule, the identity and conformation of one or more residues of interest in the protein molecule. In particular, in some cases, the patch design computation model may be trained to operate on embeddings of sample patches that includes a description of the atomic context surrounding the one or more residues of interest in each sample patch. As noted, the patch design computation model may be trained to operate at the patch level instead of the full length scale of protein molecules in order to leverage the greater availability of data at the patch level. That is, in some cases, the patch design computation model trained in accordance with the process 400 may better capture the relationships that exist between the identity and conformation of amino acid residues at the patch level. For example, a protein molecule in which the residue of interest exhibits the identity (or type) and conformation (e.g., sidechain conformation) determined by the patch design computation model may be more likely to exhibit one or more desirable properties, such as binding affinity towards a target molecule (e.g., anotherNAI-5002497275vl 53Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 protein molecule, a small molecule, and / or the like), than the same protein molecule in which the identity and conformation of the residue of interest is determined by the molecule design computation. As such, once trained, the patch design computation model may be deployed to perform operation 204 of the process 200 shown in FIG. 2 and / or operation 304 of the process 300 shown in FIG. 3.

[0117] At 402, a sample patch including one or more amino acid residues may be extracted from a complex formed by an interaction between a sample protein molecule and a sample target molecule. In some example embodiments, a sample patch for training a patch design computation model may be extracted from a protein molecule that exhibits one or more properties of interest including, in some cases, a binding affinity towards a target molecule. As such, in some cases, the sample patch may be extracted from a complex that is formed by the protein molecule bound to the target molecule, in which case the conformation of the protein molecule complements that of the target molecule. In some cases, the sample patch may be generated to include a residue of interest from a protein molecule as well as one or more additional amino acid residues from the protein molecule and / or the target molecule. For example, in some cases, the residue of interest from the protein molecule may be an amino acid residue that is located at an interface between the protein molecule and the target molecule. Furthermore, the one or more additional amino acid residues may include amino acid residues from both the protein molecule and the target molecule or, in some cases, only amino acid residues from the protein molecule but not amino acid residues from the target molecule. In instances where the target molecule is a small molecule and not another protein molecule, the sample patch may include one or more of the atoms in the small molecule. That the residue of interest is located at the interface between the protein molecule and the target molecule means that the sample patch may include at least a portion of the interfaceNAI-5002497275vl 54Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 between the protein molecule and the target molecule. As such, the sample patch may capture at least a portion of the atomic context at the interface between the protein molecule and the target molecule such that the patch design computation model may be trained to learn the atomic context that exists in the interface between two molecules with sufficient binding affinity.

[0118] At 404, an embedding of the sample patch may be generated to include a description of the one or more amino acid residues. In some example embodiments, the sample patch may be encoded to generate an embedding that includes a description of the atomic context captured in the sample patch. In some cases, a description of the atomic context captured in the sample patch may include an encoding of at least one of the identity, sequential position, and spatial position of each amino acid residue in the sample patch. As noted, in some cases, the description of the identity of an amino acid residue may include a one-hot vector including an M quantity of binary values (e.g., 0 or 1) in which the type of amino acid residue is identified by setting one of the M quantity of binary value to a first value (e.g., 1) and the remaining M quantity of binary values to a second value (e.g., 0). In some cases, the description of the sequential position of the amino acid residue may be a positional encoding that indicates the position of the amino acid residue relative to those of the other amino acid residues in the same sequence of amino acid residues. The description of the spatial position of the amino acid residue may include one or more three-dimensional coordinates (e.g., (x,y, z) coordinates) to indicate the position of one or more of the atoms forming the amino acid residue. An embedding of the sample patch that includes a fine-grained representation of the spatial position of the amino acid residue may include the three- dimensional coordinates (e.g., (x,y, z) coordinates) of a greater proportion of atoms in the amino acid residue than an embedding of the sample patch with a coarse-grained representation of the spatial position of the amino acid residue.NAI-5002497275vl 55Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0119] At 406, a training sample in which an identity and conformation of one or more amino acid residues in the embedding of the sample patch are masked may be generated based on the embedding of the sample patch. In some example embodiments, a training sample may be generated by masking the identity and conformation of at least one amino acid residue in the sample patch, such as the residue of interest located at the center of the sample patch. For example, in some cases, the identity and conformation of the residue of interest, which may be located at a center of the sample patch, may be replaced with a special mask token. As described in more details below, the training of the patch design computation model may include training the patch design computation model to recover, based on the atomic context of other amino acid residues in the sample patch, the identity and conformation of the residue of interest.

[0120] At 408, a patch design computation model may be trained, based at least on the sample patch, to recover the identity and conformation of the one or more masked amino acid residues in the embedding of the sample patch. In some example embodiments, the patch design computation model may be trained to capture the relationships that exist between the identity and conformation of amino acid residues at the patch level. In particular, the patch design computation model may be trained to capture the relationships between the identity and conformation of the residue of interest, which is masked in each sample patch, and the atomic context of the other amino acid residues included in the sample patch. As noted, in some cases, the output of the patch design computation model may include a probability distribution across the possible types of amino acid residues (e.g., 20 canonical amino acid residues) and a probability distribution (e.g., categorical distribution) across the possible rotamers of each type of amino acid residue. In this context, it should be appreciated that the identity (or type) and conformation (e.g., sidechain conformation) of the masked residue of interest may be considered the ground-truth annotation (orNAI-5002497275vl 56Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 label) of the training sample. Accordingly, in some cases, the training of the patch design computation model may include adjusting one or more parameters (e.g., weights, biases, and / or the like) of the patch design computation model to increase (or maximize), in the probability distribution output by the patch design computation model, the probability of the ground truth type of amino acid residue of the masked residue of interest relative to the probabilities of the other possible types of amino acid residues. Furthermore, in some cases, the training of the patch design computation model may include adjusting one or more parameters (e.g., weights, biases, and / or the like) of the patch design computation model to increase (or maximize), in the probability distribution output by the patch design computation model for a particular type of amino acid residue (e.g., the ground truth type of amino acid residue of the masked residue of interest), the probability of the ground truth rotamer of the masked residue of interest relative to the probabilities of the other possible rotamers of the masked residue of interest.

