Predicting intragenic complementation

Abstract

Methods and systems of using a machine learning model to identify pairs of loss of function protein variants which when combined may make up a functional multimeric protein are described. The machine learning model may be trained to identify combinations of protein variants which form an at least partially functional multimeric protein activity as compared to the wild-type multimeric protein

Description

PNDRI.020WO PATENTPREDICTING INTRAGENIC COMPLEMENTATIONCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Application No. 63 / 790,633, filed April 17, 2025, which is hereby incorporated by reference in its entirety.REFERENCE TO SEQUENCE LISTING SUBMITTED ELECTRONICALLY

[0002] The instant application contains a Sequence Listing which has been submitted electronically in XML file format and is hereby incorporated by reference in its entirety. Said XML copy, created on April 13, 2026, is named “PNDRI.020WO.xml” and is 12,532 bytes in size.STATEMENT REGARDING FEDERALLY SPONSORED R&D

[0003] This invention was made with government support under Grant No. 2310113 awarded by the National Science Foundation and Grant Nos. R01HD114863, R01GM134274 and 5T32HG000035-27 awarded by the National Institutes of Health. The government has certain rights in the invention.FIELD OF THE INVENTION

[0004] The disclosed technology relates to the fields of protein structure and machine learning. More particularly, the disclosed technology relates to predicting intragenic complementation in enzyme active sites using machine learning models.BACKGROUND

[0005] Genetic interactions may be obstacles to the goals of precision medicine, impeding accurate genotype-phenotype inferences. One type of genetic interaction, intragenic complementation, is a positive non-linear interaction between two loss-of-function alleles of the same gene. In a multimeric protein with multiple active sites, two variants that are individually deleterious may together restore partial or full protein function when their positions are spatially segregated across subunits such that intact active sites are preserved, a mechanism known as variant sequestration. There exists a continuing need for improving the characterization and prediction of intragenic complementation in proteins.PNDRI.02QWO PATENT

[0006] Human argininosuccinate lyase (ASL), a homotetrameric enzyme of the urea cycle, provides an illustrative example of intragenic complementation. Systematic experimental study of pairwise combinations of loss-of-function variants in ASL demonstrated that intragenic complementation is pervasive, with over 60% of approximately 3,600 tested combinations exhibiting restored growth. The occurrence of complementation w as found to depend entirely on variant position rather than amino acid identity, with variant positions clustering into discrete spatial groups corresponding to distinct local regions that assemble into the active sites of the tetramer.

[0007] Because intragenic complementation follows predictable structural rules, machine learning approaches can be applied to predict whether any two loss-of-function variants will complement one another. However, existing approaches have been limited in scope, and there remains a need in the art for computational methods capable of accurately predicting intragenic complementation across diverse multimeric proteins without requiring explicit three-dimensional protein structure as input. The present disclosure addresses these needs and others.SUMMARY

[0008] In some embodiments, provided herein is a computer-implemented method for predicting complementation or non-complementation between at least two variants of a multimeric protein, the method comprising: (a) obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site: (b) identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence; (c) obtaining, for each of the first variant and the second variant, a local structural context representation; (d) determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output; and (e) classifying, from the complementation prediction output, the first variant and the second variant as a complementing pair or a non-complementing pair, wherein steps (a) through (e) may be performed in any order, concurrently, or iteratively.

[0009] In some embodiments, provided herein is a computer-implemented method of generating a dataset for modeling complementation in multimeric proteins, the method comprising: generating, via a high-throughput functional assay, experimental measurements of functional activity for a plurality of pairw ise combinations of variants of a multimeric protein, wherein each variant comprises a modification at a respective residue position in an amino acid sequence; constructing a dataset comprising a plurality of entries,PNDRI.020WO PATENTeach entry corresponding to a respective pairwise combination of variants and comprising: identifiers of the pairwise combination of variants, a quantitative measurement of functional activity associated with the pairwise combination, and a label indicating a degree of non-additive interaction between the variants; and organizing the dataset such that variant pairs exhibiting complementation are distinguishable from variant pairs that do not exhibit complementation, wherein the dataset encodes relationships between spatially distinct residue positions that contribute to restoration of protein function upon multimer assembly.

[0010] In some embodiments, provided herein is a non-transitory computer-readable medium storing a dataset for modeling complementation in multimeric proteins, the dataset comprising a plurality of entries, each entry corresponding to a pairwise combination of variants of a multimeric protein, each entry comprising: (i) identifiers of a first variant and a second variant at respective residue positions, (ii) a quantitative functional activity measurement obtained from experimental evaluation of the painvise combination, and (iii) a non-additivity metric representing deviation from an expected additive functional outcome; wherein the dataset is structured to encode complementation relationships between variants located in different spatial regions of the protein, such that the dataset enables identification of variant pairs that restore at least partial functional activity7upon multimer assembly.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Features of examples of the present disclosure will become apparent by reference to the following detailed description and drawings, in which like reference numerals correspond to similar, though perhaps not identical, components. For the sake of brevity7, reference numerals or features having a previously described function may or may not be described in connection with other drawings in which they appear.

[0012] FIG. 1A illustrates examples of intragenic complementation by variant sequestration in a dimeric protein.

[0013] FIG. IB shows a surface representation of human ASL protein (PDB: 1AOS) monomers and a homotetramer.

[0014] FIG. 2A illustrates a machine learning (ML) model that predicts intragenic complementation of ASL monomers according to some embodiments of the technology.

[0015] FIG. 2B shows predicted ASL diploid grow th versus experimentally measured ASL diploid growth in connection with the ML model illustrated in FIG. 2A.PNDRI.020WO PATENT

[0016] FIG. 2C shows an AUC-ROC curve of the predicted ASL diploid growth from 20-fold cross-validation in connection with the ML model illustrated in FIG. 2A.

[0017] FIG. 3 shows a series of charts illustrating the results of ML-predicted growth of ASL diploids having particular amino acid mutations from L1AAPO cross-validation in connection with the ML model illustrated in FIG. 2A.

[0018] FIG. 4A illustrates a three dimensional protein structure alignment of human ASL and human fumarase (FH) homotetramers.

[0019] FIG. 4B is a bar chart showing the experimental results of the distribution of FH compound heterozygous diploid yeast grow th.

[0020] FIG. 4C is a heatmap of the normalized diploid growth of all variant combinations tested in FH.

[0021] FIG. 4D is a three-dimensional illustration of residues on a surface representation of a human FH monomer based on clusters identified in the heatmap of FIG. 4C.

[0022] FIG. 4E is a chart showing ML model-predicted versus experimentally-measured FH compound heterozy gous diploid growth.

[0023] FIG. 4F is a chart showing the ROC-AUC curve of the predicted FH compound heterozygous diploid growth.

[0024] FIG. 5A is an ovendew of a high throughput process for yeast-based growth assays for characterizing the functional impact of SNV-accessible ASL missense variants.

[0025] FIG. 5B is a bar chart which shows the experimental data of the growth of humanized yeast harboring wild-type human A SL, Q286, R12Q, and T131M variants.

[0026] FIG. 5C is a chart show ing the normalized experimental data of the growth of 2,431 SNV-accessible ASL missense variants.

[0027] FIG. 6 is a chart showing the normalized experimental data of the growth of ASL missense variants in ClinVar.

[0028] FIG. 7A is a bar chart showing the distribution of ASL compound heterozygote diploid yeast growth.

[0029] FIG. 7B is a heatmap of the average normalized diploid growth of all variant combinations between two amino acid positions.

[0030] FIG. 7C is a three-dimensional surface representation of an ASL monomer.

[0031] FIG. 7D is a three-dimensional surface representation of an ASL tetramer wdth a magnified view- of an active site.

[0032] FIG. 8A is a bar chart which shows the experimental data of the growth of ASL variant combinations.PNDRI.020WO PATENT

[0033] FIG. 8B is a plot which illustrates the additive allelic effect in complementing compound heterozy gous variant combinations.

[0034] FIG. 9 is a chart which summarizes the experimental data of the normalized growth for all tested ASL variant combinations.

[0035] FIG. 10A is an example system overview for a predictive system, in accordance with an embodiment.

[0036] FIG. 10B depicts an example block diagram for predicting complementation or noncomplementation betw een at least two variants of a multimeric protein, in accordance with an embodiment.

[0037] FIG. 11 depicts a process for predicting complementation or non-complementation between at least two variants of a multimeric protein, in accordance with an embodiment.

[0038] FIG. 12 illustrates an exemplary computing device 600 for implementing systems and methods.DETAILED DESCRIPTION

[0039] All patents, patent applications, and other publications, including all sequences disclosed within these references, referred to herein are expressly incorporated herein by reference, to the same extent as if each individual publication, patent or patent application was specifically and individually indicated to be incorporated by reference. All documents cited are, in relevant part, incorporated herein by reference in their entireties for the purposes indicated by the context of their citation herein. However, the citation of any document is not to be construed as an admission that it is prior art with respect to the present disclosure.Definitions

[0040] Unless defined otherwise, technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. See, e.g. Singleton et al., Dictionary of Microbiology and Molecular Biology' 2nd ed., J. Wiley & Sons (New' York, NY 1994); Sambrook et al., Molecular Cloning, A Laboratory Manual, Cold Spring Harbor Press (Cold Spring Harbor, NY 1989). For purposes of the present disclosure, the following terms are defined below.

[0041] The articles "a" and "an" are used herein to refer to one or to more than one (i.e., to at least one) of the grammatical object of the article. By way' of example, "an element" means one element or more than one element.PNDRI.020WO PATENT

[0042] As used herein, "about" will be understood by persons of ordinary' skill and will vary to some extent depending on the context in which it is used. If there are uses of the term which are not clear to persons of ordinary skill given the context in which it is used, "about" will mean up to plus or minus 10% of the particular value.

[0043] The term "and / or" is used in this disclosure to mean either "and" or "or" unless indicated otherwise.

[0044] It should be understood that the expression "at least one of includes individually each of the recited objects after the expression and the various combinations of two or more of the recited obj ects unless otherwise understood from the context and use. The expression "and / or" in connection with three or more recited objects should be understood to have the same meaning unless otherwise understood from the context.

[0045] Conditional language used herein, such as, among others, “can,” “might,” “may,” “e.g.,” and the like, unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments include, while other embodiments do not include, certain features, elements and / or states. Thus, such conditional language is not generally intended to imply that features, elements and / or states are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without author input or prompting, whether these features, elements and / or states are included or are to be performed in any particular embodiment. The terms “comprising,” “including.” “having,” “involving.” and the like are synonymous and are used inclusively, in an open-ended fashion, and do not exclude additional elements, features, acts, operations, and so forth. Also, the term “or” is used in its inclusive sense (and not in its exclusive sense) so that when used, for example, to connect a list of elements, the term “or” means one, some, or all of the elements in the list.

[0046] Disjunctive language such as the phrase “at least one of X, Y or Z,” unless specifically stated otherwise, is otherw ise understood with the context as used in general to present that an item, term, etc., may be either X, Y or Z, or any combination thereof (e.g., X, Y and / or Z). Thus, such disjunctive language is not generally intended to, and should not, imply that certain embodiments require at least one of X, at least one of Y or at least one of Z to each be present.

[0047] The terms “about” or “approximate” and the like are synonymous and are used to indicate that the value modified by the term has an understood range associated with it, where the range can be ±20%, ±15%, ±10%, ±5%, or±l%. The term “substantially” is used to indicate that a result (e.g., measurement value) is close to a targeted value, where closePNDRI.020WO PATENTcan mean, for example, the result is within 80% of the value, within 90% of the value, within 95% of the value, or within 99% of the value.

[0048] The use of the term "include," "includes," "including," "have," "has," "having," "contain," "contains," or "containing," including grammatical equivalents thereof, should be understood generally as open-ended and non-limiting, for example, not excluding additional unrecited elements or steps, unless otherwise specifically stated or understood from the context.

[0049] It should be understood that the order of steps or order for performing certain actions is immaterial so long as the present disclosure remains operable. Moreover, two or more steps or actions may be conducted simultaneously.

[0050] At various places in the present specification, variables or parameters are disclosed in groups or in ranges. It is specifically intended that the description include each and every individual subcombination of the members of such groups and ranges. For example, an integer in the range of 0 to 5 is specifically intended to individually disclose 0, 1, 2, 3, 4, 5, and an integer in the range of 1 to 3 is specifically intended to individually disclose 1, 2, and 3.

[0051] The use of any and all examples, or exemplary language herein, for example, "such as" or "including," is intended merely to illustrate better the present disclosure and does not pose a limitation on the scope of any disclosure(s) unless claimed. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of that provided by the present disclosure.

[0052] Throughout the description, where systems and compositions are described as having, including, or comprising specific components, or where processes and methods are described as having, including, or comprising specific steps, it is contemplated that, additionally, there are systems, compositions, and kits of the present disclosure that consist essentially of, or consist of, the recited components, and that there are processes and methods according to the present disclosure that consist essentially of, or consist of, the recited processing steps.

[0053] In the disclosure, where an element or component is said to be included in and / or selected from a list of recited elements or components, it should be understood that the element or component can be any one of the recited elements or components, or the element or component can be selected from a group consisting of two or more of the recited elements or components.PNDRI.02QWO PATENT

[0054] Further, it should be understood that elements and / or features of a system or a method provided and described herein can be combined in a variety of ways without departing from the spirit and scope of the present disclosure, whether explicit or implicit herein. For example, where reference is made to a particular system, that system can be used in various embodiments of systems of the present disclosure and / or in methods of the present disclosure, unless otherwise understood from the context. In other words, within this application, embodiments have been described and depicted in a way that enables a clear and concise application to be written and rawn, but it is intended and wall be appreciated that embodiments may be variously combined or separated without departing from the present teachings and disclosure(s).

[0055] As used herein, a “nucleotide” includes a nitrogen-containing heterocyclic base, a sugar, and one or more phosphate groups. Nucleotides are monomeric units of a nucleic acid sequence. Examples of nucleotides include, for example, ribonucleotides or deoxyribonucleotides. In ribonucleotides (RNA), the sugar is a ribose, and in deoxyribonucleotides (DNA), the sugar is a deoxyribose, i.e., a sugar lacking a hydroxyl group that is present at the 2' position in ribose. The nitrogen-containing heterocyclic base can be a purine base or a pyrimidine base. Purine bases include adenine (A) and guanine (G), and modified derivatives or analogs thereof. Pyrimidine bases include cytosine (C), thymine (T), and uracil (U), and modified derivatives or analogs thereof. The C-l atom of deoxyribose is bonded to N-l of a pyrimidine or N-9 of a purine. The phosphate groups may be in the mono-, di-, or tri-phosphate form.

[0056] A “genome,” as used herein refers to the complete genetic information of an organism or virus, expressed in nucleic acid sequences.

[0057] As used herein, a “protein” refers to a large biomolecule that comprise one or more long chains of amino acid residues where two or more amino acids joined together by an amide bond (that is, a “peptide bond”).

[0058] As used herein, the term "multimeric protein" refers to a protein complex comprising two or more polypeptide subunits. A multimeric protein can be a homomultimeric protein, in which all subunits are identical, or a heteromultimeric protein, in which at least two subunits are non-identical. Examples of homomultimeric proteins include homodimers, homotrimers, homotetramers, homopentamers, and higher-order homomultimers. The subunits of a multimeric protein associate through non-covalent and / or covalent interactions to form a functional complex.PNDRI.020WO PATENT

[0059] As used herein, the term "variant" refers to an amino acid sequence that differs from a reference sequence at one or more residue positions. A variant can comprise one or more amino acid substitutions, insertions, or deletions relative to the reference sequence. A "loss-of-function variant" refers to a variant that exhibits substantially reduced or absent protein activity relative to the wild-type protein, for example, exhibiting no more than 10% of wildtype protein activity when present alone in a homozygous context.

[0060] As used herein, the term "complementation" refers to the phenomenon in which two loss-of-function variants of a multimeric protein, when present together in a compound heterozygous organism or cell, restore at least partial protein function. Complementation occurs when the two variants are located at different functional regions or spatial clusters of the multimeric protein, such that each subunit harboring one variant can compensate for the deficiency of the other subunit, resulting in a non-additive restoration of functional activity'. A "complementing pair" refers to a pair of variants that together exhibit restoration of functional activity', for example at least 50% of wild-type protein activity. A "noncomplementing pair" refers to a pair of variants that together fail to restore substantial functional activity.

[0061] As used herein, the term "functional site" refers to a region of a multimeric protein that is involved in the protein's biological activity', including but not limited to an active site, ligand-binding site, catalytic site, or multichain active site formed by residues contributed by at least two subunits of the multimeric protein.

[0062] As used herein, the term "local structural context representation" refers to a computational representation of the amino acid sequence and / or structural environment surrounding a particular residue position in a protein. In various embodiments, the local structural context representation is generated by extracting a sequence window centered on a variant residue position and computing an embedding of that sequence window using a pretrained protein language model. The local structural context representation encodes information about the local sequence and structural context of the variant without requiring explicit three-dimensional protein structure coordinates as input.

[0063] As used herein, the term "machine learning model” refers to a computer-implemented model that can be tuned (e.g., trained) based on inputs to approximate unknown functions. In particular, in some embodiments, a machine-learning model includes a model that utilizes computer processes or procedures to learn from, and make predictions on. known data by analyzing the known data to leam to generate outputs that reflect patterns and attributes of the known data. For instance, in some cases, a machine-PNDRI.02QWO PATENTleaming model includes, but is not limited to, aneural network (e.g., a convolutional neural network, recurrent neural network or other deep learning network), a decision tree (e.g., a gradient boosted decision tree, such as XGBoost), association rule learning, inductive logic programming, support vector learning, a Bayesian network, a regression-based model (e.g., censored regression), principal component analysis, or a combination thereof.

[0064] Further, as used herein, the term “neural network” refers to a model of interconnected artificial neurons (e.g., organized in layers) that communicate and learn to approximate complex functions and generate outputs based on inputs provided to the model. In some instances, a neural network includes one or more machine learning processes or algorithms. Further, in some cases, a neural network includes process (or set of processes) that implement deep learning techniques that utilize a set of procedures to model high-level abstractions in data. To illustrate, in some embodiments, a neural network includes a convolutional neural network, a recurrent neural network (e.g., a long short-term memory neural network), a generative adversarial network, a graph neural network, a multilayer perceptron, or a diffusion neural network. In some embodiments, a neural network includes a combination of neural networks or neural network components.

[0065] A “deep learning neural network” uses multiple layers (ranging from three to several hundred or thousands, or more) in the neural network, where the hierarchy of layers is used to transform input data into a progressively more abstract and composite representation. Examples of deep learning neural network include convolutional neural networks and transformers.