[0121] In some cases, the output of the patch design computation model may include a probability distribution across the possible values of the torsion angles present in each possible type of amino acid residue. For example, in some cases, for each torsion angle present in the sidechain of an amino acid residue, the output of the patch design computation model may include a continuous distribution across the possible values of the torsion angle (e.g., between —180° and + 180°). Alternatively, the output of the patch design computation model may be a categorical distribution across a set of discrete angular values, such as —180°, —90°, +90°, and +180°. In this case, the ground-truth annotations of the training sample may include probability distributions in which the actual values of the torsion angles present in the sidechain of the training sample are assigned a higher (or maximum) probability than the other possible values. In some cases, the training of the patch design computation model may include adjusting one or more parametersNAI-5002497275vl 57Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1(e g., weights, biases, and / or the like) of the patch design computation model to increase (or maximize), in the probability distribution output by the patch design computation model, the probability of the ground truth torsion angle value.

[0122] FIG. 5A depicts a flowchart illustrating an example of a process 500 for computational small molecule design, in accordance with some example embodiments. Referring to FIGS. 1 and 5A, the process 500 may be performed by the molecule design engine 110 to generate one or more small molecules (e.g., drug compounds), which may then undergo synthesis and testing (e.g., in vitro measurements, in vivo characterization, and / or the like) by the one or more laboratory equipment 140. In some cases, the one or more small molecules generated by the molecule design engine 110 may undergo in silico analysis at the analysis engine 130 before a set of small molecules whose properties (e.g., drug-like properties) satisfy one or more thresholds are selected for synthesis and testing (e.g., in vitro measurements, in vivo characterization, and / or the like) by the one or more laboratory equipment 140. Due to the high cost of wet lab assessment, the process 500 may be performed to increase the likelihood that a small molecule (e.g., drug molecule) generated by the molecule design computation model 113 exhibits one or more properties of interest (e.g., drug-like properties). Due to operating in a low data regime, the performance of the molecule design computation model 113 in generating small molecules at the full molecule scale may be inadequate. Accordingly, as described in more details below, the likelihood of that the small molecule (e.g., drug molecule) generated by the molecule design computation model 113 exhibits the one or more properties of interest (e.g., drug-like properties) may be increased (or maximized) by validating and, in some cases, correcting the chemical composition and / or conformation (e.g., functional groups) in the small molecule generated by theNAI-5002497275vl 58Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 molecule design computation model 113 based on the output of the patch design computation model 115.

[0123] At 502, a molecule design computation model may be applied to generate a small molecule by at least determining an identity of one or more functional groups in the small molecule. In some example embodiments, a molecule design computation model trained to operate at the full molecule scale of small molecules (e g., drug molecules) may be applied to generate a small molecule including one or more functional groups attached to a scaffold (or core structure). In some cases, the molecule design computation model may generate the small molecule by at least modifying another molecule (e.g., a lead molecule) exhibiting one or more properties of interest (e.g., drug-like properties). Alternatively, the molecule design computation model may generate the small molecule de novo, meaning that the small molecule is generated from scratch instead of by modifying another molecule (e g., a lead molecule). In some cases, the molecule design computation model may generate the small molecule by determining the identity of the one or more functional groups or, alternatively, one or more of the constituent atoms in the small molecule. In some cases, the molecule design computation model may be trained based on sample small molecules exhibiting one or more properties of interest (e.g., drug-like properties). As such, in some cases, the identity (or type) of the atoms and / or functional groups in the small molecule generated by the molecule design computation model should be consistent with those present in small molecules exhibiting the one or more properties of interest. However, due to an insufficient quantity of data at the full molecule scale of small molecules exhibiting one or more properties of interest, the likelihood that the small molecule with the types of atoms and / or functional groups determined by the molecule design computation model actually exhibits the one or more properties of interest (e.g., drug-like properties) may not be adequate. In particular, whereNAI-5002497275vl 59Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 the small molecule is generated to be a candidate molecule for synthesis and testing (e.g., in vitro measurements, in vivo characterization, and / or the like), the subpar performance of the molecule design computation model in generating the small molecule at the full molecule scale may increase the risk that precious wet lab resources are squandered on a candidate molecule with low viability as a therapeutic. Contrastingly, as described in more details below, generating at least a portion of a small molecule at a patch scale or, alternatively, performing patch scale validation of the small molecule may increase the likelihood of the small molecule generated at the full molecule scale exhibits the one or more properties of interest (e.g., drug-like properties).

[0124] At 504, a patch design computation model may be applied to determine an identity of a functional group of interest in a patch that includes the functional group and at least a portion of one or more adjacent functional groups in the small molecule. In some example embodiments, a patch design computation model trained to operate at the patch level may be applied to determine the identity (or type) of at least a portion of one or more functional groups in the small molecule (e.g., drug molecule) generated at the full molecule scale by the molecule design computation model. For small molecules, the patch design computation model may operate on a patch that includes a functional group of interest. In some cases, the functional group of interest may be a functional group participating in an interaction, such as binding or modulation, with a target molecule (e.g., a protein molecule and / or the like). In some cases, the patch may further include one or more atoms, functional groups, and / or amino acid residues forming an atomic context around the functional group of interest. For example, in some cases, the patch may include one or more atoms and / or functional groups located within a threshold distance (e.g., radius and / or the like) of the functional group of interest. In some cases, the functional group of interest may be located at an interface between the small molecule and a target molecule. In instancesNAI-5002497275vl 60Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 where the target molecule is a protein molecule, the patch may also include one or more amino acid residues in the protein molecule that are located at the interface between the small molecule and the protein molecule.

[0125] In some example embodiments, the patch design computation model may be trained to operate at the patch level in order to leverage the greater abundance of data at the patch level than at the full molecule scale of small molecules (e.g., drug molecules). In particular, the greater abundance of data at the patch level may enable the patch design computation model to better capture the many nuanced relationships that exist between functional groups at the patch level than the molecule design computation model is able to at the full molecule scale. Those relationships, which reflect the physics that govern protein structure, stability, and association, may be paramount to generating small molecules (e.g., drug molecules) that are more likely to exhibit one or more properties of interest (e.g., drug-like properties). However, the paucity of such data for entire small molecules means that the molecule design computation model is unable to capture these relationships at the full molecule scale as well as the patch design computation model is able to do at the patch scale. As such, in some cases, the type of the functional groups determined by the patch design computation model may be more likely to be consistent with those present in small molecules that exhibit the one or more properties of interest (e.g., drug-like properties).