[0066] A “transformer” is a deep learning architecture that is based on the multi-head attention mechanism. Transformers have no recurrent units, therefore requiring less training time than recurrent neural architectures. In a transformer, text is converted to numerical representations called tokens, and each token is converted into a vector via lookup from a word embedding table. At each layer, each token is then contextualized within the scope of the context window with other (unmasked) tokens via a parallel multihead attention mechanism, allowing the signal for key tokens to be amplified and less important tokens to be diminished. Transformers may be used to train (large) language models.

[0067] Transformers may use an encoder-decoder architecture. The encoder consists of encoding layers that process all the input tokens together one layer after another, while the decoder consists of decoding layers that iteratively process the encoder's output and the decoder's output tokens so far. The purpose of each encoder layer is to create contextualizedPNDRI.020WO PATENTrepresentations of the tokens, where each representation corresponds to a token that mixes information from other input tokens via self-attention mechanism. Each decoder layer contains two attention sublayers: (1) cross-attention for incorporating the output of encoder (contextualized input token representations), and (2) self-attention for mixing information among the input tokens to the decoder (i.e. the tokens generated so far during inference time). Both the encoder and decoder layers have a feed-forward neural network for additional processing of their outputs and contain residual connections and layer normalization steps. These feed-forward layers may contain most of the parameters in a transformer model.

[0068] A “protein language model” is language model for protein sequence representation learning that distills information from massive protein sequence databases to capture complex dependencies between amino acids. Protein language models may be used to generate protein sequences with a predictable function and facilitate optimization of a given protein.

[0069] As used herein, the term "embedding" refers to a numerical vector representation of a sequence or structural element, generated by a computational model such as a pretrained protein language model. An embedding encodes properties of the input sequence in a high-dimensional vector space, enabling downstream computational analysis. In various embodiments, the embedding comprises a vector of about 500 to about 5,000 parameters, for example about 1,280 parameters.

[0070] As used herein, the term "pretrained protein language model" refers to a machine learning model that has been trained on large corpora of protein sequence data to leam representations of amino acid sequences. Non-limiting examples of pretrained protein language models include ESM-2, ESM-lb, ESM-lv, ProtTrans, ProtBERT, TAPE, UniRep, SeqVec, and AlphaFold-based language models, as well as any functionally equivalent model capable of generating local structural context representations of amino acid sequences.

[0071] As used herein, the term "trained computational model" refers to a machine learning model that has been trained on experimental data to leam relationships between inputs and outputs and to generate predictions on new, unseen data. In various embodiments, the trained computational model is a neural network, including a transformer-based neural network, configured to receive as input the local structural context representations of two variants and to output a complementation prediction. The trained computational model canPNDRI.020WO PATENTbe an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.

[0072] As used herein, the term "complementation prediction output" refers to the output generated by a trained computational model that indicates whether a pair of variants is predicted to complement one another. The complementation prediction output can comprise a continuous predicted growth value, a probability of complementation, a binary complementation class label, or a direct classification or ranking generated without a separately computed interaction score.

[0073] As used herein, the term "compound heterozygous" refers to an organism or cell harboring two different variants at the same genetic locus, one variant on each allele. In the context of the present disclosure, compound heterozygosity refers to the state in which a first variant is present on one allele and a second variant is present on the second allele of the same gene encoding a multimeric protein subunit.

[0074] As used herein, the term "additive model" refers to a mathematical model that predicts the combined functional effect of two variants based on the sum or linear combination of the individual effects of each variant alone, without accounting for nonadditive interactions. A deviation from the additive model indicates a non-additive interaction between variants, which can reflect complementation or epistasis.

[0075] As used herein, the term "non-additive interaction" or "nonadditivity" refers to a functional outcome resulting from a combination of two variants that is greater or lesser than the sum of the individual effects of each variant alone. In the context of complementation, a non-additive interaction typically refers to a greater-than-additive restoration of functional activity when two loss-of-function variants are combined in a compound heterozygous context.

[0076] As used herein, the term "high-throughput functional assay" refers to an experimental method capable of measuring the functional activity of a large number of variants or variant combinations simultaneously or in rapid succession. Non-limiting examples include yeast-based growth assays under selective conditions and deep mutational scanning.

[0077] As used herein, the term "deep mutational scanning" refers to a high-throughput experimental method that simultaneously measures the functional effects of a large number of protein variants by combining site-directed mutagenesis or random mutagenesis with a functional selection or screening assay and next-generation sequencing.PNDRI.02QWO PATENT

[0078] As used herein, the phrase "obtaining, or having obtained" encompasses both actively obtaining information or data (e.g., by performing an experiment or retrieving data from a database) and passively receiving information or data that has previously been obtained by another party.

[0079] As used herein, the term "subject" refers to any organism, including a human or non-human animal. In embodiments involving validation of a complementation prediction, a subject can be an organism expressing both a first variant and a second variant of a multimeric protein in a compound heterozygous configuration, including but not limited to a yeast diploid strain or a mammalian organism.Introduction

[0080] Disclosed herein are methods, systems, and non-transitory computer-readable media for predicting complementation or non-complementation between variants of a multimeric protein. Multimeric proteins assemble from two or more subunits to form functional complexes, and disease-associated variants of such proteins are frequently compound heterozygous in affected individuals. Accurately predicting whether two loss-of-function variants will complement one another when present together in a compound heterozygous genotype is a problem of both fundamental biological importance and direct clinical relevance. The methods, systems, and non-transitory computer-readable media disclosed herein enable, for example: (i) prediction of complementation or noncomplementation for pairwise combinations of variants of a multimeric protein; (ii) identification of candidate multimeric proteins for complementation based on structural features of their multichain active sites; (iii) precomputation of complementation predictions for all pairwise variant combinations at scale; (iv) generation and organization of datasets for training computational models to predict complementation; and (v) use of complementation predictions to inform therapeutic approaches and predict phenotypic severity of compound heterozy gous genotypes in subjects.

[0081] By representing each variant as a local structural context representation derived from a sequence window centered on the variant residue position using a pretrained protein language model, and training a computational model on experimental measurements of functional activity for pairwise variant combinations, it is possible to generate accurate predictions of complementation that generalize across structurally related proteins. The approach yields complementation prediction outputs constructed from the local structural context representations of two variants, which are the substrate for classification via aPNDRI.020WO PATENTtrained computational model, including transformer-based neural networks and ensemble models; each pairwise variant combination is evaluated using the local structural context representations of both variants, and the resulting complementation prediction output is used to classify the pair as complementing or non-complementing. The complementation prediction outputs resulting from this approach can be further combined with an additive model to quantify the predicted level of functional restoration for complementing pairs. This approach enables automated, high-throughput assessment of variant interactions in multimeric proteins and large-scale precomputation of complementation landscapes across entire variant libraries.

[0082] When characterizing the functional consequences of genetic variation in multimeric proteins, it is valuable to understand the structural context in which variant interactions occur. In compound heterozygous individuals, many of the differences in phenotypic severity are attributable to whether the two variants present on opposite alleles complement one another upon multimer assembly. The biological mechanisms underlying these differences are incompletely understood by existing single-variant effect prediction approaches. Traditional approaches to analyzing protein variants often fail to capture the pairwise interactions between variants that determine functional outcome for compound heterozygous genotypes. Moreover, many disease-associated variants occur at low7frequency in human populations and have been observed only rarely in compound heterozygous combinations, making it difficult to identify meaningful patterns from clinical observation alone. Additionally, variant-to-variant differences in complementation interactions mean that predictions of phenotypic severity effective for one variant pair may not apply to another, and understanding these differences requires the ability to model and test hypotheses about variant interactions in a protein-specific structural context.

[0083] These challenges make the prediction of complementation fundamentally different from traditional single-variant effect prediction approaches that do not account for pairwise variant interactions. At a traditional molecular level, it may be sufficient to consider the functional effect of individual variants in isolation. However, this analysis necessarily presupposes that variant interactions are additive, which is not generally true for multimeric proteins in which active sites are formed by residues contributed by more than one subunit. Variants that are individually deleterious may together restore partial or full protein activity when their residue positions reside in different spatial regions of the protein that contribute to distinct active site assemblies, producing a non-additive functional outcome that cannot be predicted from single-variant scores alone. This problem compounds when consideringPNDRI.020WO PATENTclinical interpretation of compound heterozy gous genotypes, as a given variant pair may complement one another in one structural context but not another, producing unreliable predictions of phenotypic severity if only individual variant effects are considered.

[0084] If structural context is informative for predicting variant interactions, then additional parameters beyond individual variant effect scores may be considered. One approach is to generate local structural context representations for each variant, which are numerical representations of the sequence and structural environment surrounding a variant residue position, derived from a pretrained protein language model, even without requiring explicit three-dimensional protein structure coordinates as input. A sequence window centered on the variant residue position provides the local context, conditioning the representation to capture information about the structural neighborhood of each variant. By computing local structural context representations for both variants in a pair and providing them as input to a trained computational model, a prediction of whether the pair will complement one another can be generated. The salient output is not the individual effect of either variant in isolation, but rather the complementation prediction output that reflects the joint structural relationship of both variants within the multimeric protein.

[0085] At the variant-pair level, this property of structural context-dependent complementation is representative of phenotypic heterogeneity in compound heterozy gous individuals; complementation predictions are informative across different multimeric proteins and variant types, enabling quantification of how strongly the structural positions of two variants are predictive of their functional interaction. Complementation predictions constructed from local structural context representations can be used for large-scale variant pair classification, generation of complementation maps or heatmaps identify ing which variant combinations are predicted to complement one another, clustering of variants by residue position into spatially distinct complementation groups, and prediction of phenotypic severity for compound heterozygous genotypes in subjects.

[0086] Disclosed herein are methods, systems, and non-transitory computer-readable media for predicting complementation or non-complementation between variants of a multimeric protein and generating complementation datasets and trained computational models for use in such predictions. The approaches described herein involve obtaining a local structural context representation for each variant by extracting a sequence window centered on the variant residue position and computing an embedding of that window using a pretrained protein language model, which in non-limiting embodiments is ESM-2 or a functionally equivalent model trained on large corpora of protein sequence data. LocalPNDRI.020WO PATENTstructural context representations are generated for both variants in a pair and provided as input to a trained computational model, such as a transformer-based neural network with an attention pooling layer, to generate a complementation prediction output. The trained computational model can be trained on experimental measurements of functional activity for pairwise variant combinations obtained by a high-throughput functional assay, such as a yeast-based growth assay or deep mutational scanning, and may include models trained using cross-validation, including leave-one-amino-acid-position-out cross-validation, or ensemble models comprising predictions averaged across models trained during cross-validation. Altogether, this approach enables more accurate prediction of the functional consequences of compound heterozygous genotypes, large-scale precomputation of complementation landscapes across entire variant libraries, generalization of complementation predictions to structurally related proteins beyond the protein used for training, and clinically actionable classification of variant pairs to inform therapeutic approaches and patient-specific genoty pe-phenotype predictions.

[0087] FIG. 10A is an exemplary system overview for generating complementation predictions based on local structural context representations of protein variants. The system overview includes a predictive system 130 and one or more third party7entities 110A and / or HOB in communication with one another through a network 120. FIG. 10A depicts one embodiment of the overall system environment. In other embodiments, additional or fewer third party entities 110 in communication with the predictive system 130 can be included. Generally, the predictive system 130 implements methods disclosed herein for generating complementation predictions for pairwise variant combinations of multimeric proteins. As referenced herein, methods for generating complementation predictions can include one or more of: classifying variant pairs as complementing or non-complementing; generating complementation maps or heatmaps identifying which variant combinations are predicted to complement one another; clustering variants by residue position into spatially distinct complementation groups; quantifying predicted levels of functional restoration for complementing pairs using an additive model; predicting phenotypic severity of compound heterozygous genotypes in subjects; and identifying candidate multimeric proteins for complementation from protein structure databases. The third party7entities 110 communicate with the predictive system 130 for purposes associated with using the complementation predictions generated therefrom, including, for example, clinical variant interpretation, genetic counseling, patient stratification, rational drug discovery, or therapeutic target identification in disease genes encoding multimeric proteins.PNDRI.020WO PATENT

[0088] In various embodiments, the third party' entity 110 represents a partner entity' of the predictive system 130 that operates either upstream or downstream of the predictive system 130. As one example, the third party entity 110 operates upstream of the predictive system 130 and provides information to the predictive system 130 to enable the generation of complementation predictions. In this scenario, the predictive system 130 receives data from the third party entity 110, examples of which include variant identity and position data for one or more multimeric proteins of interest, a subset of which may have experimental functional activity measurements for some pairwise variant combinations (e.g., growth values from a yeast-based assay, percent wild-type activity', or deep mutational scanning enrichment scores). In some embodiments, these data comprise amino acid sequences of multimeric proteins of interest and identifiers of variants to be evaluated in pairwise combinations. In some embodiments, the data further comprises three-dimensional structural information used to identify candidate proteins for complementation based on multichain active site geometry', optionally supplemented by ortholog databases to identify human orthologs of candidate proteins from model organisms. The predictive system 130 performs methods disclosed herein to generate complementation predictions from the provided data. As another example, the third party entity 110 operates downstream of the predictive system 130. In this scenario, the predictive system 130 generates complementation predictions and provides those predictions to the third party entity 110. The third party entity 110 can subsequently use the information for their own purposes. For example, the third party entity 110 may be a clinical diagnostic entity or genetic counseling service. Therefore, given the complementation predictions generated by the predictive system 130, the third party entity' 110 may predict the phenotypic severity of a compound heterozygous genotype in a subject and inform clinical decision-making regarding therapeutic approaches, including gene therapy, enzyme replacement therapy, or small molecule targeting. If the predictions identify' variant combinations associated with a milder disease phenotype due to complementation, the third party entity 110 may receive the predictions and use them to stratify patients for clinical trials or design new therapeutics that seek to modulate the relevant protein function and ameliorate one or more symptoms of the associated disease.

[0089] Referring to the network 120, any suitable network 120 may be implemented that enables connection between the predictive system 130 and third party' entities 110. The network 120 may comprise any combination of local area and / or wide area networks, using both wired and / or wireless communication systems. In one embodiment, the network 120PNDRI.020WO PATENTuses standard communications technologies and / or protocols. For example, the network 120 includes communication links using technologies such as Ethernet, 802.11. worldwide interoperability for microwave access (WiMAX), 3G, 4G, 5G, code division multiple access (CDMA), digital subscriber line (DSL), etc. Examples of networking protocols used for communicating via the network 120 include multiprotocol label switching (MPLS), transmission control protocol / Intemet protocol (TCP / IP), hypertext transport protocol (HTTP), simple mail transfer protocol (SMTP), and file transfer protocol (FTP). Data exchanged over the network 120 may be represented using any suitable format, such as hypertext markup language (HTML) or extensible markup language (XML). In some embodiments, all or some of the communication links of the network 120 may be encrypted using any suitable technique or techniques.

[0090] FIG. 10B is an exemplary system architecture for the predictive system 130. The predictive system 130 includes input data 200, a complementation prediction module 300, and an output module 400. In various embodiments, the predictive system 130 may include additional or fewer modules than those shown in FIG. 10B. The input data 200 comprises, for each variant pair to be evaluated, the amino acid identity and position of each variant within the primary sequence of the multimeric protein of interest, along with the full amino acid sequence of the multimeric protein sufficient to extract a sequence window centered on each variant residue position. In some embodiments, the variant input data 150 further comprises three-dimensional structural information for the multimeric protein, such as coordinates from the Protein Data Bank, used to identify candidate proteins for complementation based on multichain active site geometry or to validate complementation predictions against known structural clusters. The complementation prediction module 300 receives as input the embedding vectors for both variants in a pair and inputs them into a trained computational model, such as a transformer-based neural netw ork with an attention pooling layer and a prediction layer, to generate a complementation prediction output comprising a predicted growth value, a probability of complementation, or a binary complementation class label. In some embodiments, the complementation prediction module 300 comprises an embedding module 350, which extracts a sequence window of a defined length centered on each variant residue position and computes a local structural context representation for each window7using a pretrained protein language model, producing a numerical embedding vector for each variant. In some embodiments, the complementation prediction module 300 applies an ensemble approach in which predictions are averaged across a lurality of models trained during cross-validation. ThePNDRI.020WO PATENToutput module 400 returns the complementation prediction output for each queried variant pair and, optionally, generates a complementation map or heatmap visualizing predicted complementation across all pairwise combinations of variants in a protein of interest, a clustering of variant positions into spatially distinct complementation groups, or a ranked list of all pairwise combinations sorted by predicted complementation probability, enabling downstream clinical interpretation or large-scale precomputation across a variant library.

[0091] The methods and systems disclosed herein provide specific technological improvements to computer functionality and the technical field of variant effect prediction and computational genetics. Unlike conventional approaches that predict the functional effect of individual variants in isolation without accounting for pairwise variant interactions, the disclosed methods generate complementation predictions conditioned on the local structural context representations of both variants in a pair, capturing non-additive epistatic interactions that are not accessible from single-variant effect scores alone. This pairwise structural conditioning represents a concrete technical constraint that ties the computational process to the physical organization of residues within the multimeric protein complex. In various embodiments, the trained computational model is not a generic variant effect predictor, but rather a specialized transformer architecture trained on experimental measurements of pairwise variant functional activity' to leam the structural rules governing complementation across multimeric proteins, enabling high-fidelity prediction of whether two variants will complement one another without requiring explicit three-dimensional protein structure as input. By training on pairwise functional activity data and using local structural context representations derived from a pretrained protein language model, the model learns to predict complementation conditioned on the structural neighborhood of each variant residue position, enabling predictions that reflect spatial patterns of variant interactions that are obscured by additivity assumptions in conventional approaches. Complementation predictions generate outputs at higher resolution and biological specificity' than conventional single-variant approaches, as demonstrated by accurate classification of variant pairs whose individual effects are each strongly deleterious but whose combination restores substantial functional activity. The model's ability to integrate local structural context representations from two variants simultaneously enables systematic interrogation of how the spatial positions of variants within the multimeric protein determine their functional interaction across large variant libraries. This computational framework can evaluate thousands to millions of pairwise variant combinations, generate complementation maps across entire variant libraries, clusterPNDRI.020WO PATENTvariant positions into spatially distinct complementation groups, and enable automated predictions for clinical variant interpretation, patient stratification, and identification of compound heterozygous genotypes associated with milder or more severe disease phenotypes. These are tasks that require processing scale, speed, and numerical precision far beyond human mental steps and that provide actionable clinical insights not achievable through manual inspection of variant data or experimental characterization of individual variant pairs alone. Furthermore, this computational framework avoids having to experimentally characterize every possible pairwise variant combination, thereby avoiding the significant costs associated with high-throughput functional assays, including reagent costs, sequencing costs, and other costs associated with performing large-scale wet lab experiments.