[0126] In some example embodiments, the patch design computation model may determine the identity (or type) of a functional group of interest in a patch from the small molecule generated by the molecule design computation model at a full molecule scale. As noted, in some cases, the functional group of interest may participate in an interaction with the target molecule (e.g., protein molecule) and may therefore be located at an interface between the small molecule and the target molecule. In some cases, the patch may also include one or more atoms and / or otherNAI-5002497275vl 61Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 functional groups that are located within a threshold distance (e.g., radius and / or the like) of the functional group of interest, thus providing an atomic context for the functional group of interest. In some cases, the atomic context of the functional group of interest may also include a portion of the target molecule, such as one or more amino acid residues in instances where the target molecule is a protein molecule, that is located at the interface between the small molecule and the target molecule. In some cases, the patch design computation model may determine the type of the functional group of interest based at least on an embedding of the patch including the functional group of interest. In some cases, the embedding of the patch may include a description of the atomic context of the functional group of interest, which may include one or more atoms, other functional groups, amino acid residues in the target molecule, and / or the like. In some cases, the patch design computation model may determine the type of the functional group of interest by at least determining a probability distribution (e.g., categorical distribution) across the possible types of functional groups. Examples of different types of functional groups may include alcohol, amine, carbonyl, halide, aromatic ring, and / or the like. In some cases, the type of the functional group of interest may be determined by at least sampling from the probability distribution. For instance, in some cases, the type of the functional group of interest may correspond to the type of functional group associated with a threshold (or highest) probability in the probability distribution (e.g., categorical distribution) output by the patch design computation model.

[0127] At 506, the identity of one or more functional groups in the small molecule generated by the molecule design computation model may be validated based at least on an output of the patch design computation model. In some example embodiments, the type of one or more functional groups in the small molecule (e.g., drug molecule) generated by the molecule design computation model, such as that of the functional group of interest, may be validated based on theNAI-5002497275vl 62Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 probability distributions (e.g., categorical distribution) included in the output of the patch design computation model. As noted, the molecule design computation model may lack sufficient data to generate small molecules at the full molecule scale with adequate performance. For example, in some cases, the small molecules generated by the molecule design computation model may include errors in the type of one or more functional groups. Such errors may be especially prevalent when the molecule design computation model is applied to generate, at the full molecule scale, small molecules de novo. Accordingly, in some cases, at least a portion of the small molecule generated by the molecule design computation model operating at the full molecule scale may be validated based on the output of the patch design computation model. For instance, as described in more details below, a type of one or more functional groups in the small molecule generated by the molecule design computation model may be corrected based on the output of the patch design computation model.

[0128] In some example embodiments, the output of the patch design computation model may include a probability distribution (e.g., categorical distribution) across the possible types of functional groups for the functional group of interest in the small molecule (e.g., drug molecule) generated by the molecule design computation model. In some cases, this probability distribution may indicate that the functional group of interest may be a different type of functional group than what is included in the small molecule generated by the molecule design computation model operating at the full molecule scale. For instance, the functional group of interest may be a first type of functional group in the small molecule generated by the molecule design computation model but the output of the patch design computation model may indicate that the probability of the functional group of interest being the first type of functional group fails to satisfy one or more thresholds. Instead, the output of the patch design computation model may indicate that theNAI-5002497275vl 63Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 probability of the functional group of interest being a second type of functional group is higher than the probability of the functional group of interest being the first type of functional group. Accordingly, in some cases, the type of the functional group of interest in the small molecule (e.g., drug molecule) generated by the molecule design computation model may be corrected, for example, by replacing the first type of functional group identified by the molecule design computation model operating at the full molecule scale with the second type of functional group indicated by the output of the patch design computation model operating at the patch scale.

[0129] FIG. 5B depicts a flowchart illustrating another example of a process 550 for computational small molecule design, in accordance with some example embodiments. Referring to FIGS. 1 and 5A-5B, the process 550 may be performed by the molecule design engine 110 to train a patch design computation model to determine, based at least on a patch that includes a functional group of interest (e.g., a functional group participating in an interaction with a target molecule) and at least a portion of one or more adjacent functional groups in a small molecule (e.g., drug compound), the patch level identity of the functional group of interest. In some cases, the patch design computation model may be trained to operate on embeddings of sample patches that includes a description of the atomic context surrounding the functional group of interest in each sample patch. As noted, the patch design computation model may be trained to operate at the patch level instead of the full molecule scale in order to leverage the greater availability of data at the patch level. That is, in some cases, the patch design computation model trained in accordance with the process 550 may better capture the relationships that exist between different functional groups and the properties of the corresponding small molecules at the patch level. For example, a small molecule (e g., drug compound) in which the identity of the functional group of interest is determined by the patch design computation model may be more likely to exhibit oneNAI-5002497275vl 64Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 or more desirable properties (e.g., drug-like properties) than the same small molecule generated by a molecule design computation trained to operate at the full molecule scale. As such, once trained, the patch design computation model may be deployed to perform operation 504 of the process 500 shown in FIG. 5A.

[0130] At 552, a sample patch including a functional group of interest from a sample small molecule may be extracted from a complex formed by a binding interaction between the sample small molecule and a sample target molecule. In some example embodiments, a sample patch for training a patch design computation model may be extracted from a small molecule (e.g., drug molecule) that exhibits one or more properties of interest including, in some cases, a binding affinity or biological activity towards a target molecule (e.g., a protein molecule). For example, in some cases, the sample patch may be extracted from a complex that is formed by the small molecule (e.g., drug molecule) bound to the target molecule. In some cases, the sample patch may be generated to include a functional group of interest from the small molecule. In some cases, the functional group of interest may be a functional group participating in an interaction (e.g., binding, modulation, and / or the like) with the target molecule. In some cases, the sample patch may be generated to include at least a portion of one or more additional functional groups from the small molecule. For instance, in some cases, the one or more additional functional groups may be adjacent functional groups attached to a core structure (or scaffold) of the small molecule. In some cases, it may be possible for the sample patch to be generated to also include one or more atoms (or amino acid residues) in the target molecule that occupy an interface between the small molecule and the target molecule. By including atoms, functional groups, and / or amino acid residues located at the interface between the small molecule and the target molecule, the sample patch may be generated to capture at least a portion of the atomic context at the interface between the smallNAI-5002497275vl 65Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 molecule and the target molecule such that the patch design computation model may be trained to learn the atomic context that exists in the interface between two molecules with sufficient binding affinity.