[0092] The disclosed methods provide a technical improvement in computational protein analysis by generating local structural context representations directly from sequence data using pretrained protein language models, thereby eliminating the need for experimentally derived or predicted three-dimensional structures. This enables efficient and scalable prediction of inter-residue functional complementation across structurally similar but sequence-divergent proteins.

[0093] Unlike conventional approaches that rely on explicit structural coordinates or additive models, the claimed techniques model non-additive interactions between residue positions using learned representations and a trained computational model, thereby improving predictive accuracy and generalization across proteins.Overview of the Predictive Model

[0094] Although sporadic examples of complementation consistent with the variant sequestration model have been observed in several multimeric proteins, the scope and scale of this phenomenon have yet to be comprehensively characterized across the human proteome, and existing computational tools are not designed to predict pairwise variant interactions of this kind. Embodiments of the invention therefore involve predicting a priori whether two variants of a multimeric protein will complement one another upon multimer assembly, and quantifying the strength of their functional interaction. A model using the human argininosuccinate lyase (ASL) protein was chosen as a primary training and validation system. The ASL protein forms a homotetramer displaying the structural features required for variant sequestration, wherein each active site is formed by residuesPNDRI.020WO PATENTcontributed by more than one subunit, such that variants at spatially distinct residue positions are segregated into different active sites upon oligomerization.

[0095] FIG. IB shows a surface representation of human ASL protein (PDB: 1AOS) monomers and a homotetramer. The active site residues are darker grey on the grey monomer. ASL ligands, argininosuccinate (AlphaFill: P04424), are shown as sticks relative to the grey monomer and active site residues. Each monomer has three local regions of active and ligand-binding sites. Oligomerization of four monomers forms a homotetramer having four active and ligand-binding sites, each with a contribution from three monomers.

[0096] ASL is a clinically important enzyme. In the liver, ASL is part of the urea cycle that detoxifies waste nitrogen and converts it into excretable urea. In non-hepatic tissues, ASL takes part in de novo synthesis of L-arginine, and in specific tissues, the citrulline-nitric oxide cycle for cell-autonomous arginine synthesis and nitric oxide synthesis. Deficiency of ASL causes the urea cycle disorder argininosuccinic aciduria (ASA), a rare autosomal recessive disease. Because ASA patients are frequently compound heterozy gous, accurate prediction of whether two loss-of-function variants will complement one another is directly relevant to predicting the phenotypic severity of compound heterozygous genotypes in affected subjects.

[0097] Using high-throughput functional measurements of pairwise variant combinations in ASL, embodiments of the invention test and validate the variant sequestration model of complementation at scale. As described in more detail in the Examples section below, it was discovered that more than 60% (2,219 out of 3,599) of the loss-of-function variant combinations that were tested displayed substantial restoration of ASL activity when different variants were combined in an in vitro yeast assay. Consistent with the variant sequestration model, the patern of complementation was found to be dependent primarily on the position of the amino acid pairs in the protein structure, and not on the biochemical identities of the substituted amino acids.

[0098] The structural features that enable variant sequestration were found to be common among multimeric proteins possessing multichain active sites, and therefore complementation may affect a large number of variant combinations across a large number of human proteins. For example, it was estimated that variant sequestration could potentially affect approximately 4-6% of all genes in the human genome. Because complementation produces non-additive functional outcomes, models that assume additivity of individual variant effects will systematically mispredict the phenotypic consequences of compound heterozy gous genotypes in this class of proteins.PNDRI.020WO PATENT

[0099] Because of the scale of this problem and its implications for human health, embodiments of the invention relate to the development of machine learning models to predict complementation or non-complementation between pairs of variants of a multimeric protein. For example, one model was trained to leam the structural rules governing complementation in ASL. Using this model, it was found that it was possible to predict complementation between different ASL variant pairs with exceptionally high accuracy of 99.6%, including for variant combinations at residue positions not included in the training data. In addition, an ensemble model trained on ASL data was able to accurately predict (91%) complementation in a structurally similar but evolutionarily divergent protein, fumarase (FH), which shares only approximately 21% amino acid sequence identity with ASL. Accordingly, embodiments are not limited to predicting complementation between variants of ASL and FH, but may be extended to any multimeric protein possessing multichain functional sites, using experimentally-derived data on the functional impact of variant combinations from a relatively parsimonious set of diverse structures to predict this large class of genetic interactions.

[0100] Disclosed herein is an approach that performs pairwise variant-level analysis using local structural context representations to predict complementation or noncomplementation between variants of a multimeric protein. In some embodiments, complementation between variants is governed by the spatial positions of the variant residues within the multimeric protein complex, such that variants at residue positions contributing to different multichain active site assemblies complement one another, while variants at residue positions contributing to the same active site assembly do not. In some embodiments, the multimeric protein is a homomultimeric protein, such as a homodimer, homotrimer, homotetramer, or higher-order homomultimer, though the approaches described herein may be applied to heteromultimeric proteins as well. However, the predictive model described herein may be used to predict complementation for variants of any multimeric protein possessing at least one functional site formed by residues contributed by more than one subunit, including without limitation enzymes, receptors, transporters, structural proteins, and regulatory proteins from any species.

[0101] The pairwise variant analysis models each variant as an element defined by its local structural context within the amino acid sequence of the multimeric protein. A local structural context representation is a numerical representation of the sequence and structural environment surrounding a variant residue position, derived from a pretrained protein language model, without requiring explicit three-dimensional protein structurePNDRI.020WO PATENTcoordinates as input. Local structural context representations are obtained independently for each variant in a pair, and both representations are provided as input to a trained computational model to generate a complementation prediction output. In some embodiments, modelling pairwise variant interactions using local structural context representations enables determination of the individual contribution of each variant's structural neighborhood to the predicted functional interaction, as well as identification of the spatial cluster membership of variant residue positions within the multimeric protein. An understanding of spatial cluster membership can be useful, for example, in rational drug design, as variants within different spatial clusters that complement one another produce functional protein complexes that may be distinct drug targets from non-complementing variant combinations.

[0102] Generally, the disclosed methodology for predicting complementation between variants of a multimeric protein, also referred to herein as the predictive model, involves using variant input data to generate complementation prediction outputs for pairwise combinations of variants. Input data comprise the amino acid sequence of the multimeric protein of interest and identifiers of variants at respective residue positions, a subset of which may have experimental functional activity measurements for pairwise combinations available for model training or validation. In some embodiments, the input data further comprise three-dimensional structural information for the multimeric protein, used to identify candidate proteins for complementation or to validate predicted spatial cluster assignments.

[0103] The process begins with obtaining, for each variant, a local structural context representation by extracting a sequence window centered on the variant residue position and computing an embedding of that window using a pretrained protein language model. In some embodiments, the sequence window comprises from 10 to 50 amino acids centered on the respective variant residue position, for example a window of 21 amino acids. In some embodiments, the pretrained protein language model is ESM-2 or a functionally equivalent model, and the embedding comprises a vector of about 1,280 parameters encoding local structural context information from the amino acid sequence. In some embodiments, a local structural context representation is precomputed for each variant once and reused across multiple pairwise predictions, enabling efficient large-scale precomputation of complementation prediction outputs across entire variant libraries.

[0104] Before generating complementation predictions, parameters are chosen that determine, for each protein of interest: (1) which variant pairs will be evaluated; (2) whichPNDRI.020WO PATENTpretrained protein language model will be used to generate local structural context representations; (3) the length of the sequence window centered on each variant residue position; and (4) whether predictions will be generated for all pairwise combinations of a set of variants, a user-specified subset of variant pairs, or variants identified from sequencing data of a subject.

[0105] Model inference consists of running a trained computational model on the local structural context representations of both variants in each pair to generate a complementation prediction output. The trained computational model receives as input the local structural context representations of the first variant and the second variant, and outputs a complementation prediction comprising a continuous predicted growth value, a probability of complementation, or a binary complementation class label. In some embodiments, the trained computational model is a transformer-based neural network comprising an attention pooling layer and a prediction layer, with a Sigmoid function applied to generate a predicted growth value. The trained computational model does not use the numerical positions of the variant residues in the amino acid sequence as inputs, and does not require three-dimensional structural coordinates of the multimeric protein.

[0106] The trained computational model is trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein, obtained by a high-throughput functional assay such as a yeast-based growth assay or deep mutational scanning. In some embodiments, the dataset used for training comprises at least 1,000 pairwise variant combinations, for example at least 3,000 pairwise variant combinations. In some embodiments, the trained computational model is trained using cross-validation, including 20-fold cross-validation and leave-one-amino-acid-position-out cross-validation, and the final complementation prediction output is generated by an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.

[0107] A binary complementation classification is generated by applying a predetermined decision threshold to the complementation prediction output, such that predicted values above the threshold are classified as complementing pairs and predicted values below the threshold are classified as non-complementing pairs. In some embodiments, a complementing pair is defined as one exhibiting restoration of at least 50% of wild-type protein activity, and in some embodiments at least 60% of wild-type protein activity. For pairs classified as complementing, the predicted level of functional restoration may be further quantified by applying an additive model that accounts for the independentPNDRI.020WO PATENTallelic contributions of each variant. In some embodiments, combining the complementation classification with the additive model achieves an R2 of at least 0.99 in predicting quantitative functional outcomes for complementing pairs.

[0108] Once complementation predictions are generated for a plurality of pairwise combinations, the results may be organized and visualized in various ways. In some embodiments, a complementation map, heatmap, or interaction network is generated that identifies which variant combinations are predicted to complement one another across a protein of interest. In some embodiments, variants are clustered by residue position based on predicted or observed complementation patterns, thereby identifying two or more spatially distinct regions of the amino acid sequence that contribute to multichain active site assembly, wherein variants within the same local region do not complement each other and variants in different local regions do complement each other. Prior to clustering, dimensionality reduction may be applied to the complementation prediction matrix using standard techniques such as Principal Components Analysis or hierarchical clustering. In some embodiments, the dimensionality reduction is only ever fit on training data, and not on held-out data used for cross-validation, to prevent leakage of testing data into the classifier training sets.

[0109] The full space of complementation prediction outputs and variant cluster assignments may be visualized with standard techniques such as, and without limitation, t-Stochastic Neighbor Embedding (tSNE). Uniform Manifold Approximation and Projection (UMAP), or similar methods. Once two-dimensional or three-dimensional visualization coordinates are constructed, the complementation landscape can be visually inspected for clusters of residue positions and regions with distinct complementation behavior, as well as individual components of the local structural context representations that most strongly predict complementation.

[0110] As discussed herein, an understanding of complementation between variants of a multimeric protein can be useful for, for example, clinical variant interpretation and rational drug design. For example, a large-scale precomputation of complementation predictions across all pairwise combinations of known loss-of-function variants in a disease gene encoding a multimeric protein can be used to identify which compound heterozygous genotypes are predicted to produce a milder disease phenotype due to complementation, informing therapeutic decision-making including gene therapy, enzyme replacement therapy, protein replacement therapy, substrate reduction therapy, and small molecule targeting. Variant pairs predicted to complement one another, and thereby restore partial orPNDRI.02QWO PATENTfull protein function, may represent distinct therapeutic targets relative to noncomplementing pairs in which no functional restoration is predicted.Implementation of the Predictive Model

[0111] In some embodiments, the predictive model is designed to reveal pairwise variant interactions in multimeric proteins that determine complementation or noncomplementation, in contrast to methodologies that treat variant effects as additive or that analyze individual variant effects without accounting for pairwise structural context. In compound heterozygous individuals, many of the differences in phenotypic severity are attributable to whether the two variants present on opposite alleles complement one another upon multimer assembly, and the biological mechanisms underlying these differences are not accurately captured by single-variant effect prediction approaches. Self-supervised pretraining of protein language models on large corpora of protein sequence data, combined with supervised training on experimental measurements of pairwise variant functional activity, can reveal the structural rules governing complementation without requiring explicit three-dimensional protein structure as input. Using the predictive model as a discovery engine allows for the evaluation of pairwise variant interactions in a protein of interest without recourse to experimental characterization of every possible variant combination, which is particularly relevant for rare disease genes in which the number of known disease-associated variants is large relative to the number of variant combinations that have been experimentally characterized. Since the predictive model is trained on pairwise variant functional activity data and uses local structural context representations derived from a pretrained protein language model trained on tens of millions of protein sequences via self-supervised learning, it is applicable to diverse multimeric proteins and disease contexts, making it applicable even to proteins for which limited experimental complementation data are available, provided that sufficient sequence similarity exists to a protein on which the model was trained.

[0112] Accordingly, in various embodiments, the predictive model may be used to classify variant pairs as complementing or non-complementing, or to assign a continuous predicted growth value or probability of complementation for each pair, by providing local structural context representations of both variants as input to a trained computational model. The generative capacity of the predictive model allows predictions about precise pairwise variant interactions, identification of structural differences between variant positions that determine complementation, and large-scale precomputation ofPNDRI.020WO PATENTcomplementation landscapes across entire variant libraries. Precomputed complementation predictions can help stratify compound heterozygous genotypes by predicted phenotypic severity and inform understanding of why some compound heterozygous individuals exhibit milder disease phenotypes than expected from the individual effects of each variant alone, pointing toward better approaches for therapeutic positioning and patient stratification. Large-scale precomputation of complementation predictions across all pairwise combinations of known loss-of-function variants in a disease gene can be performed by computing local structural context representations for each variant once and evaluating all pairwise combinations using the trained computational model, with these predictions simulating the functional outcome of each possible compound heterozy gous genotype. The full space of complementation prediction outputs can be visually inspected for clusters of residue positions with distinct complementation behavior, using standard visualization techniques such as complementation maps, heatmaps, or two-dimensional embeddings generated by t-Stochastic Neighbor Embedding (tSNE) or Uniform Manifold Approximation and Projection (UMAP), enabling identification of spatially distinct local regions of the amino acid sequence that contribute to multichain active site assembly.Methods for Predicting Complementation or Non-Complementation Between Variants of a Multimeric Protein

[0113] Disclosed herein are methods, systems, and non-transitory computer readable media for predicting complementation or non-complementation between at least two variants of a multimeric protein.

[0114] FIG. 11 is an example flow diagram of the computer-implemented method for predicting complementation or non-complementation between at least two variants of a multimeric protein, in accordance with an embodiment. As shown in FIG. 11 , the method comprises: (i) obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site (step 310); (ii) identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence (step 320); (iii) obtaining, for each of the first variant and the second variant, a local structural context representation (step 330); (iv) determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output (step 340); and (v) classifying, from the complementation prediction output, the first variant and the second variant as a complementing pair or a non-PNDRI.020WO PATENTcomplementing pair (step 350). In some embodiments, the steps shown in FIG. 11 may be performed in any order, concurrently, or iteratively.

[0115] In some embodiments, the method further comprises, prior to step 310, identifying the multimeric protein as a candidate for complementation. In some embodiments, identifying the multimeric protein as a candidate for complementation comprises obtaining from a protein structure database at least one homomultimeric protein assembly comprising at least one ligand in contact with more than one subunit. In some embodiments, the step further comprises selecting at least one homomultimeric protein assembly in which a ligand contacts at least two identical subunits within a heavy atom-to-heavy atom distance of 6 A or less. In some embodiments, the ligand comprises more than 5 atoms and more than 50% of the atoms of the ligand are within 3.6 A of at least two protein chains of the homomultimeric assembly. In some embodiments, the method further comprises identifying additional candidate multimeric proteins by obtaining, from a protein structure database, a structurally similar candidate multimeric protein, wherein the structurally similar candidate multimeric protein has a structure similarity score of at least 0.5. In some embodiments, the method further comprises identifying human orthologs of candidate proteins from model organisms using an ortholog database.

[0116] With reference to FIG. 11, step 310 comprises obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site. In some embodiments, obtaining the amino acid sequence comprises retrieving the amino acid sequence from a sequence database, receiving user-specified input, or extracting the amino acid sequence from sequencing data. In some embodiments, the method further comprises preprocessing the amino acid sequence to identify active site residues, ligand-binding residues, or structurally important positions prior to identifying the first variant and the second variant.

[0117] In some embodiments, the multimeric protein is a homomultimeric protein comprising two or more identical subunits. In some embodiments, the homomultimeric protein is a homodimer, homotrimer, homotetramer, homopentamer, or higher-order homomultimer. In some embodiments, the multimeric protein is a heteromultimeric protein comprising at least two non-identical subunits. In some embodiments, the multimeric protein comprises at least two subunits. In some embodiments, the multimeric protein is an enzyme, receptor, transporter, structural protein, or regulatory protein. In some embodiments, the at least one functional site is a multichain active site, ligand-binding site,PNDRI.020WO PATENTor catalytic site formed by residues contributed by at least two subunits of the multimeric protein.

[0118] With reference to FIG. 11, step 320 comprises identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence. In some embodiments, identifying the first variant and the second variant comprises receiving user-specified variants, analyzing sequencing data from a subject, or computationally generating variant combinations from a set of known loss-of-function variants. In some embodiments, the first variant and the second variant are each loss-of-function variants exhibiting no more than 10% of wild-type protein activity when present alone. In some embodiments, the first variant and the second variant are each independently an amino acid substitution, insertion, or deletion relative to a reference sequence. In some embodiments, the first residue position and the second residue position are in different regions of the amino acid sequence, wherein the regions are defined structurally, empirically through experimental complementation data, or implicitly by a trained computational model. In some embodiments, the first residue position and the second residue position are in different spatial clusters of the protein structure, as determined by hierarchical clustering of complementation data or three-dimensional structural analysis. In some embodiments, the first residue position and the second residue position are each located at or near an active site, ligand-binding site, or catalytic residue of the multimeric protein. In some embodiments, the method is performed without experimental screening of variant combinations prior to prediction.

[0119] In some embodiments, the method further comprises identifying a plurality of additional variants and generating pairwise or higher-order complementation predictions. In some embodiments, the method further comprises identifying a plurality of additional variants at additional residue positions in the amino acid sequence, and determining complementation prediction outputs for a plurality of pairwise or higher-order combinations of variants. In some embodiments, the plurality of pairwise combinations comprises all pairwise combinations of loss-of-function variants at two or more active site or ligand-binding site residue positions.

[0120] With reference to FIG. 11, step 330 comprises obtaining, for each of the first variant and the second variant, a local structural context representation. In some embodiments, the local structural context representation is generated by extracting a sequence window centered on the respective variant residue position. In some embodiments, the sequence window comprises from 10 to 50 amino acids centered on thePNDRI.02QWO PATENTrespective variant residue position. In some embodiments, the sequence window comprises 21 amino acids. In some embodiments, the local structural context representation is generated by computing an embedding of the sequence window using a pretrained protein language model. In some embodiments, the pretrained protein language model is selected from the group consisting of ESM-2, ESM-lb, ESM-lv, ProtTrans, ProtBERT, TAPE, UniRep, SeqVec, an AlphaFold-based language model, and any functionally equivalent pretrained protein language model capable of generating local structural context representations of amino acid sequences. In some embodiments, the pretrained protein language model is ESM-2. In some embodiments, the embedding comprises a vector of about 500 to about 5,000 parameters. In some embodiments, the embedding comprises a vector of about 1,280 parameters. In some embodiments, the embedding encodes local structural context information from the amino acid sequence, without three-dimensional protein structure coordinates as input. In some embodiments, the local structural context representations of the first variant and the second variant are generated independently of one another. In some embodiments, the local structural context representations are precomputed and reused across multiple predictions.

[0121] With reference to FIG. 11, step 340 comprises determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output. In some embodiments, the complementation prediction output is determined by using a trained computational model that receives the local structural context representations of both the first variant and the second variant as input. In some embodiments, the trained computational model comprises a neural network configured to receive as input the concatenated local structural context representations of the first variant and the second variant. In some embodiments, the neural network is a transformer-based neural network. In some embodiments, the complementation prediction output is determined without using numerical positions of the first residue position or the second residue position in the amino acid sequence as inputs to the trained computational model. In some embodiments, complementation between the first variant and the second variant is determined by the first residue position and the second residue position in the protein structure and is independent of the biochemical identities of the substituted amino acids. In some embodiments, the complementation prediction output informs whether the first variant and the second variant produce a nonadditive functional outcome. In some embodiments, the complementation prediction output comprises a continuous predictedPNDRI.02QWO PATENTgrowth value, a probability of complementation, a binary complementation class label, or a direct classification or ranking generated without a separately computed interaction score.

[0122] In some embodiments, the trained computational model used in step 340 has been trained on experimental data. In some embodiments, the trained computational model has been trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein. In some embodiments, the experimental measurements of functional activity are obtained by a high-throughput yeast-based growth assay, deep mutational scanning, or other high-throughput functional screening method. In some embodiments, the experimental measurements of functional activity comprise measurements for at least 1,000 pairwise variant combinations. In some embodiments, the experimental measurements of functional activity comprise measurements for at least 3,000 pairwise variant combinations. In some embodiments, the trained computational model has been trained using cross-validation, including 20-fold cross-validation, leave-one-amino-acid-position-out cross-validation, or a combination thereof. In some embodiments, the trained computational model is an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.

[0123] In some embodiments, the trained computational model achieves high performance on test data. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.90 on test data. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.95 on test data from the same multimeric protein used to train the trained computational model. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.95 on test data comprising variant combinations at residue positions not included in the training data. In some embodiments, the trained computational model achieves an area under the receiver operating characteristic curve (AUC-ROC) of at least 0.85. In some embodiments, the trained computational model achieves an R2of at least 0.84.

[0124] In some embodiments, the trained computational model may be applied across structurally related proteins. In some embodiments, the trained computational model has been trained on variant combinations from a first multimeric protein and is applied to predict complementation in a second multimeric protein that is structurally similar to but sequence-divergent from the first multimeric protein. In some embodiments, the first multimeric protein and the second multimeric protein share less than 25% amino acid sequence identity. In some embodiments, the first multimeric protein and the second multimeric protein are members of the same structural family or superfamily. In somePNDRI.020WO PATENTembodiments, the trained computational model achieves a classification accuracy of at least 0.90 on variant combinations from the second multimeric protein.

[0125] With reference to FIG. 11, the final step of the method shown comprises classifying, from the complementation prediction output, the first variant and the second variant as a complementing pair or a non-complementing pair. In some embodiments, classifying comprises applying a predetermined decision threshold to the complementation prediction output. In some embodiments, the decision threshold is selected such that complementation prediction output above the threshold is classified as complementing pairs and complementation prediction output below the threshold is classified as noncomplementing pairs. In some embodiments, a complementing pair exhibits restoration of at least 50% of wild-type protein activity. In some embodiments, a complementing pair exhibits restoration of at least 60% of wild-type protein activity.

[0126] In some embodiments, the method further comprises quantifying the predicted level of functional restoration for pairs classified as complementing. In some embodiments, the method further comprises, for a pair classified as a complementing pair, quantifying the predicted level of functional restoration by applying an additive model, wherein the additive model accounts for the independent allelic contributions of the first variant and the second variant to the quantitative level of restored protein activity. In some embodiments, combining the complementation classification with the additive model achieves an R2of at least 0.99. In some embodiments, the method further comprises determining a predicted functional score for the first variant, determining a predicted functional score forthe second variant, and quantify ing the degree of nonadditivity by comparing the interaction score to an expected additive combination of the individual predicted functional scores.

[0127] In some embodiments, the method further comprises generating a visual representation of predicted complementation relationships across a plurality of variants. In some embodiments, the method further comprises generating a complementation map, heatmap, interaction network, or visual representation identifying which variant combinations are predicted to complement one another. In some embodiments, the method further comprises clustering the variants by residue position based on predicted or observed complementation patterns, thereby identifying two or more spatially different regions of the amino acid sequence that contribute to multichain active site assembly. In some embodiments, variants residing within the same local region do not complement each other, and variants residing in different local regions complement each other.PNDRI.020WO PATENT

[0128] In some embodiments, the classification generated by the method has dow nstream clinical and therapeutic utility. In some embodiments, the method further comprises using the classification to inform a therapeutic approach selected from the group consisting of gene therapy, cell therapy, enzyme replacement therapy, protein replacement therapy, substrate reduction therapy, and small molecule targeting. In some embodiments, the method further comprises using the classification to predict the phenotypic severity of a compound heterozygous genotype in a subject, wherein a classification of complementation pair is associated with a predicted milder disease phenotype relative to a homozygous loss-of-function genotype. In some embodiments, the method further comprises validating the classification by experimentally measuring functional activity of a compound heterozygous organism or cell expressing both the first variant and the second variant. In some embodiments, the compound heterozygous organism is a yeast diploid strain harboring the first variant on one allele and the second variant on a second allele, and wherein functional activity is measured by a growth assay under selective conditions.Methods for Predicting Complementation Between Variants of a Multimeric Protein

[0129] Disclosed herein are computer-implemented methods for predicting complementation between at least two variants of a multimeric protein. As shown generally in FIG. 10B, the method involves providing input data (200) to a Complementation Prediction Module (300), which generates a complementation prediction output that is communicated via an Output Module (400). In some embodiments, the method comprises: (i) obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site; (ii) identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence; (iii) obtaining, for each of the first variant and the second variant, a local structural context representation; (iv) determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output; and (v) classifying, from the complementation prediction output, the first variant and the second variant as a complementing pair. In some embodiments, the foregoing steps may be performed in any order, concurrently, or iteratively.

[0130] The method comprises obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site, which serves as input data (200) to the Complementation Prediction Module (300) as depicted generally in FIG. 10B. In some embodiments, obtaining the amino acid sequence comprises retrieving the aminoPNDRI.020WO PATENTacid sequence from a sequence database, receiving user-specified input, or extracting the amino acid sequence from sequencing data. In some embodiments, the method further comprises preprocessing the amino acid sequence to identify active site residues, ligandbinding residues, or structurally important positions prior to identifying the first variant and the second variant.

[0131] In some embodiments, the multimeric protein is a homomultimeric protein comprising two or more identical subunits. In some embodiments, the homomultimeric protein is a homodimer, homotrimer, homotetramer, homopentamer, or higher-order homomultimer. In some embodiments, the multimeric protein is a heteromultimeric protein comprising at least two non-identical subunits. In some embodiments, the multimeric protein is an enzyme, receptor, transporter, structural protein, or regulatory’ protein. In some embodiments, the at least one functional site is a multichain active site, ligand-binding site, or catalytic site formed by residues contributed by at least two subunits of the multimeric protein.

[0132] In some embodiments, prior to obtaining the amino acid sequence, the method further comprises a step of identifying the multimeric protein as a candidate for complementation. In some embodiments, identifying the multimeric protein as a candidate for complementation comprises obtaining from a protein structure database at least one homomultimeric protein assembly comprising at least one ligand in contact with more than one subunit. In some embodiments, identifying the multimeric protein as a candidate for complementation further comprises selecting at least one homomultimeric protein assembly in which a ligand contacts at least two identical subunits within a heavy atom-to-heavy atom distance of 6 A or less. In some embodiments, the ligand comprises more than 5 atoms and more than 50% of the atoms of the ligand are within 3.6 A of at least two protein chains of the homomultimeric assembly. In some embodiments, the method further comprises identifying additional candidate multimeric proteins by obtaining, from a protein structure database, a structurally similar candidate multimeric protein, wherein the structurally similar candidate multimeric protein has a structure similarity’ score of at least 0.5. In some embodiments, the method further comprises identifying human orthologs of candidate proteins from model organisms using an ortholog database.

[0133] The method further comprises identify ing a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence. In some embodiments, identifying the first variant and the second variant comprises receiving user-specified variants, analyzing sequencing data from a subject, or computationally generatingPNDRI.020WO PATENTvariant combinations from a set of known loss-of-function variants. In some embodiments, the method is performed without experimental screening of variant combinations prior to prediction.

[0134] In some embodiments, the first variant and the second variant are each loss-of-function variants exhibiting no more than 10% of wild-type protein activity7when present alone. In some embodiments, the first variant and the second variant are each independently an amino acid substitution, insertion, or deletion relative to a reference sequence. In some embodiments, the first residue position and the second residue position are each located at or near an active site, ligand-binding site, or catalytic residue of the multimeric protein.

[0135] In some embodiments, the first residue position and the second residue position are in different regions of the amino acid sequence, wherein the regions are defined structurally, empirically through experimental complementation data, or implicitly by a trained computational model. In some embodiments, the first residue position and the second residue position are in different spatial clusters of the protein structure, as determined by hierarchical clustering of complementation data or three-dimensional structural analysis. In some embodiments, complementation between the first variant and the second variant is determined by the first residue position and the second residue position in the protein structure and is independent of the biochemical identities of the substituted amino acids.

[0136] In some embodiments, the method further comprises identifying a plurality7of additional variants at additional residue positions in the amino acid sequence, and determining complementation prediction outputs for a plurality of pairwise or higher-order combinations of variants. In some embodiments, the plurality of pairwise combinations comprises all pairwise combinations of loss-of-function variants at two or more active site or ligand-binding site residue positions.

[0137] The method further comprises obtaining, for each of the first variant and the second variant, a local structural context representation, which is processed by the Complementation Prediction Module (300) as depicted generally in FIG. 10B. In some embodiments, the local structural context representation is generated by extracting a sequence window centered on the respective variant residue position. In some embodiments, the sequence window comprises from 10 to 50 amino acids centered on the respective variant residue position. In some embodiments, the sequence window comprises 21 amino acids.PNDRI.02QWO PATENT

[0138] In some embodiments, the local structural context representation is generated by computing an embedding of the sequence window using a pretrained protein language model. In some embodiments, the pretrained protein language model is selected from the group consisting of ESM-2, ESM-lb, ESM-lv, ProtTrans, ProtBERT, TAPE, UniRep, SeqVec, an AlphaFold-based language model, and any functionally equivalent pretrained protein language model capable of generating local structural context representations of amino acid sequences. In some embodiments, the pretrained protein language model is ESM-2.

[0139] In some embodiments, the embedding comprises a vector of about 500 to about 5,000 parameters. In some embodiments, the embedding comprises a vector of about 1,280 parameters. In some embodiments, the embedding encodes local structural context information from the amino acid sequence, without three-dimensional protein structure coordinates as input. In some embodiments, the local structural context representations of the first variant and the second variant are generated independently of one another. In some embodiments, the local structural context representations are precomputed and reused across multiple predictions.

[0140] The method further comprises determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output. As depicted generally in FIG. 10B, this determination is performed by the Complementation Prediction Module (300). In some embodiments, the complementation prediction output is determined by using a trained computational model that receives the local structural context representations of both the first variant and the second variant as input. In some embodiments, the trained computational model comprises a neural network configured to receive as input the concatenated local structural context representations of the first variant and the second variant. In some embodiments, the neural network is a transformer-based neural network. In some embodiments, the complementation prediction output is determined without using numerical positions of the first residue position or the second residue position in the amino acid sequence as inputs to the trained computational model.

[0141] In some embodiments, the complementation prediction output informs whether the first variant and the second variant produce a nonadditive functional outcome. In some embodiments, the complementation prediction output comprises a continuous predicted growth value, a probability of complementation, a binary complementation class label, or a direct classification or ranking generated without a separately computed interaction score.PNDRI.020WO PATENT

[0142] In some embodiments, the trained computational model has been trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein. In some embodiments, the experimental measurements of functional activity are obtained by a high-throughput yeast-based growth assay, deep mutational scanning, or other high-throughput functional screening method. In some embodiments, the experimental measurements of functional activity comprise measurements for at least 1 ,000 pairwise variant combinations. In some embodiments, the experimental measurements of functional activity comprise measurements for at least 3,000 pairwise variant combinations.

[0143] In some embodiments, the trained computational model has been trained using cross-validation, including 20-fold cross-validation, leave-one-amino-acid-position-out cross-validation, or a combination thereof. In some embodiments, the trained computational model is an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.

[0144] In some embodiments, the trained computational model achieves a classification accuracy of at least 0.90 on test data. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.95 on test data from the same multimeric protein used to train the trained computational model. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.95 on test data comprising variant combinations at residue positions not included in the training data. In some embodiments, the trained computational model achieves an area under the receiver operating characteristic curve (AUC-ROC) of at least 0.85. In some embodiments, the trained computational model achieves an R2of at least 0.84.

[0145] In some embodiments, the trained computational model may be applied across structurally related proteins. In some embodiments, the trained computational model has been trained on variant combinations from a first multimeric protein and is applied to predict complementation in a second multimeric protein that is structurally similar to but sequence-divergent from the first multimeric protein. In some embodiments, the first multimeric protein and the second multimeric protein share less than 25% amino acid sequence identity. In some embodiments, the first multimeric protein and the second multimeric protein are members of the same structural family or superfamily. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.90 on variant combinations from the second multimeric protein.

[0146] The method further comprises classifying, from the complementation prediction output, the first variant and the second variant as a complementing pair. As depictedPNDRI.020WO PATENTgenerally in FIG. 10B, the classification result is communicated via the Output Module (400). In some embodiments, classifying comprises applying a predetermined decision threshold to the complementation prediction output, wherein a complementation prediction output above the threshold is classified as a complementing pair. In some embodiments, a complementing pair exhibits restoration of at least 50% of wild-ty pe protein activity. In some embodiments, a complementing pair exhibits restoration of at least 60% of wild-type protein activity.

[0147] In some embodiments, the method further comprises quantitative analysis following classification. In some embodiments, for a pair classified as a complementing pair, the method further comprises quantify ing the predicted level of functional restoration by applying an additive model, wherein the additive model accounts for the independent allelic contributions of the first variant and the second variant to the quantitative level of restored protein activity. In some embodiments, combining the complementation classification with the additive model achieves an R2of at least 0.99. In some embodiments, the method further comprises determining a predicted functional score for the first variant, determining a predicted functional score for the second variant, and quantifying the degree of nonadditivity by7comparing the interaction score to an expected additive combination of the individual predicted functional scores.

[0148] In some embodiments, the method further comprises generating a visual representation of predicted complementation relationships across a plurality’ of variants. In some embodiments, the method further comprises generating a complementation map, heatmap, interaction network, or visual representation identifying which variant combinations are predicted to complement one another. In some embodiments, the method further comprises clustering the variants by residue position based on predicted or observed complementation patterns, thereby identifying two or more spatially different regions of the amino acid sequence that contribute to multichain active site assembly, wherein variants residing within the same local region do not complement each other, and variants residing in different local regions complement each other.

[0149] In some embodiments, the classification generated by the method has downstream clinical and therapeutic utility. In some embodiments, the method further comprises using the classification to inform a therapeutic approach selected from the group consisting of gene therapy, cell therapy, enzyme replacement therapy, protein replacement therapy, substrate reduction therapy, and small molecule targeting. In some embodiments, the method further comprises using the classification to predict the phenoty pic severity' of aPNDRI.020WO PATENTcompound heterozy gous genoty pe in a subject, wherein a classification of complementing pair is associated with a predicted milder disease phenotype relative to a homozygous loss-of-function genotype. In some embodiments, the method further comprises validating the classification by experimentally measuring functional activity of a compound heterozygous organism or cell expressing both the first variant and the second variant. In some embodiments, the compound heterozygous organism is a yeast diploid strain harboring the first variant on one allele and the second variant on a second allele, and wherein functional activity^ is measured by a grow th assay' under selective conditions.Methods for Predicting Non-Complementation Between Variants of a Multimeric Protein

[0150] Disclosed herein are computer-implemented methods for predicting noncomplementation between at least two variants of a multimeric protein. As shown generally in FIG. 10B, the method involves providing input data (200) to a Complementation Prediction Module (300), which generates a complementation prediction output that is communicated via an Output Module (400). In some embodiments, the method comprises: (i) obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site; (ii) identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence; (iii) obtaining, for each of the first variant and the second variant, a local structural context representation; (iv) determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output; and (v) classifying, from the complementation prediction output, the first variant and the second variant as a non-complementing pair. In some embodiments, the foregoing steps may be performed in any order, concurrently, or iteratively.

[0151] The method comprises obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site, which serves as input data (200) to the Complementation Prediction Module (300) as depicted generally' in FIG. 10B. In some embodiments, obtaining the amino acid sequence comprises retrieving the amino acid sequence from a sequence database, receiving user-specified input, or extracting the amino acid sequence from sequencing data. In some embodiments, the method further comprises preprocessing the amino acid sequence to identify active site residues, ligandbinding residues, or structurally important positions prior to identify ing the first variant and the second variant.PNDRI.020WO PATENT

[0152] In some embodiments, the multimeric protein is a homomultimeric protein comprising two or more identical subunits. In some embodiments, the homomultimeric protein is a homodimer, homotrimer, homotetramer, homopentamer, or higher-order homomultimer. In some embodiments, the multimeric protein is a heteromultimeric protein comprising at least two non-identical subunits. In some embodiments, the multimeric protein is an enzyme, receptor, transporter, structural protein, or regulatory’ protein. In some embodiments, the at least one functional site is a multichain active site, ligand-binding site, or catalytic site formed by residues contributed by at least two subunits of the multimeric protein.