[0131] At 554, an embedding of the sample patch may be generated to include a description of the functional group. In some example embodiments, the sample patch may be encoded to generate an embedding that includes a description of the atomic context captured in the sample patch. In some cases, a description of the atomic context captured in the sample patch may include an encoding of at least one of the identity, sequential position, and spatial position of each functional group (or constituent atoms) in the sample patch. In some cases, the description of the identity of a functional group may include a one-hot vector including an M quantity of binary values (e.g., 0 or 1) in which the type of functional group is identified by setting one of the M quantity of binary value to a first value (e.g., 1) and the remaining M quantity of binary values to a second value (e.g., 0). In some cases, the description of the sequential position of the functional group may be a positional encoding (e.g., a numerical value) that specifies the carbon atom in the core structure (or scaffold) of the small molecule to which it is attached. The description of the spatial position of the functional group may include one or more three-dimensional coordinates (e.g., (x, y, z) coordinates) to indicate the position of one or more of constituent atoms (e.g., heavy atoms).

[0132] At 556, a training sample in which an identity of the functional group of interest is masked may be generated based at least on the embedding of the sample patch. In some example embodiments, a training sample may be generated by masking the identity of at least one functional group in the sample patch, such as that of the functional group of interest. For example, in some cases, the identity of the functional group of interest may be replaced with a special mask token.NAI-5002497275vl 66Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1As described in more details below, the training of the patch design computation model may include training the patch design computation model to recover, based on the atomic context of functional groups in the sample patch, the identity of the masked functional group of interest.

[0133] At 558, a patch design computation model may be trained, based at least on the training sample, to recover the identity of the masked functional group in the embedding of the sample patch. In some example embodiments, the patch design computation model may be trained to capture the relationships that exist between adjacent functional groups at the patch level. In particular, the patch design computation model may be trained to capture the relationships between the functional group of interest, which is masked in each sample patch, and the atomic context of the other atoms, functional groups, and / or amino acid residues included in the sample patch. In some cases, the output of the patch design computation model may include a probability distribution (e.g., categorical distribution) across the possible types of functional groups (e.g., alcohol, amine, carbonyl, halide, aromatic ring, and / or the like). In this context, it should be appreciated that the type of the functional group of interest may be considered the ground-truth annotation (or label) of the training sample. Accordingly, in some cases, the training of the patch design computation model may include adjusting one or more parameters (e.g., weights, biases, and / or the like) of the patch design computation model to increase (or maximize), in the probability distribution output by the patch design computation model, the probability of the ground truth type of functional group of the masked functional group of interest relative to the probabilities of the other possible types of functional groups.

[0134] FIG. 6A depicts a schematic diagram illustrating an example of a process 600 for computational molecule design, in accordance with some example embodiments. Referring to FIGS. 1-2 and 6A, the process 600 shown in FIG. 6A may correspond to the process 200 shownNAI-5002497275vl 67Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 in FIG. 2 in which a protein molecule generated by a molecule design computation model trained to operate at the full length scale of protein molecules is validated based on the output of a patch design computation model trained to operate at the patch level.

[0135] As shown in FIG. 6A, in some example embodiments, the patch design computation model 115 may be applied to determine, based at least on an embedding of an input patch 610, an identity and conformation of one or more amino acid residues in the input patch 610. For example, in some cases, the patch design computation model 115 may operate on the embedding of the input patch 610 in order to determine the identity and conformation of a residue of interest included in the input patch 610. In some cases, the residue of interest may be an amino acid residue from a protein molecule. In some cases, the residue of interest may be an amino acid residue located in an interface between the protein molecule and a target molecule, which may be another protein molecule, a small molecule, and / or the like. In some cases, the patch design computation model 115 may operate on the embedding of the input patch 610, which includes a description of one or more other amino acid residues in the protein molecule and / or the target molecule, to determine the identity and conformation of the residue of interest. In some cases, the one or more other amino acid residues may include a threshold quantity of amino acid residues from either the protein molecule or the target molecule that are closest to the residue of interest or any amino acid residue. Alternatively and / or additionally, in some cases, the one or more amino acid residues may include any amino acid residue from either the protein molecule and the target molecule that are located within the threshold distance (e.g., radius and / or the like) of the residue of interest. Alternatively, in some cases, the one or more amino acid residues may include amino acid residues from the protein molecule but not amino acid residues from the target molecule, even if those amino acid residues are within sufficient proximity of the residue of interest. In instancesNAI-5002497275vl 68Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 where the target molecule is a small molecule and not another protein molecule, the embedding of the input patch 610 may include one or more of the atoms forming the small molecule and not one or more amino acid residues from the target molecule.

[0136] In some example embodiments, the embedding of the input patch 610 may provide a description of the atomic context associated with the protein molecule and / or the target molecule. For example, in instances where the residue of interest is located at the interface between the protein molecule and the target molecule, the embedding of the input patch 610 may include a description of the atomic context present in at least a portion of the interface between the protein molecule and the target molecule. In some cases, the patch design computation model 115 may ingest the embedding of the input patch 610 and determine, based at least on the atomic context described by the embedding of the input patch 610, the identity and conformation of the residue of interest in the input patch 610. As shown in FIG. 6A, in some cases, the patch design computation model 115 may generate an output 615 that indicates the identity and conformation of the residue of interest in the input patch 610 as one or more probability distributions. For instance, in some cases, the output 615 of the patch design computation model 115 may include a probability distribution across the possible types of amino acid residues (e.g., 20 canonical amino acid residues). Furthermore, in some cases, the output 615 of the patch design computation model 115 may include, for a given type of amino acid residue, a probability distribution across the possible rotamers of that type of amino acid residue.