[0153] In some embodiments, prior to obtaining the amino acid sequence, the method further comprises a step of identifying the multimeric protein as a candidate for noncomplementation analysis. In some embodiments, identifying the multimeric protein as a candidate comprises obtaining from a protein structure database at least one homomultimeric protein assembly comprising at least one ligand in contact with more than one subunit. In some embodiments, identifying the multimeric protein as a candidate further comprises selecting at least one homomultimeric protein assembly in which a ligand contacts at least two identical subunits within a heavy atom-to-heavy atom distance of 6 A or less. In some embodiments, the ligand comprises more than 5 atoms and more than 50% of the atoms of the ligand are within 3.6 A of at least two protein chains of the homomultimeric assembly. In some embodiments, the method further comprises identify ing additional candidate multimeric proteins by obtaining, from a protein structure database, a structurally similar candidate multimeric protein, wherein the structurally similar candidate multimeric protein has a structure similarity score of at least 0.5. In some embodiments, the method further comprises identifying human orthologs of candidate proteins from model organisms using an ortholog database.

[0154] The method further comprises identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence. In some embodiments, identifying the first variant and the second variant comprises receiving user-specified variants, analyzing sequencing data from a subject, or computationally generating variant combinations from a set of known loss-of-function variants. In some embodiments, the method is performed without experimental screening of variant combinations prior to prediction.

[0155] In some embodiments, the first variant and the second variant are each loss-of-function variants exhibiting no more than 10% of wild-type protein activity when presentPNDRI.020WO PATENTalone. In some embodiments, the first variant and the second variant are each independently an amino acid substitution, insertion, or deletion relative to a reference sequence. In some embodiments, the first residue position and the second residue position are each located at or near an active site, ligand-binding site, or catalytic residue of the multimeric protein.

[0156] In some embodiments, the first residue position and the second residue position are in the same region of the amino acid sequence, wherein the region is defined structurally, empirically through experimental complementation data, or implicitly by a trained computational model. In some embodiments, the first residue position and the second residue position are within the same spatial cluster of the protein structure, as determined by hierarchical clustering of complementation data or three-dimensional structural analysis. In some embodiments, non-complementation between the first variant and the second variant is determined by the first residue position and the second residue position being located within the same structural region of the protein and is independent of the biochemical identities of the substituted amino acids.

[0157] In some embodiments, the method further comprises identifying a plurality of additional variants at additional residue positions in the amino acid sequence, and determining non-complementation prediction outputs for a plurality of pairwise or higher-order combinations of variants. In some embodiments, the plurality7of pairwise combinations comprises all pairwise combinations of loss-of-function variants at two or more active site or ligand-binding site residue positions within the same structural region.

[0158] The method further comprises obtaining, for each of the first variant and the second variant, a local structural context representation, which is processed by the Complementation Prediction Module (300) as depicted generally in FIG. 10B. In some embodiments, the local structural context representation is generated by extracting a sequence window centered on the respective variant residue position. In some embodiments, the sequence window7comprises from 10 to 50 amino acids centered on the respective variant residue position. In some embodiments, the sequence window7comprises 21 amino acids.

[0159] In some embodiments, the local structural context representation is generated by computing an embedding of the sequence window using a pretrained protein language model. In some embodiments, the pretrained protein language model is selected from the group consisting of ESM-2. ESM-lb, ESM-lv, ProtTrans, ProtBERT, TAPE, UniRep, SeqVec, an AlphaFold-based language model, and any functionally equivalent pretrained protein language model capable of generating local structural context representations ofPNDRI.02QWO PATENTamino acid sequences. In some embodiments, the pretrained protein language model is ESM-2.

[0160] In some embodiments, the embedding comprises a vector of about 500 to about 5,000 parameters. In some embodiments, the embedding comprises a vector of about 1,280 parameters. In some embodiments, the embedding encodes local structural context information from the amino acid sequence, without three-dimensional protein structure coordinates as input. In some embodiments, the local structural context representations of the first variant and the second variant are generated independently of one another. In some embodiments, the local structural context representations are precomputed and reused across multiple predictions.

[0161] The method further comprises determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output. As depicted generally in FIG. 10B, this determination is performed by the Complementation Prediction Module (300). In some embodiments, the complementation prediction output is determined by using a trained computational model that receives the local structural context representations of both the first variant and the second vanant as input. In some embodiments, the trained computational model comprises a neural network configured to receive as input the concatenated local structural context representations of the first variant and the second variant. In some embodiments, the neural network is a transformer-based neural network. In some embodiments, the complementation prediction output is determined without using numerical positions of the first residue position or the second residue position in the amino acid sequence as inputs to the trained computational model.

[0162] In some embodiments, the complementation prediction output informs whether the first variant and the second variant fail to produce a nonadditive functional outcome, thereby indicating non-complementation. In some embodiments, the complementation prediction output comprises a continuous predicted grow th value, a probability of noncomplementation, a binary non-complementation class label, or a direct classification or ranking generated without a separately computed interaction score.

[0163] In some embodiments, the trained computational model has been trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein. In some embodiments, the experimental measurements of functional activity are obtained by a high-throughput yeast-based growth assay, deep mutational scanning, or other high-throughput functional screening method. In some embodiments, thePNDRI.020WO PATENTexperimental measurements of functional activity' comprise measurements for at least 1 ,000 pairwise variant combinations. In some embodiments, the experimental measurements of functional activity comprise measurements for at least 3,000 pairwise variant combinations.

[0164] In some embodiments, the trained computational model has been trained using cross-validation, including 20-fold cross-validation, leave-one-amino-acid-position-out cross-validation, or a combination thereof. In some embodiments, the trained computational model is an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.

[0165] In some embodiments, the trained computational model achieves a classification accuracy of at least 0.90 on test data. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.95 on test data from the same multimeric protein used to train the trained computational model. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.95 on test data comprising variant combinations at residue positions not included in the training data. In some embodiments, the trained computational model achieves an area under the receiver operating characteristic curve (AUC-ROC) of at least 0.85. In some embodiments, the trained computational model achieves an R2of at least 0.84.

[0166] In some embodiments, the trained computational model may be applied across structurally related proteins. In some embodiments, the trained computational model has been trained on variant combinations from a first multimeric protein and is applied to predict non-complementation in a second multimeric protein that is structurally similar to but sequence-divergent from the first multimeric protein. In some embodiments, the first multimeric protein and the second multimeric protein share less than 25% amino acid sequence identity. In some embodiments, the first multimeric protein and the second multimeric protein are members of the same structural family or superfamily. In some embodiments, the trained computational model achieves a classification accuracy of at least 0.90 on variant combinations from the second multimeric protein.

[0167] The method further comprises classifying, from the complementation prediction output, the first variant and the second variant as a non-complementing pair. As depicted generally in FIG. 10B, the classification result is communicated via the Output Module (400). In some embodiments, classifying comprises applying a predetermined decision threshold to the complementation prediction output, wherein a complementation prediction output below the threshold is classified as a non-complementing pair. In some embodiments, a non-complementing pair fails to exhibit restoration of wild-type proteinPNDRI.020WO PATENTactivity above background levels. In some embodiments, a non-complementing pair exhibits less than 50% restoration of wild-type protein activity. In some embodiments, a non-complementing pair exhibits less than 10% restoration of wild-type protein activity.

[0168] In some embodiments, the method further comprises quantitative analysis following classification. In some embodiments, for a pair classified as a noncomplementing pair, the method further comprises quantifying the predicted level of functional impairment by applying an additive model, wherein the additive model accounts for the independent allelic contributions of the first variant and the second variant to the quantitative level of lost protein activity. In some embodiments, combining the noncomplementation classification with the additive model achieves an R2of at least 0.99. In some embodiments, the method further comprises determining a predicted functional score for the first variant, determining a predicted functional score for the second variant, and confirming additivity by comparing the interaction score to an expected additive combination of the individual predicted functional scores.

[0169] In some embodiments, the method further comprises generating a visual representation of predicted non-complementation relationships across a plurality of variants. In some embodiments, the method further comprises generating a noncomplementation map, heatmap, interaction network, or visual representation identifying which variant combinations are predicted not to complement one another. In some embodiments, the method further comprises clustering the variants by residue position based on predicted or observed non-complementation patterns, thereby identifying one or more spatially defined regions of the amino acid sequence within which variants fail to complement one another, wherein variants residing within the same local region do not complement each other.

[0170] In some embodiments, the classification generated by the method has downstream clinical and therapeutic utility. In some embodiments, the method further comprises using the classification to inform a therapeutic approach selected from the group consisting of gene therapy, cell therapy, enzyme replacement therapy, protein replacement therapy, substrate reduction therapy, and small molecule targeting. In some embodiments, the method further comprises using the classification to predict the phenotypic severity of a compound heterozygous genotype in a subject, wherein a classification of noncomplementing pair is associated with a predicted severe disease phenotype comparable to a homozygous loss-of-function genotype. In some embodiments, the method further comprises validating the classification by experimentally measuring functional activity ofPNDRI.020WO PATENTa compound heterozy gous organism or cell expressing both the first variant and the second variant. In some embodiments, the compound heterozygous organism is a yeast diploid strain harboring the first variant on one allele and the second variant on a second allele, and wherein functional activity7is measured by a growth assay under selective conditions.Optimizing Computational Resources for Variant Pair Evaluation

[0171] Disclosed herein are computer-implemented methods, systems, and non-transitory computer readable media for optimizing computational resources for predicting complementation or non-complementation between variants of a multimeric protein at scale. In some embodiments, such computational optimization enables the evaluation of variant pair combinations that are too numerous, structurally complex, or otherwise intractable using conventional experimental or brute-force computational approaches without the embedding-guided methods provided herein.

[0172] In some embodiments, the computer-implemented method for optimizing computational resources for complementation prediction comprises receiving, by one or more processors, amino acid sequence data derived from one or more multimeric proteins. In some embodiments, the amino acid sequence data comprises user-specified variant inputs, computationally generated variant combinations derived from a set of known loss-of-function variants, sequencing data from one or more subjects, or any combination thereof, such as those provided for herein.

[0173] In some embodiments, the computer-implemented method further comprises obtaining local structural context representations corresponding to each variant of interest in the one or more multimeric proteins. In some embodiments, the local structural context representations are derived from sequence window embeddings generated by a pretrained protein language model, such as those generated according to the methods provided for herein, including without limitation the methods for obtaining local structural context representations described herein. In some embodiments, the local structural context representations are derived from variants identified across at least 10, at least 50, at least 100, at least 500, at least 1,000, at least 3,000, or more variant positions. In some embodiments, the local structural context representations derived from a plurality of variant positions are precomputed and stored, such that they may be reused across multiple pairwise or higher-order complementation prediction queries without requiring redundant computation.PNDRI.020WO PATENT

[0174] In some embodiments, the computer-implemented method further comprises generating a representation of pairwise relationships among the variants of interest using the local structural context representations. In some embodiments, the pairwise representation encodes candidate variant pair combinations for complementation evaluation, wherein each pair corresponds to a candidate complementing or noncomplementing combination of the first variant and the second variant. In some embodiments, the pairwise representation is generated from precomputed local structural context representations that encode position-specific and sequence-context information from the amino acid sequence without requiring three-dimensional protein structure coordinates as input.

[0175] In some embodiments, the computer-implemented method further comprises reducing a computational search space associated with evaluation of candidate variant pair combinations, relative to the computational search space that would be required for exhaustive experimental or brute-force computational screening of all possible variant combinations. In some embodiments, the local structural context representations constrain the set of candidate variant pairs that must be evaluated by the one or more processors, thereby reducing the number of computational operations required to identify complementing pairs. In some embodiments, reducing the computational search space comprises generating a vector of probabilities calculated from the local structural context representations and, optionally, observed experimental data, wherein each element of the vector corresponds to the probability that a given variant pair exhibits complementation. In some embodiments, low-probability variant pairs are deprioritized or excluded from further evaluation, thereby reducing the computational search space. In some embodiments, the probabilities are derived from population-level variant frequencies, structural clustering of residue positions, recombination or co-occurrence data, or other statistical models of variant interactions embedded in the trained computational model.

[0176] In some embodiments, the computer-implemented method further comprises generating a vector of pairwise complementation probabilities calculated from the local structural context representations and observed experimental data. In some embodiments, each element of the vector of pairwise complementation probabilities represents the probability that a given variant pair exhibits complementation in the multimeric protein. In some embodiments, the vector of pairwise complementation probabilities is used during evaluation operations to preferentially assess high-probability complementing pairs and to exclude low-probability pairs from further evaluation, thereby further reducing thePNDRI.020WO PATENTcomputational search space. In some embodiments, the vector of pairwise complementation probabilities is calculated by combining the probability of each variant pair as determined from the local structural context representations with any available observed experimental measurements of functional activity supporting each pair. In some embodiments, the vector of pairwise complementation probabilities enables classification of variant pairs as complementing or non-complementing, as described herein with respect to the methods for predicting complementation and non-complementation.

[0177] In some embodiments, the computer-implemented method further comprises performing evaluation operations on the pairwise representation with the local structural context representations to generate one or more complementation prediction outputs for the variant pairs of interest. In some embodiments, the evaluation operations comprise inputting the concatenated local structural context representations of the first variant and the second variant into a trained computational model in a manner guided by the vector of probabilities and / or the vector of pairwise complementation probabilities, such that the resulting output corresponds to a high-confidence complementation prediction. In some embodiments, the evaluation operations are performed by a trained neural network, such as a transformer-based neural network, that processes pairwise embeddings without requiring the numerical residue positions of the first variant or the second variant as inputs. In some embodiments, the one or more complementation prediction outputs correspond to continuous predicted functional activity values, binary complementation class labels, or probability scores for each evaluated variant pair.

[0178] In some embodiments, the computer-implemented method further comprises separating the evaluated variant pairs into two or more classes based on the complementation prediction outputs. In some embodiments, separating the variant pairs into two or more classes comprises assigning each variant pair a complementation class label based on the vector of painvise complementation probabilities, the observed experimental data, and the local structural context representations. In some embodiments, the two or more classes correspond to complementing pairs and non-complementing pairs. In some embodiments, separating the variant pairs into two or more classes is performed by applying a predetermined decision threshold to the complementation prediction output, such as those provided for herein with respect to the methods for predicting complementation and non-complementation.

[0179] Without wishing to be bound by a particular theory, the computational resources required for evaluation of all possible variant pair combinations in a multimeric protein arePNDRI.020WO PATENTsubstantially reduced by the methods provided for herein as compared to approaches that rely on exhaustive experimental screening or brute-force computational evaluation of all possible combinations. Specifically, in proteins with a large number of loss-of-function variants, the total number of candidate pairwise combinations grows combinatorially and rapidly becomes intractable for experimental characterization. By using precomputed local structural context representations embedded in the trained computational model to calculate vectors of probabilities and pairwise complementation probabilities, the methods provided for herein substantially reduce the number of candidate variant pairs that must be evaluated by the one or more processors, reducing both computational time and experimental resource requirements.

[0180] In some embodiments, the methods for optimizing computational resources for complementation prediction provided for herein are performed on one or more computers, such as the computer system described herein. In some embodiments, the steps of the methods are performed in silico. In some embodiments, the methods are implemented in software as one or more program modules stored on a non-transitory computer readable medium and executed by one or more processors..Methods for Generating a Machine-Readable Protein Variant Data Structure

[0181] Disclosed herein are computer-implemented methods for generating a machine-readable protein variant data structure from identified variant positions and their associated complementation prediction outputs, such as those generated using the methods for predicting complementation and non-complementation provided for herein. In some embodiments, the machine-readable protein variant data structure is a structured complementation data record. In some embodiments, the structured complementation data record enables representation of pairwise variant complementation relationships without duplication of sequence embedding data and with reduced computational processing required for accessing complementation prediction information, relative to approaches that require recomputation of embeddings and predictions for each query.

[0182] In some embodiments, the computer-implemented method for generating a machine-readable protein variant data structure comprises encoding a pairwise variant complementation relationship without reliance on a fixed reference protein sequence. In some embodiments, encoding a pairwise variant complementation relationship without reliance on a fixed reference protein sequence means that the variant records stored in the structured complementation data record are defined relative to the local structural contextPNDRI.020WO PATENTrepresentations of the variants themselves, rather than relative to the coordinates of a single fixed linear reference sequence. In some embodiments, this reference-free encoding enables representation of complementation relationships involving novel variant alleles that are absent from standard protein databases, including variants discovered through population-scale sequencing or deep mutational scanning, such as those identified using the methods provided for herein.

[0183] In some embodiments, the computer-implemented method for generating a machine-readable protein variant data structure further comprises encoding pairwise complementation relationships and position-level variant information occurring within functionally defined regions within coordinated record structures that preserve the sequence diversity of the variant positions. In some embodiments, coordinated record structures are organized units of the data structure that jointly encode a pairwise complementation relationship and all position-level variant information occurring at or flanking the variant residue positions in a manner that preserves the local structural context and sequence information of all variants. In some embodiments, the coordinated record structures enable representation of complex variant interactions, including higher-order combinations of variants across multiple residue positions, without loss of sequence diversity information.

[0184] In some embodiments, the computer-implemented method for generating a machine-readable protein variant data structure further comprises assigning each record a coordinate within a virtual residue coordinate system that preserves linear sequence locality and enables positional indexing of the records. In some embodiments, the virtual residue coordinate system is derived from the amino acid sequence of the multimeric protein and the local structural context representations of the variant positions, such that records at adjacent residue positions are assigned adjacent or nearby coordinates in the virtual residue coordinate system. In some embodiments, the coordinates assigned within the virtual residue coordinate system are compatible with standard sequence interval arithmetic, enabling positional queries, windowed analyses, and linkage or co-occurrence calculations using the structured complementation data record.

[0185] In some embodiments, the computer-implemented method for generating a machine-readable protein variant data structure further comprises storing the data structure in a format compatible with standard variant annotation and parsing software. In some embodiments, the data structure is stored in a file format that can be read and processed by standard variant annotation tools, thereby enabling interoperability of the structuredPNDRI.020WO PATENTcomplementation data record with existing protein variant analysis pipelines. In some embodiments, the structured complementation data record comprises a plurality of structured records, wherein each record encodes the pairwise complementation relationship between a first variant and a second variant, including without limitation the local structural context representations of each variant, the virtual residue coordinates of each variant position, the complementation prediction output, the complementation class label, and the identities of the variants for which each prediction was generated.

[0186] In some embodiments, the structured complementation data record enables representation of pairwise complementation relationships without duplication of sequence embedding data. Without wishing to be bound by a particular theory, approaches that recompute local structural context representations for each complementation prediction query require redundant processing of the same variant embeddings across many pairwise prediction operations, which can result in substantial duplication of computational effort in applications involving large numbers of variant combinations. By contrast, the structured complementation data record provided for herein encodes local structural context representations relative to the variant positions themselves, without requiring re-encoding of the sequence embedding at each record, thereby reducing storage and computational requirements. In some embodiments, the structured complementation data record further reduces computational processing required for accessing complementation prediction information, by enabling direct retrieval of precomputed complementation prediction outputs and local structural context representations from the data structure without requiring re-execution of the trained computational model for each query.

[0187] In some embodiments, the structured complementation data record comprises a plurality of structured records encoding pairwise complementation relationships among the variants of interest. In some embodiments, each structured record comprises one or more fields that encode the virtual residue coordinate of each variant position, the local structural context representation of each variant, the complementation prediction output, the complementation class label, the identities of the variants carrying each allele, and any additional annotations associated with the variant pair, such as predicted functional activity score, structural cluster assignment, population-level variant frequency, or clinical annotation.

[0188] In some embodiments, the computer-implemented method for generating the structured complementation data record is performed on one or more computers, such as the computer system described herein, and is implemented as one or more program modulesPNDRI.020WO PATENTstored on a non- transitory computer readable medium and executed by one or more processors.Embodiments of the Machine Learning Models

[0189] The disclosed machine learning models offer a scalable approach for predicting the non-linear genetic interaction of intragenic complementation. In some cases, because the pattern of intragenic complementation was consistent with variant sequestration, which obeys a specific set of structural rules, in these cases the intragenic complementation interactions should be predictable given information about the assembly of the monomers in the homomultimer and the positions of the active site residues.

[0190] In some embodiments, the disclosed technology provides a two-step transformer machine learning model architecture for predicting non-additive function of intragenic complementation. Specifically, the two-step ML model as illustrated in FIG. 2A may be used to predict whether intragenic complementation will occur between two loss-of-function ASL variants.

[0191] In some embodiments, the first step involves generation of a variant protein region embedding for each variant. The first step may capture the protein structure context of each amino acid position using embeddings from ESM-2, a large protein language model developed by Meta Al, and containing up to ~15 billion parameters. ESM-2 was pre-trained on more than 65 million protein sequences and has been shown to accurately capture protein structural information. For each of the two variants in the compound heterozygote, an embedding consisting of 1,280 parameters was generated for a 21 -amino-acid window centered on the variant residue. This embedding was able to capture the local protein structural context in which each variant resides. Therefore, the embedding dimension was 1280 parameters from ESM-2 for a 21-aa context window centered on the variant residue of interest corresponding to the local region in which the variant resides. In some embodiments, the ML model did not explicitly use the solved or predicted structure of a protein but rather learned these important structural features implicitly through the ESM-2 embeddings.

[0192] In some embodiments, the second step of the ML model was a deep learning neural network that iteratively weighed the embedding parameters in the context of the ASL variant combination data. The second step concatenated two embeddings corresponding to the variants and applied an attention-weighted average and subsequently a Sigmoid function to predict intragenic complementation based on the protein structuralPNDRI.020WO PATENTcontext from the embeddings of the two variant regions generated in the first step. Specifically, the Sigmoid function was applied to predict the growth of the compound heterozygote, with intragenic complementation indicated by predicted high growth. In some embodiments, the accuracy of predicting these non-additive interactions was maximized using a transformer neural network with an attention pooling layer and a prediction layer that was parameterized with variant protein region embeddings and trained on ASL intragenic complementation data.Intragenic Complementation

[0193] Intragenic complementation is a type of genetic interaction in which two variants, each displaying severe loss of function as homozygotes, show' significant restoration of protein function when brought together in combination. In 1964 Crick and Orgel proposed several models by which intragenic complementation could occur. One model describes what is now commonly referred to as compensatory mutation, in which the function of a protein complex is partially or fully restored by reciprocal amino acid substitutions that recreate molecular interactions such as salt bridges or hydrogen bonds. Under this model, both the position of the two amino acid substitutions and their biochemical properties underlie the ability to restore function in the protein. Crick and Orgel also proposed a model for intragenic complementation with a fundamentally different mechanism. In this model, protein function is restored when non-functional variants are sequestered into a subset of its multiple active sites, leaving the remaining active sites functional (See FIG. 1 A for some examples), and as such it is the positions, but not the biochemical properties of the two substituted amino acids, that drive restoration of protein function. Intragenic complementation through variant sequestration may occur when the active form of the protein is a homo-multimer with multiple active sites where each active site is formed from different monomers.

[0194] To illustrate protein structure properties permitting intragenic complementation by variant sequestration, FIG. 1A illustrates examples of intragenic complementation by variant sequestration in a dimeric protein. The wild-type protein is a homodimer with two active sites. Homodimers of the three active site variants (A, B. and C) are non-functional. Variant A and C disrupt the same local region of the active site (square cutout) in the protein monomer, while variant B disrupts a different local region (triangle cutout). Dimers consisting of variant B in combination with variant A or variant C result in a partially functional protein, since the variants are sequestered in one active site, leaving the onePNDRI.020WO PATENTfunctional active site. Proteins consisting of variant A and variant C contain only nonfunctional active sites.

[0195] In the examples shown in FIG. 1 A, the wild-type homodimer has two functional active sites comprised of different local regions (square and triangle cutouts) of each monomer. For a loss-of-function active site variant, A, an A / A dimer has two nonfunctional active sites. Similarly, a dimer consisting of variant monomers A and C, where both variants are in the same local region of the protein, will also have two non-functional active sites. In contrast, in a dimer consisting of two variant monomers in different local regions of the protein, such as A / B dimer or B / C dimer, the loss-of-function variants are sequestered in a single active site, leaving the other active site functional.

[0196] Embodiments of the invention relate to an electronic system or method of identifying variant monomers of a multimeric protein which may complement each other, such that when they come together to form a multimeric protein, some or all of the multimeric protein's function is restored. For example, a multimeric protein may be formed from two identical proteins (homodimer) or two different proteins (heterodimer). In some cases, the variants, when paired with a wild-type protein may form a partially, or completely, non-functional multimeric protein. However, when two particular variant proteins come together to form the multimeric protein, they may partially, or fully, restore the function of the multimeric protein due to complementation between the two variants. Embodiments of the invention include systems and methods which use a trained machine learning model to predict which variant proteins within a multimeric protein will complement each other and therefore be able to restore at least some, or all, of the wildt pe activity7of the native multimeric protein.

[0197] In these embodiments, a machine learning model is trained to identify combinations of protein variants which form an at least partially functional multimeric protein. For example, the machine learning model may be trained by first analyzing a particular multimeric protein and its variants to determine which variants do, and do not, complement each other. This analysis may be performed in a high throughput manner so that hundreds, thousands, or tens of thousands of variants may be analyzed to determine which variants were able to complement each other to form a multimeric protein with at least partial function. In one embodiment, the high throughput system may analyze the results of experiments using the Argininosuccinate Lyase (ASL) tetrameric protein, as discussed in more detail below, to determine which variants of the monomers making up this ASL protein may complement each other. For example, yeast cells containingPNDRI.02QWO PATENTmodified genes encoding the yeast ASL ortholog (ARG4) are unable to grow on medium lacking arginine. Thus, any yeast cells having variants which complement each other will be found to grow on such arginine-lacking media. The data from which variants do, and do not, complement each other may be used a train a ML model to recognize untested ASL variants to predict whether they would be able to complement each other.

[0198] Thus, embodiments may involve querying the ML model with a set of possible loss of function protein variants so the model can predict combinations of loss of function protein variants which would result in a multimeric protein which is at least partially functional. The system or method may then identify pairs of loss of function protein variants which when combined will result in an at least partially functional multimeric protein. For example, the improved function may be a 10. 20. 30. 40, 50, 60, 70, 80, 90, or 100% improvement in protein function as compared to two variants which do not complement each other.

[0199] In other embodiments, the ML model was found to be able to accurately predict intragenic complementation by learning and utilizing generalizable structural features of the protein. For example, the ML model was found to be able to accurately predict a phenotype of variant combinations that involved novel and untrained amino acid positions and accurately classify7interactions involving untrained positions as complementing or noncomplementing. Thus, embodiments are not limited to using the trained ML model to only predict whether particular trained positions of a protein may be able to complement a variant, but also accurately predict whether novel variants may also be able to complement a variant to result in a protein having improved function.Example Implementation of the Two-Step Machine Learning Model

[0200] For predicting the non-additive intragenic complementation effect between two variants, such as the ASL or FH variants, an example implementation of the two-step model may be represented as IF = {fega} where the parameters are denoted as 9 and o. These two sets of parameters are exclusive of each other and optimized separately. ferepresents generated embeddings of ESM2, a large protein language model (PLM) that is pretrained on the objective of masked language modeling or token / span corruption over millions of protein sequences from the UniRef database. To create one example implementation of the two-step model, ESM2 (esm2_t33_650M_UR50D; / / github.com / facebookresearch / esm) was selected as the first step of model development because it was used in many proteomic prediction tasks, such as secondary structure prediction, mutation effect estimation, andPNDRI.020WO PATENTevolutional information encoding. In the example implementation of the two-step model, the second step consisted of a two-layer Transformer encoder parameterized by o as in ga. Finally, an attention pooling layer was deployed, followed by a linear projection head to predict the final scores of the mutation effect.

[0201] Specifically, given the wild-type protein sequence of length, I, and two variants, the two variant protein sequences representing the two variant chains. var_seql and var_seq2, were generated. These two sequences were fed to the feand the protein embeddings, embl and emb2, computed, where the embedding size was IK / xdim-embed.fe: v ar_seq 1 emb 1fe - var_seq2— > emb2

[0202] Next, the embeddings corresponding to the variant regions were cropped, specifically up to +10 and -10, from the embl and emb2 and concatenated, resulting in a matrix of size IR2( xdim-embed. This tensor represented the independently affected variant site representations. In some cases, window spans of 15, 20 and full did not produce better results compared to a window size of 10.

[0203] The concatenated variation regions O2'xdim-embed)were fed as input to gain order to learn the non-additive function corresponding to the intragenic complementation effect. In the example implementation of the two-step model, no positional information was used explicitly so that the intragenic complementation effect predictor model was independent of any positional bias of the input amino acid position in the protein sequence. As such, this example implementation of the two-step model can be applied to estimate the variant-variant effects on other proteins that are not seen or utilized during model training. This allows the results of the specific model, trained on ASL or FH to be extensible to other similar proteins. As used herein, gamay be a two-layer transformer encoder with embedding dimensions the same as the fe. specifically 1280, with 8 attention heads per layer. Next, the example implementation of the two-step model has an attention pooling layer with tanh non-linearity that computes the weighted average of the embeddings at the variant interacting positions. Finally, the example implementation of the two-step model has a linear projection head to make the final regression.

[0204] Each model may be initialized using the PyTorch framework and default weight initialization and trained for 50 epochs with a learning rate of I e5. During training, a batch size of 128 may be utilized to ensure efficient learning and stable convergence. The input to the model included concatenated variant embeddings, which were processed to generatePNDRI.020WO PATENTa prediction score. For the regression model, the mean squared error (MSE) loss function was used to quantify the discrepancy between predicted and actual values. Model parameters were optimized using the AdamW optimizer, which effectively balanced weight decay and adaptive learning rates. To ensure optimal generalization, the best-performing model was based on its performance on the validation set.Validation of the Disclosed Machine Learning Models

[0205] Two types of cross-validation with two different partition schemes were carried out to evaluate the performance of the disclosed ML models.

[0206] The first was a 20-fold cross-validation where, for each fold, 95% of the variant combinations were used to train the ML model and predictions were made for the remaining 5% of the combinations. The variant combinations were randomized and then split into 20 groups of 183 variant combinations. Model training was conducted with all groups but the left-out group was then used for prediction.

[0207] FIG. 2B shows a strong correlation between the model-predicted ASL diploid growth versus the experimentally measured ASL diploid growth. Model-predicted growth values are from the 20-fold cross-validation. FIG. 2C shows an AUC-ROC curve of the predicted ASL diploid growth from 20-fold cross-validation. These results suggest that the disclosed machine learning models can accurately predict intragenic complementation between ASL variants with a very high correlation to the experimentally measured results. The predicted growth and the measured growth were very highly correlated (R2=0.907; Table 1).PNDRI.020WO PATENTTABLE 1. ML MODEL METRICS.RegressionR2MSE RMSE PCCPC^ P‘Line.m Line-c valueASL 20-fold 0.907 0.009 0.094 0.985 0.000 1.08 0.005 AQT 1 P VP- 1 - A A -Posi- 0.847 0.031 0.176 0.920 0.000 1.03 0.007 outEnsemble FH 0.511 0.082 0.287 0.715 0.000 0.523 0.261Supervised ClassificationAccuracy Fl Precision Recall ROC-AUC ASL 20-fold 0.996 0.996 0.993 1 0.999 AQT 1 AAV' P- 1 - A A -PG - 0.980 0.984 0.988 0.980 0.995 outEnsemble FH 0.911 0.935 0.879 1 0.894 MSE, mean squared error; RMSE, root mean squared error; PCC, Pearson’s correlation coefficient; Line-m: slope of regression line; Line-c: y-intercept of regression line; ROC-AUC, receiver-operating characteristic curve area under curve

[0208] Experimentally, combinations of loss-of-function active site variants fall into two clear classes, complementing (>55% growth) or non-complementing (<10% growth). Given the excellent agreement between the values predicted from 20-fold cross-validation and observed growth values, any decision boundary in the predicted range 15-55% will correctly classify 99.6% of variant combinations as complementing or non-complementing (FIG. 2B and 2C and Table 1). The small number (15 out of 3,599 combinations) of misclassified combinations all involved variants of the highly conserved catalytic center H160, either with other variants at the same position or with the neighboring amino acid position 159 (FIG. 2B).

[0209] The second cross-validation scheme tested the ML model’s ability to predict all combinations involving a specific amino acid position where all variants at that position were excluded from the training set (leave-one-AA-position-out, L1AAPO). For each amino acid position, the model was trained on all variant combinations except those involving the amino acid in question to predict the left-out variant combinations. FIG. 3 shows the results of ML-predicted growth of ASL diploids from L1AAPO cross-validation. Each subplot in FIG. 3 plots model-predicted ASL diploid growth from L1AAPO cross-validation versus experimentally-measured ASL diploid growth for different variant combinations w here one of the variants is of the amino acid position labeled above each subplot. For example, “E86” refers to the amino acid at position 86 in the reference (wildtype) human ASL where “E” is glutamic acid. Individual data points in the “E86” subplotPNDRI.02QWO PATENTrepresent measured and predicted growth of variant combinations where one of the variants is of the E86 residue. Here, the model was able to accurately predict the phenotype of variant combinations that involved a novel (untrained) amino acid position (R2=0.847; Table 1 and FIG. 3) and very accurately classify (accuracy=0.98 using decision boundary of 15% predicted growth; Table 1) interactions involving untrained positions as complementing or non-complementing. The results suggest that models produced by the disclosed ML models are able to accurately predict intragenic complementation by learning and utilizing generalizable structural features of the protein.Assays

[0210] Disclosed herein are assays for determining the functional activity of variant combinations in a multimeric protein, including without limitation assays for generating experimental training data for the trained computational model described herein, assays for validating complementation predictions generated by the methods provided for herein, and assays for determining the functional activity of a compound heterozygous genotype in a subject. In some embodiments, the functional activity data obtained using the assays provided for herein serves as input data (200) to the Complementation Prediction Module (300) described herein. In some embodiments, the functional activity data is used to train, validate, or refine the trained computational model described herein.

[0211] In some embodiments, obtaining or having obtained functional activity data for one or more variant combinations comprises performing an assay. In some embodiments, the assay is a cell-based functional assay, a biochemical activity assay, a growth-based assay, a high-throughput screening assay, a deep mutational scanning assay, or any combination thereof. In some embodiments, the assay generates quantitative measurements of functional activity for one or more variant combinations. In some embodiments, the information from the assay is quantitative and is transmitted to a computer system for downstream analysis. In some embodiments, the information is qualitative, such as observed growth patterns, fluorescence, or colorimetric signals, which can be translated into a quantitative measure by a user or automatically by a reader or computer system.

[0212] In some embodiments, the assay is a yeast-based functional assay. In some embodiments, the yeast-based functional assay is used to measure the functional activity of one or more variants of a multimeric protein, individually or in pairwise or higher-order combinations. In some embodiments, the yeast-based functional assay is a growth-based assay in which the functional activity of a variant or variant combination is measured byPNDRI.020WO PATENTthe growth rate or fitness of a yeast strain expressing the variant or variant combination under selective conditions. In some embodiments, the yeast-based functional assay is a yeast complementation assay.

[0213] In some embodiments, the yeast complementation assay comprises generating a yeast diploid strain harboring a first variant on one allele and a second variant on a second allele of the multimeric protein. In some embodiments, generating the yeast diploid strain comprises introducing the first variant into a first haploid yeast strain and introducing the second variant into a second haploid yeast strain, and mating the first haploid yeast strain and the second haploid yeast strain to produce the diploid yeast strain. In some embodiments, the diploid yeast strain is a Saccharomyces cerevisiae diploid strain. In some embodiments, the selective conditions under which functional activity is measured comprise growth on selective media that requires functional activity of the multimeric protein for survival or growth. In some embodiments, functional activity is measured by quantifying colony size, optical density, fluorescence, or other growth-related metrics of the diploid yeast strain under the selective conditions. In some embodiments, functional activity is normalized to wild-type functional activity measured in a control yeast strain expressing the wild-type allele of the multimeric protein on both alleles.

[0214] In some embodiments, the yeast-based functional assay is a high-throughput yeast-based growth assay. In some embodiments, the high-throughput yeast-based growth assay enables simultaneous measurement of functional activity for a plurality of variant combinations, including at least 100, at least 500, at least 1,000, at least 3,000, or more pairwise variant combinations. In some embodiments, the high-throughput yeast-based growth assay comprises a pooled screening format in which a library of yeast strains expressing different variant combinations is grown under selective conditions and the relative fitness of each variant combination is determined by sequencing-based readout of strain abundance. In some embodiments, the high-throughput yeast-based growth assay is performed according to methods described in International Publication Nos. WO2019204295. WO2021076654, WO2024081592, and WO2024097899, each of which is incorporated herein by reference in its entirety'.