[0137] Referring again to FIG. 6A, in some cases, the design controller 117 may validate, based at least on the output 615 of the patch design computation model 115, an output molecule 625 generated by the molecule design computation model 113. In some cases, the molecule design computation model 113 may generate the output molecule 625 based on an input molecule 620.NAI-5002497275vl 69Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1For example, in some cases, the output molecule 625 may be generated by the molecule design computation model 113 denoising, over multiple successive timesteps, the input molecule 215. In some cases, this denoising may include updating the identity and / or conformation of one or more amino acid residues in the input molecule 620 to increase the likelihood that the output molecule 625 exhibits one or more desirable properties (e.g., drug-like properties), including a binding affinity towards the target molecule. However, due to the paucity of data at the full length scale of protein molecules for training the molecule design computation model 113 to reach optimal performance, the output molecule 625 may not exhibit a sufficiently high probability of exhibiting the one or more properties of interest. By contrast, the patch design computation model 115 may leverage the availability of data at the patch scale. Accordingly, as shown in FIG. 6A, the identity and conformation of at least one amino acid residue in the output molecule 625 generated by the molecule design computation model 113 may be validated and, in some cases, corrected based on the output 615 of the patch design computation model 115. For instance, in some cases, the design controller 117 may change the identity and / or conformation of the residue of interest in the output molecule 625 in accordance with the output 615 of the patch design computation model 115 if there are discrepancies between the identity and conformation of the residue of interest in the output molecule 625 and the identity and conformation of the residue of interest indicated by the output 615 of the patch design computation model 615. In doing so, the design controller 117 may generate a validated molecule 630 in which the identity and conformation of the residue of interest is consistent with the output 615 of the patch design computation model 115. As such, the validated molecule 630 may be identified as a candidate for wet lab assessment (e.g., in vitro measurements and / or in vivo characterization by the one or more laboratory equipment 140) atNAI-5002497275vl 7AAttorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 least because the validated molecule 630 may be more likely to exhibit the one or more properties of interest.

[0138] FIG. 6B depicts a schematic diagram illustrating an example of an amino acid residue 650, in accordance with some example embodiments. As shown in FIG. 6B, the amino acid residue 650 may include a plurality of backbone atoms, which would include the same combination of an alpha carbon (Ca) atom, an amino group (NH2), and a carboxylic group (COOH) regardless of the type of the amino acid residue 650. As further shown in FIG. 6B, the amino acid residue 650 may also include a plurality of sidechain atoms, which may vary depending on the type of the amino acid residue 650. In the example of the amino acid residue 650 show in FIG. 6B, the sidechain atoms of the amino acid residue 650 include a beta carbon ( ;), a gamma carbon a delta carbon ( ), and a zeta carbon ( ) coupled with a trio of nitrogen (N) atoms. In some cases, each rotamer of the amino acid residue 650 may correspond to a different spatial arrangement of the sidechain atoms of the amino acid residue 650. For example, in some cases, different rotamers of the amino acid residue 650 may be defined by imposing different constraints on how and to what extent the sidechain atoms of the amino acid residue 650 can move in three- dimensional space. One way in which the different rotamers of the amino acid residue 650 can be defined is based on the different torsion angles % present in the sidechain of the amino acid residue 650. In the example shown in FIG. 6B, these torsion angles include the first torsion angle Xi of the rotatable bond between the alpha carbon (C„) atom and the beta carbonatom, the second torsion angle2°f the rotatable bond between the beta carbonatom and the gamma carbon (Cy) atom, the third torsion angle / 3of the rotatable bond between the gamma carbon (Cy) atom and the delta carbon ( ) atom, and the fourth torsion angle / 4of the rotatable bond between the delta carbon (C5) atom and the nitrogen (N) atom. Defining the different rotamers of the aminoNAI-5002497275vl 71Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 acid residue 650 based on torsion angles / may limit the degree-of-freedom (DoF) of the sidechain atoms to rotations about one or more of the torsion angles . However, it should be appreciated that the different rotamers of the amino acid residue 650 may be defined differently to permit different and / or additional degrees-of-freedom (DoF), such as translations in the form of changes in the distance (e.g., bond length) between one or more sidechain atoms.

[0139] In some example embodiments, where the amino acid residue 650 is the residue of interest being identified by a patch design computation model (e.g., the patch design computation model 115 in FIG. 1), the sidechain conformation of the amino acid residue 650 may be recovered based at least on a probability distribution across the different possible rotamers of the amino acid residue 650. For example, in some cases, the patch design computation model may determine a probability distribution that includes a probability that the sidechain conformation of the amino acid residue 650 is one rotamer, a probability that the sidechain conformation of the amino acid residue 650 is a different rotamer, and / or the like. However, as noted, different rotamers of the amino acid residue 650 may be defined by imposing different constraints on how and to what extent the sidechain atoms of the amino acid residue 650 can move in three- dimensional space. In some cases, these constraints are fixed, meaning that a rotamer-based definition of the sidechain conformation of the amino acid residue 650 may yield a less granular representation of the amino acid residue 650. Accordingly, in some cases, the sidechain conformation of the amino acid residue 650 may be defined based on the individual torsion angles present in the amino acid residue 650. For instance, in the example shown in FIG. 6B, the sidechain conformation of the amino acid residue 650 may be defined based on the first torsion angle Xi of the rotatable bond between the alpha carbon (Ca) atom and the beta carbon ( ?) atom, the second torsion angle Xi of the rotatable bond between the beta carbon ( ) atom and the gammaNAI-5002497275vl 72Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 carbon (Cy) atom, the third torsion angle / 3of the rotatable bond between the gamma carbon (Cy) atom and the delta carbon (Cs) atom, and the fourth torsion angle / 4of the rotatable bond between the delta carbon (Cs) atom and the nitrogen (N) atom. In this case, each of the torsion angles / x, / 2, and Z4 may be permitted to its own individual degree-of-freedom, thus yielding a more granular representation of the amino acid residue 650.

[0140] In some example embodiments, each torsion angles Xi, Z2, Z3, and XA the amino acid residue 650 may be represented as a continuous value, such as a probability distribution across a continuous variable (e.g., a probability density function (PDF) and / or the like). Alternatively, instead of a continuous value, each of the torsion angles x13X2, Xs,and Z4 may be represented as a probability distribution (e.g., a categorical distribution) across a set of discrete values, for example, between —180° and +180°. For example, in some cases, the representation of the amino acid residue 650 may include, for each of the torsion angles Xi, Z2, Zs>and Z4, a probability distribution (e.g., a categorical distribution) across the values —180°, —90°, +90°, and + 180°. In some cases, the probability distribution (e.g., categorical distribution) associated with each of the torsion angles Xi, %2> Xs,and Z4 may include more (or fewer) possible values to increase (or decrease) the granularity of the representation of the amino acid residue 650. In some cases, instead of recovering the sidechain conformation of the amino acid residue 650 by sampling from a probability distribution (e.g., categorical distribution) across the possible rotamers of the amino acid residue 650, the sidechain conformation of the amino acid residue 650 may be recovered by sampling from the probability distributions (e.g., continuous or categorical distribution) of the possible values of each of the torsion angles Xi, X2, Xs, and XA- For instance, in some cases, the sidechain conformation of the amino acid residue 650 may be determined toNAI-5002497275vl 73Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 include may include, for each of the torsion angles Xi, / 2,and Z4,avalue a threshold probability (or a highest probability) of being the value of the torsion angle.

[0141] FIG. 7 depicts a block diagram illustrating an example of computing system 700, in accordance with some example embodiments. Referring to FIGS. 1-7, the computing system 700 may be used to implement the molecule design engine 110, the client device 120, the analysis engine 130, the one or more laboratory equipment 140, and / or any components therein.

[0142] As shown in FIG. 7, the computing system 700 can include a processor 710, a memory 720, a storage device 730, and input / output devices 740. The processor 710, the memory 720, the storage device 730, and the input / output devices 740 can be interconnected via a system bus 750. The processor 710 is capable of processing instructions for execution within the computing system 700. Such executed instructions can implement one or more components of, for example, the molecule design engine 110, the client device 120, the analysis engine 130, the one or more laboratory equipment 140, and / or the like. In some example embodiments, the processor 710 can be a single-threaded processor. Alternatively, the processor 710 can be a multithreaded processor. The processor 710 is capable of processing instructions stored in the memory 720 and / or on the storage device 730 to display graphical information for a user interface provided via the input / output device 740.

[0143] The memory 720 is a computer readable medium such as volatile or non-volatile that stores information within the computing system 700. The memory 720 can store data structures representing configuration object databases, for example. The storage device 730 is capable of providing persistent storage for the computing system 700. The storage device 730 can be a floppy disk device, a hard disk device, an optical disk device, or a tape device, or other suitable persistent storage means. The input / output device 740 provides input / output operations for theNAI-5002497275vl 74Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 computing system 700. In some example embodiments, the input / output device 740 includes a keyboard and / or pointing device. In various implementations, the input / output device 740 includes a display unit for displaying graphical user interfaces.

[0144] According to some example embodiments, the input / output device 740 can provide input / output operations for a network device. For example, the input / output device 740 can include Ethernet ports or other networking ports to communicate with one or more wired and / or wireless networks (e.g., a local area network (LAN), a wide area network (WAN), the Internet).

[0145] In some example embodiments, the computing system 700 can be used to execute various interactive computer software applications that can be used for organization, analysis and / or storage of data in various formats. Alternatively, the computing system 700 can be used to execute any type of software applications. These applications can be used to perform various functionalities, e.g., planning functionalities (e.g., generating, managing, editing of spreadsheet documents, word processing documents, and / or any other objects, etc.), computing functionalities, communications functionalities, etc. The applications can include various add-in functionalities or can be standalone computing products and / or functionalities. Upon activation within the applications, the functionalities can be used to generate the user interface provided via the input / output device 740. The user interface can be generated and presented to a user by the computing system 700 (e.g., on a computer screen monitor, etc.).

[0146] One or more aspects or features of the subject matter described herein can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs, field programmable gate arrays (FPGAs) computer hardware, firmware, software, and / or combinations thereof. These various aspects or features can include implementation in one or more computerNAI-5002497275vl 75Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device. The programmable system or computing system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0147] These computer programs, which can also be referred to as programs, software, software applications, applications, components, or code, include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object- oriented programming language, and / or in assembly / machine language. As used herein, the term “machine-readable medium” refers to any computer program product, apparatus and / or device, such as for example magnetic discs, optical disks, memory, and Programmable Logic Devices (PLDs), used to provide machine instructions and / or data to a programmable processor, including a machine — readable medium that receives machine instructions as a machine-readable signal. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor. The machine-readable medium can store such machine instructions non-transitorily, such as for example as would a non-transient solid-state memory or a magnetic hard drive or any equivalent storage medium. The machine-readable medium can alternatively or additionally store such machine instructions in a transient manner, such as for example, as would a processor cache or other random access memory associated with one or more physical processor cores.NAI-5002497275vl 76Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1

[0148] To provide for interaction with a user, one or more aspects or features of the subject matter described herein can be implemented on a computer having a display device, such as for example a cathode ray tube (CRT) or a liquid crystal display (LCD) or a light emitting diode (LED) monitor for displaying information to the user and a keyboard and a pointing device, such as for example a mouse or a trackball, by which the user may provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well. For example, feedback provided to the user can be any form of sensory feedback, such as for example visual feedback, auditory feedback, or tactile feedback; and input from the user may be received in any form, including acoustic, speech, or tactile input. Other possible input devices include touch screens or other touch-sensitive devices such as single or multi-point resistive or capacitive track pads, voice recognition hardware and software, optical scanners, optical pointers, digital image capture devices and associated interpretation software, and the like.

[0149] In the descriptions above and in the claims, phrases such as “at least one of’ or “one or more of’ may occur followed by a conjunctive list of elements or features. The term “and / or” may also occur in a list of two or more elements or features. Unless otherwise implicitly or explicitly contradicted by the context in which it is used, such a phrase is intended to mean any of the listed elements or features individually or any of the recited elements or features in combination with any of the other recited elements or features. For example, the phrases “at least one of A and B;” “one or more of A and B;” and “A and / or B” are each intended to mean “A alone, B alone, or A and B together.” A similar interpretation is also intended for lists including three or more items. For example, the phrases “at least one of A, B, and C;” “one or more of A, B, and C;” and “A, B, and / or C” are each intended to mean “A alone, B alone, C alone, A and B together, A and C together, B and C together, or A and B and C together.” Use of the term “based on,” aboveNAI-5002497275vl 77Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 and in the claims is intended to mean, “based at least in part on,” such that an unrecited feature or element is also permissible.