[0215] In some embodiments, the yeast-based functional assay further comprises a deep mutational scanning component. In some embodiments, the deep mutational scanning component comprises introducing a library of variants into a yeast strain expressing the multimeric protein and measuring the functional activity of each variant individually or in combination with one or more additional variants. In some embodiments, the deepPNDRI.02QWO PATENTmutational scanning component enables systematic characterization of the functional consequences of all possible single amino acid substitutions at one or more residue positions of the multimeric protein. In some embodiments, the deep mutational scanning component is performed according to methods described in International Publication Nos. WO2019204295, WO2021076654, WO2024081592, and WO2024097899, each of which is incorporated herein by reference in its entirely.

[0216] In some embodiments, the assay is a biochemical activity assay. In some embodiments, the biochemical activity assay measures the catalytic activity, binding affinity, substrate turnover, or other biochemical property of the multimeric protein comprising the first variant, the second variant, or both. In some embodiments, the biochemical activity assay is an enzyme activity assay, a binding assay, a fluorescencebased assay, a colorimetric assay, a luminescence-based assay, a mass spectrometry-based assay, or any combination thereof. In some embodiments, the biochemical activity assay is performed on purified protein, cell lysate, or intact cells expressing the variant combination of interest.

[0217] In some embodiments, the assay is a cell-based functional assay. In some embodiments, the cell-based functional assay measures the functional activity of one or more variant combinations in a cellular context that recapitulates the endogenous environment of the multimeric protein. In some embodiments, the cell-based functional assay is performed in a mammalian cell line, a primary cell culture, an induced pluripotent stem cell-derived cell type, or a model organism cell. In some embodiments, the cell-based functional assay comprises introducing the first variant and the second variant into a cell that is null or hypomorphic for the endogenous multimeric protein, and measuring restoration of functional activity relative to wild-type controls. In some embodiments, functional activity is measured by enzymatic assay, reporter gene assay, metabolite quantification, proliferation assay, or any other readout that is sensitive to the activity of the multimeric protein.

[0218] In some embodiments, the assay is a sequencing-based assay. In some embodiments, obtaining or having obtained molecular data from one or more variant combinations comprises performing a sequencing-based assay. Examples of sequencingbased assays include, without limitation, Sanger sequencing, next-generation sequencing (NGS), Illumina sequencing, Ion Torrent sequencing, PacBio single-molecule real-time (SMRT) sequencing, Oxford Nanopore sequencing, whole-genome sequencing, whole-exome sequencing, RNA-seq, ChlP-seq, methyl-seq, targeted sequencing, and single-cellPNDRI.020WO PATENTsequencing. In some embodiments, the sequencing-based assay is used to confirm the identity of the first variant and the second variant in a yeast strain or cell line prior to a functional assay. In some embodiments, the sequencing-based assay is used to determine the relative abundance of variant combinations in a pooled screening format, as described herein with respect to the high-throughput yeast-based grow th assay.

[0219] In some embodiments, the assay is an amplification-based assay. Examples of amplification-based assays include, without limitation, polymerase chain reaction (PCR), reverse transcription PCR (RT-PCR), real-time PCR (qPCR), reverse transcription quantitative PCR (RT-qPCR), digital PCR (dPCR), reverse transcription digital PCR (RT-dPCR), loop-mediated isothermal amplification (LAMP), nucleic acid sequence-based amplification (NASBA), rolling circle amplification (RCA), and strand displacement amplification (SDA). In some embodiments, the amplification-based assay is followed by a sequencing-based assay, such as those provided for herein. In some embodiments, performing the assay comprises contacting a test sample with a plurality of reagents. In some embodiments, the plurality of reagents comprises primers, probes, reverse transcriptase, polymerase, or any combination thereof.

[0220] In some embodiments, the assay is a microarray-based assay. In some embodiments, the microarray-based assay is a gene chip, biochip, or protein microarray that enables simultaneous measurement of one or more molecular properties of a plurality of variant combinations. In some embodiments, the microarray-based assay is used in conjunction with a sequencing-based assay or amplification-based assay to provide complementary molecular characterization of variant combinations of interest.

[0221] In some embodiments, the assay is a high-throughput screening assay. In some embodiments, the high-throughput screening assay enables simultaneous evaluation of functional activity for a plurality of variant combinations in a parallelized format. In some embodiments, the high-throughput screening assay is a pooled library screen, an arrayed screen, a flow cytometry-based screen, a fluorescence-activated cell sorting (FACS)-based screen, or any combination thereof. In some embodiments, the high-throughput screening assay generates functional activ ity measurements for at least 100, at least 1,000, at least 3,000, at least 10,000, or more variant combinations in a single experiment.

[0222] In some embodiments, the high-throughput screening assay is performed according to methods described in International Publication Nos. WO2019204295, WO2021076654. WO2024081592, and WO2024097899, each of which is incorporated herein by reference in its entirety. In some embodiments, the functional activity dataPNDRI.020WO PATENTgenerated by the high-throughput screening assay is used to train the trained computational model described herein. In some embodiments, the functional activity data generated by the high-throughput screening assay is used to validate complementation predictions generated by the methods provided for herein, including without limitation the methods for predicting complementation and non-complementation described herein.

[0223] In some embodiments, performing any of the assays provided for herein comprises contacting a test sample with a plurality of reagents. In some embodiments, the plurality of reagents comprises primers, probes, enzymes, substrates, antibodies, fluorescent labels, or any combination thereof. In some embodiments, the test sample is a cell, cell lysate, purified protein, nucleic acid extract, tissue biopsy, blood sample, saliva sample, buccal swab sample, or other biological sample obtained from a subject or derived from a yeast or other model organism strain.

[0224] In some embodiments, the information generated by the assay is quantitative and is transmitted to a computer system for downstream analysis, including without limitation training of the trained computational model, validation of complementation predictions, or generation of structured complementation data records, such as those provided for herein. In some embodiments, the information is qualitative, such as observed growth patterns, fluorescence signals, colorimetric readouts, or colony morphology7, which can be translated into a quantitative measure by a user or automatically by a reader or computer system. In some embodiments, the quantitative measure is a normalized functional activity score, a growth score, an enzymatic activity value, a binding affinity value, or any other metric that reflects the functional activity of the multimeric protein comprising the variant combination of interest.

[0225] In some embodiments, the assays provided for herein are performed in accordance with methods described in International Publication Nos. WO2019204295, WO2021076654, WO2024081592, and WO2024097899, each of which is incorporated herein by reference in its entirety.Applying the Two-Step Machine Learning Model to Predict a Structurally Similar Protein

[0226] The above results demonstrate that this form of intragenic complementation (i.e. , variant sequestration) obeys a set of structural rules in ASL. Here, how well the rules generalize to a protein with a similar structure w as tested using the human fumarase (FH) protein. The FH protein is a clinically important enzyme that underlies a rare metabolic disorder (fumarase deficiency) and is also associated with rare inherited cancers. FH andPNDRI.02QWO PATENTASL are both members of the Lyase class I family (CDD: cd01334). Like ASL, FH is active as a homotetramer with four multichain active and ligand-binding sites, each formed by residues from three monomers. Additionally, the two proteins are so structurally similar (RMSD=3.4, TM-score=0.79, 392 residues aligned, see FIG. 4A) that the complementing positions in ASL to FH may be mapped. However, despite their structural similarity', FH and ASL are quite diverged in primary’ sequence, showing only 21% amino acid sequence identity.

[0227] Thus, the predictions discovered in FH using the ML model developed with ASL data may be based primarily on FH’s structural similarity' with ASL. Since the FH data is small compared to the model size, the ensemble method may be applied to compute the final prediction scores. In some embodiments, the ensemble model generates outputs as the average of the models’ predictions developed during the 20-fold ASL cross-validation. The performance of the ensemble model may be tested on the FH variant combinations.

[0228] Experimental testing of the FH protein may utilize the fact that deletion of the yeast FH ortholog. FUM1, confers a growth defect on rich medium with glycerol as the sole carbon source, and expression of the human FH restores growth. To experimentally test whether variants of active site residues in FH participate in intragenic complementation, FH residues comparable to the ASL active site residues in both sequence and structure space were identified. The experimental testing included selecting nine missense FH variants at eight amino acid positions and confirming that yeast strains harboring these variants resulted in null growth on solid rich media with glycerol, where the variants of residues at or near the active-Zligand-binding sites had previously been associated with FH deficiency and other FH-associated diseases since these may result in loss of function.

[0229] FIG. 4A illustrates the protein structure alignment of human ASL (PDB: laos) and human FH (PDB: 5upp,) homotetramers. FIG. 4B shows the experimental results of the distribution of FH compound heterozygous diploid yeast growth normalized to homozygous fuml deletion and homozygous wild-type human FH complementation diploid strains (n=2-4). FIG. 4C is a heatmap of the normalized diploid growth of all variant combinations tested in FH. Top left triangle and bottom right triangle along the diagonal are mirrored and reflect the same data. Hierarchical clustering of normalized diploid growth identified four clusters. FIG. 4D illustrates residues of surface representation of human FH monomer (PDB: 5upp) based on clusters identified in FIG. 4C. FIG. 4E shows model-predicted versus experimentally-measured FH compound heterozygous diploid growth.PNDRI.02QWO PATENTModel-predicted growth values are from the ensemble model of the ASL 20-fold cross-validation models. The horizontal dashed line indicates the 0.35 decision threshold for intragenic complementation classification. FIG. 4F is a chart showing the ROC-AUC curve of the predicted FH compound heterozygous diploid growth.

[0230] The above results demonstrate intragenic complementation in FH with patterns consistent with the variant sequestration model. Of the 36 unique compound heterozygous missense variant combinations tested, 29 combinations exhibited intragenic complementation (FIG. 4B and FIG. 4C). The intragenic complementation effects were strong, with activity in complementing combinations restored to >65% WT. Even with fewer variants and amino acid positions as compared to ASL, the pattern by which complementation occurred in FH clearly emulated the pattern of ASL intragenic complementation (FIG. 4C and FIG. 4D, and FIG. 7B and FIG. 7C. Variants were localized to distinct regions of the FH monomer, and variant combinations from different local regions exhibited intragenic complementation while variant combinations within the same local region did not. The only deviation from the prediction based on the ensemble model was with the combination of amino acid positions 233 and 235. The prediction expected variants at these two positions to behave as one cluster, sharing the same intragenic complementation pattern, but instead the two positions behaved as two separate clusters (FIG. 4C). Overall, the behavior of 34 of the 36 variant combinations tested agreed with the results expected based on the ensemble model, suggesting that non-additive interactions measured for one protein were largely transferable to structurally similar proteins.

[0231] Although the relationship betw een the predicted and the measured growth values was noisier for FH than ASL, applying a classification threshold of 0.35 achieved an accuracy of greater than 91% (FIG. 4E and FIG. 4F, and Table 1). Therefore, the ML models, trained on experimental data from one protein, were found to be transferrable to make accurate predictions of genetic interactions in a novel protein with a similar structure.

[0232] By leveraging the PDB to identify other proteins with structural features consistent with the variant sequestration model, it is estimated that >6% of human genes (N=I,272) may display this form of intragenic complementation. Given the pervasiveness of this type of genetic interaction, many potential human genoty pes may be erroneously classified by models assuming additivity7. While the number of variant combinations that participate in this phenomenon is very large, the frequency of any specific genotype appearing in the human population is expected to be very low. As a result, it would be very difficult to learn the set of genetic and structural rules of intragenic complementationPNDRI.020WO PATENTuncovered in this study from conventional human genotype-phenotype associations. This is particularly true for rare inherited diseases, such as ASL or FH deficiency.

[0233] Experimentally testing large numbers of variant combinations to identify these instances would be costly, time-intensive, and in some systems, infeasible. The disclosed ML models can solve this problem by capturing the structural rules that permit intragenic complementation by variant sequestration, enabling accurate prediction of these nonadditive interactions. The disclosed ML models can be applied to multiple members of the same protein family. Given the wealth of protein structures, including Al-predicted structures, a relatively small number of experimentally-derived complementation datasets could be leveraged to predict non-additive allelic interactions across a large number of proteins. This is potentially valuable for precision medicine because scalable machinelearning solutions would bypass the need for domain-specific expertise, for example in the area of protein structure.Computer Implementation

[0234] The methods disclosed herein, such as the methods for predicting complementation or non-complementation between variants of a multimeric protein, the methods for optimizing computational resources for variant pair evaluation, and the methods for generating a machine-readable protein variant data structure, are, in some embodiments, performed on one or more computers. For example, the building and deployment of a system for predicting complementation or non-complementation between variants of a multimeric protein, and database storage, can be implemented in hardware or software, or a combination of both. In one embodiment, a machine-readable storage medium is provided, the medium comprising a data storage material encoded with machine readable data which, when using a machine programmed with instructions for using said data, is capable of displaying any of the datasets and execution and results of a complementation prediction model. Such data can be used for a variety' of purposes, such as clinical variant interpretation, therapeutic decision-making, patient monitoring, treatment considerations, and the like. Methods disclosed herein can be implemented in computer programs executing on programmable computers, comprising a processor, a data storage system (including volatile and non-volatile memory and / or storage elements), a graphics adapter, a pointing device, a network adapter, at least one input device, and at least one output device. Program code may be applied to input data to perform the functions described above and generate output information. The output information is applied to onePNDRI.020WO PATENTor more output devices, in known fashion. The computer can be, for example, a personal computer, microcomputer, or workstation of conventional design.

[0235] Each program can be implemented in a high level procedural or object oriented programming language to communicate with a computer system. However, the programs can be implemented in assembly or machine language, if desired. In any case, the language can be a compiled or interpreted language. Each such computer program is preferably stored on a storage media or device (e.g., ROM or magnetic diskette) readable by a general or special purpose programmable computer, for configuring and operating the computer when the storage media or device is read by the computer to perform the procedures described herein. The sy stem can also be considered to be implemented as a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.

[0236] The complementation prediction data and databases thereof can be provided in a variety of media to facilitate their use. "Media" refers to a manufacture that contains the complementation prediction information, including without limitation structured complementation data records, local structural context representations, trained computational model parameters, and complementation class label outputs, such as those described herein. The databases as described herein can be recorded on computer readable media, e.g. any medium that can be read and accessed directly by a computer. Such media include, but are not limited to: magnetic storage media, such as floppy discs, hard disc storage medium, and magnetic tape; optical storage media such as CD-ROM; electrical storage media such as RAM and ROM; and hybrids of these categories such as magnetic / optical storage media. One of skill in the art can readily appreciate how any of the presently known computer readable mediums can be used to create a manufacture comprising a recording of the present database information. "Recorded" refers to a process for storing information on computer readable medium, using any such methods as known in the art. Any convenient data storage structure can be chosen, based on the means used to access the stored information. A variety of data processor programs and formats can be used for storage, e.g. word processing text file, database format, etc.

[0237] FIG. 12 illustrates an example computer 600 for implementing the entities and modules described herein, including without limitation the Input Data Module (200), the Complementation Prediction Module (300), and the Output Module (400) as depicted generally in FIG. 10B. The computer 600 includes at least one processor 602 coupled to aPNDRI.020WO PATENTchipset 604. The chipset 604 includes a memory controller hub 620 and an input / output (I / O) controller hub 622. A memory 606 and a graphics adapter 612 are coupled to the memory controller hub 620, and a display 618 is coupled to the graphics adapter 612. A storage device 608, an input device 614, and network adapter 616 are coupled to the I / O controller hub 622. Other embodiments of the computer 600 have different architectures.

[0238] The storage device 608 is a non-transitory computer-readable storage medium such as a hard drive, compact disk read-only memory (CD-ROM). DVD, or a solid-state memory device. The memory 606 holds instructions and data used by the processor 602. The input device 614 is a touch-screen interface, a mouse, track ball, or other type of pointing device, a keyboard, or some combination thereof, and is used to input data into the computer 600. In some embodiments, the computer 600 may be configured to receive input (e.g., commands) from the input device 614 via gestures from the user. The graphics adapter 612 displays images and other information on the display 618. The network adapter 616 couples the computer 600 to one or more computer networks.

[0239] The computer 600 is adapted to execute computer program modules for providing functionality described herein. As used herein, the term “module” refers to computer program logic used to provide the specified functionality. Thus, a module can be implemented in hardware, firmware, and / or software. In one embodiment, program modules are stored on the storage device 608. loaded into the memory 606, and executed by the processor 602.

[0240] The types of computers 600 used by the entities of FIG. 12 can vary' depending upon the embodiment and the processing power required by the entity7. For example, the embodiment can run in a single computer 600 or multiple computers 600 communicating with each other through a network such as in a server farm. The computers 600 can lack some of the components described above, such as graphics adapters 612, and displays 618.

[0241] In some embodiments, the non-transitory' computer readable medium comprising instructions that, when executed by a processor, cause the processor to perform one or more of the methods provided for herein, includes: the method for predicting complementation between at least two variants of a multimeric protein; the method for predicting noncomplementation between at least two variants of a multimeric protein; the method for optimizing computational resources for variant pair evaluation; the method for generating a machine-readable protein variant data structure from identified variant positions and their associated complementation prediction outputs; and the method for classifying variant pairsPNDRI.020WO PATENTas complementing or non-complementing based on a trained computational model, each as described herein.