[0150] The subject matter described herein can be embodied in systems, apparatus, methods, and / or articles depending on the desired configuration. The implementations set forth in the foregoing description do not represent all implementations consistent with the subject matter described herein. Instead, they are merely some examples consistent with aspects related to the described subject matter. Although a few variations have been described in detail above, other modifications or additions are possible. In particular, further features and / or variations can be provided in addition to those set forth herein. For example, the implementations described above can be directed to various combinations and sub-combinations of the disclosed features and / or combinations and sub-combinations of several further features disclosed above. In addition, the logic flows depicted in the accompanying figures and / or described herein do not necessarily require the particular order shown, or sequential order, to achieve desired results. Other implementations may be within the scope of the following claims.NAI-5002497275vl 78

Claims

Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1CLAIMSWhat is claimed is:

1. A computer-implemented method of validating an output of a molecule design computation model using an output of a patch design computation model, comprising: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity and a conformation determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the input patch, a patch level residue identity and a patch level conformation of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity and the patch level conformation of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity and the conformation of the residue of interest determined by the molecule design computation model.

2. The method of claim 1, wherein the output of the patch design computation model includes a probability distribution across a plurality of possible types of amino acid residues for the residue of interest.

3. The method of claim 2, wherein the output of the patch design computation model further includes, for at least one possible type of amino acid residue, a probability distributionNAI-5002497275vl 79Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 across one or more possible rotamers of that type of amino acid residue.

4. The method of any of claims 2 to 3, wherein the output of the patch design computation model further includes, for each torsion angle present in each type of amino acid residue, a probability distribution across a plurality of different values.

5. The method of claim 4, wherein the probability distribution comprises a categorical distribution that includes, for each torsion angle present in a corresponding type of amino acid residue, a probability of the torsion angle being a first value and a probability of the torsion angle being a second value.

6. The method of any of claims 4 to 5, wherein the probability distribution comprises a continuous distribution.

7. The method of any of claims 1 to 6, wherein the output of the patch design computation model includes a probability of the residue of interest being a first type of amino acid residue and a probability of the residue of interest being a second type of amino acid residue.

8. The method of claim 7, wherein the output of the patch design computation model further includes a probability of the residue of interest being a rotamer of the first type of amino acid residue and a probability of the residue of interest being another rotamer of the first type of amino acid residue.

9. The method of any of claims 1 to 8, wherein the residue of interest is located at an interface between the protein molecule and a target molecule.

10. The method of claim 9, further comprising: generating the input patch by at least identifying, within the plurality of amino acid residues forming the protein molecule and / or a plurality of amino acid residues forming theNAI-5002497275vl 80Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 target molecule, the one or more additional amino acid residues for inclusion in the input patch.

11. The method of any of claims 9 to 10, further comprising: selecting, within a plurality of amino acid residues including the plurality of amino acid residues forming the protein molecule and excluding a plurality of amino acid residues forming the target molecule, the one or more additional amino acid residues for inclusion in the input patch.

12. The method of any of claims 9 to 11, wherein the target molecule is another protein molecule or a small molecule.

13. The method of any of claims 9 to 12, wherein the protein molecule is an antibody and the target molecule is an antigen.

14. The method of any of claims 1 to 13, further comprising: generating an embedding of the input patch to include a description of the atomic context of the one or more additional amino acid residues; and applying the patch design computation model to determine, based at least on the embedding of the input patch, the patch level identity and the patch level conformation of the residue of interest.

15. The method of claim 14, wherein the description of the atomic context of the one or more additional amino acid residues include an encoding of an identity of each additional amino acid residue.

16. The method of any of claims 14 to 15, wherein the description of the atomic context of the one or more additional amino acid residues include a positional encoding of a sequential position of each additional amino acid residue in a corresponding sequence of amino acid residues.NAI-5002497275vl 81Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-117. The method of any of claims 14 to 16, wherein the description of the atomic context of the one or more additional amino acid residues include one or more three-dimensional coordinates of a position of at least one backbone atom in each additional amino acid residue.

18. The method of claim 17, wherein at least one backbone atom includes an alpha carbon (Ca) atom.

19. The method of any of claims 1 to 18, further comprising: extracting, from a complex formed by a binding interaction between a sample protein molecule and a sample target molecule, a sample patch including one or more amino acid residues; generating a training sample in which an identity and a conformation of one or more amino acid residues in the sample patch are masked; and training, based at least on the training sample, the patch design computation model to recover an identity and a conformation of the one or more masked amino acid residues.

20. The method of any of claims 1 to 19, further comprising: applying the molecule design computation model to generate the protein molecule, where the molecule design computation model is trained at a scale of protein molecules, and where the molecule design computation model identifies a first type of amino acid residue as the identity of the residue of interest.

21. The method of claim 20, wherein the validating the identity and the conformation of the residue of interest includes determining, based at least on the output of the patch design computation model, a probability of the residue of interest being the first type of amino acid residue.

22. The method of claim 21, wherein the validating the identity and the conformationNAI-5002497275vl 82Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 of the residue of interest further includes in response to the probability of the residue of interest being the first type of amino acid residue satisfying one or more thresholds, determining, based at least on the output of the patch design computation model, a probability of the residue of interest being a rotamer of the first type of amino acid residue identified by the molecule design computation model, and in response to the probability of the residue of interest being the rotamer of the first type of amino acid residue failing to satisfy one or more thresholds, replacing, based at least on the output of the patch design computation model, the rotamer of the first type of amino acid residue with a different rotamer of the first type of amino acid residue.

23. The method of claim 22, wherein the configuration of the protein molecule comprises a conformation of one or more amino acid residues forming the protein molecule.

24. The method of claim 23, wherein the probability of the configuration of the protein molecule is further determined by at least applying the patch design computation model to determine a probability of a different residue of interest in the protein molecule being a rotamer of a second type of amino acid residue identified by the molecule design computation model.

25. The method of any of claims 21 to 24, wherein the validating the identity and conformation of the residue of interest further includes determining, based at least on the output of the patch design computation model, a probability of an angular value of one or more torsion angles present in the first type of amino acid residue identified by the molecule design computation model.

26. The method of claim 25, wherein the validating the identity and conformation of the residue of interest further includesNAI-5002497275vl 83Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 in response to the probability of the one or more torsion angles present in the first type of amino acid residue identified by the molecule design computation model failing to satisfy one or more thresholds, adjusting, based at least on the output of the patch design computation model, a value of the one or more torsion angles determined by the molecule design computation model.

27. The method of any of claims 21 to 26, further comprising: in response to the probability of the residue of interest being the first type of amino acid residue failing to satisfy one or more thresholds, replacing, based at least on the output of the patch design computation model, the first type of amino acid residue with a second type of amino acid residue, and identifying, based at least on the output of the patch design computation model, a rotamer of the second type of amino acid residue as the residue of interest.

28. The method of any of claims 21 to 27, further comprising: determining, based at least on the output of the patch design computation model, a probability of a configuration of the protein molecule generated by the molecule design computation model.

29. The method of claim 28, wherein the probability of the configuration of the protein molecule is determined by at least determining, based at least on the output of the patch design computation model, a probability of the residue of interest being the rotamer of the first type of amino acid residue.

30. The method of any of claims 1 to 29, further comprising: identifying, as the one or more additional amino acid residues in the patch, one or more amino acid residues located within a threshold distance of the residue of interest.31 . The method of any of claims 1 to 30, further comprising:NAI-5002497275vl 84Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 identifying, as the one or more additional amino acid residues in the patch, a threshold quantity of amino acid residues that are closest to the residue of interest.

32. The method of any of claims 1 to 31, wherein the atomic context around the residue of interest includes the one or more additional amino acid residues.

33. The method of any of claims 1 to 32, wherein the atomic context around the residue of interest includes one or more of a fluid, a ligand, a cofactor, or a crystallographic adjuvant.

34. A system, comprising: at least one data processor; and at least one memory storing instructions, which when executed by the at least one data processor, result in operations comprising the method of any of claims 1 to 33.

35. A non-transitory computer readable medium storing instructions, which when executed by at least one data processor, result in operations comprising the method of any of claims 1 to 33.

36. A computer-implemented method of validating an output of a molecule design computation model using an output of a patch design computation model, comprising: selecting, for validation, a residue of interest from a plurality of amino acid residues forming a protein molecule, where the residue of interest is associated with a residue identity determined by the molecule design computation model; generating an input patch to include the residue of interest and one or more additional amino acid residues; applying the patch design computation model to determine, based at least on the inputNAI-5002497275vl 85Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 patch, a patch level residue identity of the residue of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the residue of interest, an output indicative of the patch level residue identity of the residue of interest; and validating, based at least on the output of the patch design computation model, the residue identity of the residue of interest determined by the molecule design computation model.

37. The method of claim 36, wherein the patch design computation model has been further trained to generate, based at least on the atomic context around the residue of interest, the output to be indicative of a patch level conformation of the residue of interest.

38. The method of any of claims 36 to 37, further comprising: applying the patch design computation model to determine, based at least on the input patch, a patch level conformation of the residue of interest; and validating, based at least on the output of the patch design computation model, a conformation of the residue of interest determined by the molecule design computation model.

39. The method of any of claims 36 to 38, wherein the output of the patch design computation model includes a probability distribution across a plurality of possible types of amino acid residues.

40. The method of claim 39, wherein the output of the patch design computation model further includes, for each type of amino acid residue of the plurality of possible types of amino acid residues, a probability distribution across a plurality of possible rotamers.

41. The method of any of claims 36 to 40, wherein the output of the patch design computation model further includes, for each torsion angle present in the residue of interest, a probability distribution across a plurality of possible angular values.NAI-5002497275vl 86Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-142. The method of claim 42, wherein the probability distribution includes a probability of a first torsion angle present in the residue of interest having one angular value and a probability of the first torsion angle having a different angular value.

43. The method of claim 43, wherein the probability distribution further includes a probability of a second torsion present in the residue of interest having the one angular value and a probability of the second torsion angle having the different angular value.

44. A system, comprising: at least one data processor; and at least one memory storing instructions, which when executed by the at least one data processor, result in operations comprising the method of any of claims 36 to 43.

45. A non-transitory computer readable medium storing instructions, which when executed by at least one data processor, result in operations comprising the method of any of claims 36 to 43.

46. A computer-implemented method of validating an output of a molecule design computation model using an output of a patch design computation model, comprising: selecting, for validation, a functional group of interest from a small molecule generated by the molecule design computation model; generating an input patch to include the functional group of interest; applying the patch design computation model to determine, based at least on the input patch, a patch level identity of the functional group of interest, where the patch design computation model has been trained to generate, based at least on an atomic context around the functional group of interest, an output indicative of the patch level residue identity of the functional group of interest; andNAI-5002497275vl 87Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 validating, based at least on the output of the patch design computation model, an identity of the functional group of interest determined by the molecule design computation model.

47. The method of claim 46, wherein the output of the patch design computation model includes a probability distribution across a plurality of possible functional groups.

48. The method of claim 47, wherein the plurality of possible functional groups include one or more of an alcohol, an amine, a carbonyl, a halide, or an aromatic ring.

49. The method of any of claims 47 to 48, wherein the validating of the identity of the functional group of interest determined by the molecule design computation model includes in response to the probability of the functional group of interest a first type of functional group determined by the molecule design computation model failing to satisfy one or more thresholds, replacing, based at least on the output of the patch design computation model, the first type functional group determined by the molecule design computation model with a second type of functional group having a threshold probability of being the functional group of interest.

50. The method of any of claims 46 to 49, wherein the patch further includes at least a portion of one or more adjacent functional groups in the small molecule.

51. The method of any of claims 46 to 50, wherein the functional group of interest is located at an interface between the small molecule and a protein molecule.

52. The method of claim 51, wherein the patch further includes at least a portion of one or more amino acid residues from the protein molecule, and wherein the one or more amino acid residues are located at an interface between the small molecule and the protein molecule.

53. The method of any of claims 46 to 52, wherein the patch design computation model determines the patch level identity of the functional group of interest based at least on an embedding of the patch that includes a description of the atomic context surrounding theNAI-5002497275vl 88Attorney Ref.: 14786-062-228(103963-228062) / P39284-WO-1 functional group of interest.

54. A system, comprising: at least one data processor; and at least one memory storing instructions, which when executed by the at least one data processor, result in operations comprising the method of any of claims 46 to 53.

55. A non-transitory computer readable medium storing instructions, which when executed by at least one data processor, result in operations comprising the method of any of claims 46 to 53.NAI-5002497275vl 89

Images