[0242] In some embodiments, a system for performing the methods provided for herein comprises a processor and a non-transitory computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform one or more of: the method for predicting complementation between at least two variants of a multimeric protein; the method for predicting non-complementation between at least two variants of a multimeric protein; the method for optimizing computational resources for variant pair evaluation, including generating vectors of pairwise complementation probabilities and reducing the computational search space for variant pair evaluation; the method for generating a machine-readable protein variant data structure from identified variant positions and their associated complementation prediction outputs; and the method for classifying variant pairs as complementing or non-complementing pairs based on local structural context representations processed by a trained computational model, including without limitation a transformer-based neural network trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein.ENUMERATED EMBODIMENTS1. An electronic method of identifying which pairs of loss of function protein variants making up a multimeric protein will provide an at least partially functional multimeric protein, comprising:providing a machine learning model trained to identify combinations of protein variants which form an at least partially functional multimeric protein, wherein a wild type multimeric protein is made of two or more protein subunits, and each protein variant when replacing one protein subunit of the wild type multimeric protein has a reduced enzy matic activity as compared to the wild type multimeric protein;querying the model with a set of possible loss of function protein variants to determine combinations of loss of function protein variants which would result in a multimeric protein which is at least partially functional; andidentifying pairs of loss of function protein variants which when combined result in an at least partially functional multimeric protein.PNDRI.020WO PATENT2. The method of embodiment 1, wherein the at least partially functional multimeric protein has at least 35% of enzymatic activity as compared to the wild type multimeric protein.3. The method of embodiment 1, wherein the at least partially functional multimeric protein has at least 50% of enzy matic activity as compared to the wild ty pe multimeric protein.4. The method of embodiment 1, wherein the at least partially functional multimeric protein has at least 80% of enzymatic activity as compared to the wild type multimeric protein.5. The method of embodiment 1, wherein the at least partially functional multimeric protein has the complete enzymatic activity of the wild type multimeric protein.6. The method of embodiment 1, wherein the at least partially functional multimeric protein has a higher enzymatic activity7as compared to the wild type multimeric protein.7. The method of embodiment 1, wherein the machine learning model comprises a protein language model.8. The method of embodiment 7, wherein querying the model comprises inputting the protein sequences of the set of possible loss of function protein variants to the protein language model.9. The method of embodiment 7, wherein the protein language model comprises embeddings to represent the protein sequences of the protein variants as structural information.10. The method of embodiment 7, wherein the protein language model is an ESM2 protein language model.11. The method of embodiment 1, wherein providing the machine learning model comprises providing a tw o-step transformer machine learning model.12. The method of embodiment 11, wherein providing the two-step transformer machine learning model comprises a first step of generating variant protein region embedding parameters for each protein variant.13. The method of embodiment 12, wherein providing the two-step transformer machine learning model comprises a second step of providing a deep learning neural network configured to iteratively weigh the embedding parameters with weights obtained from training the deep learning neural network on experimental data.PNDRI.020WO PATENT14. The method of embodiment 13, wherein the second step concatenates two embeddings corresponding to pairs of protein variants and applies an attention-weighted average and a Sigmoid function to predict intragenic complementation between the protein variants in the pairs.15. The method of embodiment 1, wherein the machine learning model is trained to predict whether intragenic complementation occurs between two loss-of-function protein variants.16. The method of embodiment 1, wherein the protein variants are human protein variants.17. The method of embodiment 1, wherein the protein variants are ASL variants.18. The method of embodiment 17, wherein the ASL variants make up at least two subunits of a tetrameric ASL enzyme.19. The method of embodiment 1 , wherein the protein variants are human fumarase (FH) protein variants.20. The method of embodiment 1, wherein the loss of function protein variants are derived from missense single nucleotide variants in genes encoding the protein variants.21. The method of embodiment 1, wherein the identified loss of function protein variants comprise a first protein variant having a first affected local structure and a second protein variant having a second, different affected local structure, and wherein the combination of the identified loss of function protein variants forms at least one enzy atic active site.22. The method of embodiment 1, wherein the machine learning model is trained to identify combinations of protein variants that make up the at least partially functional multimeric protein based on the local protein structural feature and / or amino acid position of one or more amino acids altered in the protein variants.23. The method of embodiment 22, wherein the one or more amino acids altered in the protein variants relate to enzymatic active sites of the multimeric protein.24. A system for identifying which pairs of loss of function protein variants making up a multimeric protein will provide an at least partially functional multimeric protein, comprising:a machine learning model trained to identify combinations of protein variants which form an at least partially functional multimeric protein, wherein aPNDRI.020WO PATENTwild type multimeric protein is made of two or more protein subunits, and each protein variant when replacing one protein subunit of the wild type multimeric protein has a reduced enzymatic activity as compared to the wild type multimeric protein;an interface for inputting a set of possible loss of function protein variants into the machine learning model to determine combinations of loss of function protein variants which would result in a multimeric protein which is at least partially functional; andone or more processors configured to identify pairs of loss of function protein variants which when combined together would result in an at least partially functional multimeric protein.25. A computer-implemented method for predicting complementation or non-complementation between at least two variants of a multimeric protein, the method comprising:(a) obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site;(b) identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence;(c) obtaining, for each of the first variant and the second variant, a local structural context representation;(d) determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output; and (e) classify ing, from the complementation prediction output, the first variant and the second variant as a complementing pair or a non-complementing pair, wherein steps (a) through (e) may be performed in any order, concurrently, or iteratively,26. The method of embodiment 25, further comprising, prior to step (a), identifying the multimeric protein as a candidate for complementation by:obtaining from a protein structure database at least one homomultimeric protein assembly comprising at least one ligand in contact with more than one subunit; andselecting at least one homomultimeric protein assembly in which a ligand contacts at least two identical subunits within a heavy atom-to-heavy atom distance of 6 A or less.PNDRI.020WO PATENT27. The method of embodiment 26, wherein the ligand comprises more than 5 atoms and more than 50% of the atoms of the ligand are within 3.6 A of at least two protein chains of the homomultimeric assembly.28. The method of embodiment 26 or 27, further comprising identifying additional candidate multimeric proteins by obtaining, from a protein structure database, a structurally similar candidate multimeric protein, wherein the structurally similar candidate multimeric protein has a structure similarity score of at least 0.5.29. The method of embodiment 28, further comprising identify ing human orthologs of candidate proteins from model organisms using an ortholog database.30. The method of any one of embodiments 25-29, wherein the local structural context representation is generated by extracting a sequence window centered on the respective variant residue position.31. The method of embodiment 30, wherein the sequence window comprises from 10 to 50 amino acids centered on the respective variant residue position.32. The method of embodiment 31, wherein the sequence window comprises 21 amino acids.33. The method of any one of embodiments 30-32, wherein the local structural context representation is generated by computing an embedding of the sequence window using a pretrained protein language model.34. The method of embodiment 33. wherein the pretrained protein language model is selected from the group consisting of ESM-2, ESM-lb, ESM-lv, ProtTrans, ProtBERT, TAPE, UniRep, SeqVec, AlphaF old-based language model, and any functionally equivalent pretrained protein language model capable of generating local structural context representations of amino acid sequences.35. The method of embodiment 34, wherein the pretrained protein language model is ESM-2.36. The method of any one of embodiments 33-35, wherein the embedding comprises a vector of about 500 to about 5,000 parameters.37. The method of embodiment 36, wherein the embedding comprises a vector of about 1,280 parameters.38. The method of any one of embodiments 33-37, wherein the embedding encodes local structural context information from the amino acid sequence, without three-dimensional protein structure coordinates as input.PNDRI.020WO PATENT39. The method of any one of embodiments 25-38, wherein the local structural context representations of the first variant and the second variant are generated independently of one another.40. The method of any one of embodiments 25-39, wherein the local structural context representations are precomputed and reused across multiple predictions.41. The method of any one of embodiments 25-40, wherein the complementation prediction output is determined by using a trained computational model that receives the local structural context representations of both the first variant and the second variant as input.42. The method of embodiment 41, wherein the trained computational model comprises a neural network configured to receive as input the concatenated local structural context representations of the first variant and the second variant.43. The method of embodiment 42, wherein the neural network is a transformer-based neural network.44. The method of any one of embodiments 41-43, wherein the complementation prediction output is determined without using numerical positions of the first residue position or the second residue position in the amino acid sequence as inputs to the trained computational model.45. The method of any one of embodiments 41-44. wherein the trained computational model has been trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein.46. The method of embodiment 45, wherein the experimental measurements of functional activity’ are obtained by a high-throughput yeast-based growth assay, deep mutational scanning, or other high-throughput functional screening method.47. The method of embodiment 45 or 46, wherein the experimental measurements of functional activity comprise measurements for at least 1,000 pairwise variant combinations.48. The method of embodiment 47, wherein the experimental measurements of functional activity’ comprise measurements for at least 3,000 pairwise variant combinations.49. The method of any one of embodiments 41-48, wherein the trained computational model has been trained using cross-validation, including 20-fold cross-validation, leave-one-amino-acid-position-out cross-validation, or a combination thereof.PNDRI.020WO PATENT50. The method of any one of embodiments 41-49, wherein the trained computational model is an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.51. The method of any one of embodiments 25-50, wherein the multimeric protein is a homomultimeric protein comprising two or more identical subunits.52. The method of embodiment 51, wherein the homomultimeric protein is a homodimer, homotrimer, homotetramer, homopentamer, or higher-order homomultimer.53. The method of any one of embodiments 25-50, wherein the multimeric protein is a heteromultimeric protein comprising at least two non-identical subunits.54. The method of any one of embodiments 25-53, wherein the multimeric protein comprises at least two subunits.55. The method of any one of embodiments 25-54, wherein the at least one functional site is a multichain active site, ligand-binding site, or catalytic site formed by residues contributed by at least two subunits of the multimeric protein.56. The method of any one of embodiments 25-55. wherein the first residue position and the second residue position are in different regions of the ammo acid sequence, wherein the regions are defined structurally, empirically through experimental complementation data, or implicitly by a trained computational model.57. The method of any one of embodiments 25-56. wherein the first residue position and the second residue position are in different spatial clusters of the protein structure, as determined by hierarchical clustering of complementation data or three-dimensional structural analysis.58. The method of any one of embodiments 25-57, wherein the multimeric protein is an enzy me, receptor, transporter, structural protein, or regulatory protein.59. The method of any one of embodiments 25-58, wherein the first variant and the second variant are each loss-of-function variants exhibiting no more than 10% of wild-type protein activity when present alone.60. The method of any one of embodiments 25-59, wherein the first variant and the second variant are each independently an amino acid substitution, insertion, or deletion relative to a reference sequence.61. The method of any one of embodiments 25-60, wherein complementation between the first variant and the second variant is determined by the first residue position and the second residue position in the protein structure and is independent of the biochemical identities of the substituted amino acids.PNDRI.020WO PATENT62. The method of any one of embodiments 25-61 , wherein the first residue position and the second residue position are each located at or near an active site, ligandbinding site, or catalytic residue of the multimeric protein.63. The method of any one of embodiments 25-62, wherein the complementation prediction output informs whether the first variant and the second variant produce a nonadditive functional outcome.64. The method of any one of embodiments 25-63, wherein the complementation prediction output comprises a continuous predicted growth value, a probability7of complementation, a binary complementation class label, or a direct classification or ranking generated without a separately computed interaction score.65. The method of any one of embodiments 25-64. wherein classifying comprises applying a predetermined decision threshold to the complementation prediction output.66. The method of embodiment 65, wherein the decision threshold is selected such that predicted complementation prediction output above the threshold is classified as complementing pairs and complementation prediction output below the threshold is classified as non-complementing pairs.67. The method of any one of embodiments 25-66, wherein a complementing pair exhibits restoration of at least 50% of wild-type protein activity.68. The method of embodiment 67, wherein a complementing pair exhibits restoration of at least 60% of wild-tj pe protein activity.69. The method of any one of embodiments 41-68, wherein the trained computational model achieves a classification accuracy of at least 0.90 on test data.70. The method of embodiment 69, wherein the trained computational model achieves a classification accuracy of at least 0.95 on test data from the same multimeric protein used to train the trained computational model.71. The method of embodiment 70, wherein the trained computational model achieves a classification accuracy of at least 0.95 on test data comprising variant combinations at residue positions not included in the training data.72. The method of any one of embodiments 41-71, wherein the trained computational model achieves an area under the receiver operating characteristic curve (AUC-ROC) of at least 0.85.73. The method of any one of embodiments 41-72, wherein the trained computational model achieves an R2of at least 0.84.PNDRI.020WO PATENT74. The method of any one of embodiments 25-73, further comprising, for a pair classified as a complementing pair, quantifying the predicted level of functional restoration by applying an additive model, wherein the additive model accounts for the independent allelic contributions of the first variant and the second variant to the quantitative level of restored protein activity.75. The method of embodiment 74, wherein combining the complementation classification with the additive model achieves an R2of at least 0.99.76. The method of any one of embodiments 41-75, wherein the trained computational model has been trained on variant combinations from a first multimeric protein and is applied to predict complementation in a second multimeric protein that is structurally similar to but sequence-divergent from the first multimeric protein.77. The method of embodiment 76, wherein the first multimeric protein and the second multimeric protein share less than 25% amino acid sequence identity.78. The method of embodiment 76 or 77, wherein the first multimeric protein and the second multimeric protein are members of the same structural family or superfamily.79. The method of any one of embodiments 76-78, wherein the trained computational model achieves a classification accuracy of at least 0.90 on variant combinations from the second multimeric protein.80. The method of any one of embodiments 25-79, further comprising: identifying a plurality of additional variants at additional residue positions in the amino acid sequence; anddetermining complementation prediction outputs for a plurality of pairwise or higher-order combinations of variants.81. The method of embodiment 80, further comprising generating a complementation map, heatmap, interacton network, or visual representation identi lying which variant combinations are predicted to complement one another.82. The method of embodiment 80 or 81, further comprising clustering the variants by residue position based on predicted or observed complementation paterns, thereby identifying two or more spatally different regions of the amino acid sequence that contribute to multichain active site assembly.83. The method of embodiment 82, wherein variants residing within the same local region do not complement each other, and variants residing in different local regions complement each other.PNDRI.020WO PATENT84. The method of any one of embodiments 80-83, wherein the plurality of pairwise combinations comprises all pairwise combinations of loss-of-function variants at two or more active site or ligand-binding site residue positions.85. The method of any one of embodiments 25-84, further comprising: determining a predicted functional score for the first variant; determining a predicted functional score for the second variant; and quantifying the degree of nonadditivity by comparing the interaction score to an expected additive combination of the individual predicted functional scores.86. The method of any one of embodiments 25-85, wherein obtaining the amino acid sequence comprises retrieving the amino acid sequence from a sequence database, receiving user-specified input, or extracting the amino acid sequence from sequencing data.87. The method of any one of embodiments 25-86, wherein identify ing the first variant and the second variant comprises receiving user-specified variants, analyzing sequencing data from a subject, or computationally generating variant combinations from a set of known loss-of-function variants.88. The method of any one of embodiments 25-87, wherein the method is performed without experimental screening of variant combinations prior to prediction.89. The method of any one of embodiments 25-88, further comprising preprocessing the amino acid sequence to identify active site residues, ligand-binding residues, or structurally important positions prior to identifying the first variant and the second variant.90. The method of any one of embodiments 25-89, further comprising using the classification to inform a therapeutic approach selected from the group consisting of gene therapy, cell therapy, enzyme replacement therapy, protein replacement therapy, substrate reduction therapy, and small molecule targeting.91. The method of any one of embodiments 25-90, further comprising using the classification to predict the phenotypic severity of a compound heterozygous genotype in a subject, wherein a classification of complementation pair is associated with a predicted milder disease phenotype relative to a homozy gous loss-of-function genotype.92. The method of any one of embodiments 25-91, further comprising validating the classification by experimentally measuring functional activity of a compound heterozygous organism or cell expressing both the first variant and the second variant.PNDRI.020WO PATENT93. The method of embodiment 92, wherein the compound heterozygous organism is a yeast diploid strain harboring the first variant on one allele and the second variant on a second allele, and wherein functional activity is measured by a growth assay under selective conditions.94. A non-transitory computer readable medium comprising instructions that, when executed by a processor, cause the processor to perform the computer-implemented method of any one of embodiments 25-93.95. A system comprising:a processor; anda non-transitory computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the computer- implemented method of any one of embodiments 25-93.96. A computer-implemented method of generating a dataset for modeling complementation in multimeric proteins, the method comprising:generating, via a high-throughput functional assay, experimental measurements of functional activity for a plurality of pairwise combinations of variants of a multimeric protein, wherein each variant comprises a modification at a respective residue position in an amino acid sequence;constructing a dataset comprising a plurality of entries, each entry corresponding to a respective pairwise combination of variants and comprising:(i) identifiers of the pairwise combination of variants,(ii) a quantitative measurement of functional activity associated with the pairwise combination, and(iii) a label indicating a degree of non-additive interaction between the variants; andorganizing the dataset such that variant pairs exhibiting complementation are distinguishable from variant pairs that do not exhibit complementation, wherein the dataset encodes relationships between spatially distinct residue positions that contribute to restoration of protein function upon multimer assembly.97. The method of embodiment 96, wherein the label indicating the degree of non-additive interaction comprises a deviation from an expected additive functional effect of the individual variants.PNDRI.020WO PATENT98. The method of embodiment 96, wherein the dataset comprises at least 1,000 pairwise variant combinations.99. The method of embodiment 96, wherein the experimental measurements are obtained using a yeast-based growth assay.100. The method of embodiment 96, wherein the dataset encodes functional interactions between residue positions independent of explicit three-dimensional structural coordinates.101. A non-transitory computer-readable medium storing a dataset for modeling complementation in multimeric proteins, the dataset comprising a plurality of entries, each entry corresponding to a pairwise combination of variants of a multimeric protein, each entry comprising:(i) identifiers of a first variant and a second variant at respective residue positions,(ii) a quantitative functional activity measurement obtained from experimental evaluation of the pairwise combination, and(lii) a non-additivity metric representing deviation from an expected additive functional outcome;wherein the dataset is structured to encode complementation relationships between variants located in different spatial regions of the protein, such that the dataset enables identification of variant pairs that restore at least partial functional activity upon multimer assembly.102. The non-transitory computer-readable medium of embodiment 101, wherein the dataset is configured to train a computational model to predict complementation between variants without requiring structural coordinate input.103. The non-transitory computer-readable medium of embodiment 101, wherein the quantitative functional activity measurement is derived from deep mutational scanning.104. The non-transitory computer-readable medium of embodiment 101, wherein entries are organized to reflect clustering of residue positions into two or more spatially distinct complementation groups.105. A computer-implemented method of training a computational model, the method comprising:obtaining a dataset according to any of embodiments 96-104;PNDRI.020WO PATENTtraining the computational model using the dataset to learn relationships between variant pairs and functional activity outcomes; andgenerating, using the trained computational model, a prediction of whether a pair of variants exhibits complementation,wherein the trained computational model is configured to predict complementation for variant pairs not included in the dataset.106. The computer-implemented method of embodiment 105, wherein the trained computational model is capable of predicting complementation in a second multimeric protein that is sequence-divergent from a protein used to generate the dataset.107. The computer-implemented method of embodiment 105, wherein the prediction is generated without using three-dimensional structural coordinates of the multimeric protein.108. The computer-implemented method of embodiment 105, wherein the trained computational model achieves a classification accuracy of at least 0.90 in predicting complementation.EXAMPLES

[0243] The following examples are offered to illustrate, but not to limit the invention. In order to facilitate understanding, the specific embodiments are provided to help interpret the technical proposal, that is, these embodiments are only for illustrative purposes, but not in any way to limit the scope of the invention. Unless otherwise specified, embodiments do not indicate the specific conditions, are in accordance with the conventional conditions or the manufacturer’s recommended conditions.Example 1: Determining Functional Impact of ASL Amino Acid Substitutions

[0244] FIG. 5A is an overview of high throughput yeast-based growth assays for characterizing the functional impact of SNV-accessible ASL missense variants. In order to test the variant sequestration model at scale, strong loss-of-function variants whose activity could be tested in combination were identified. Next, the activity of all missense single nucleotide variant (SNV)-accessible amino acid substitutions in ASL were constructed and assayed, reflecting the class of variants most likely to occur in the human population.

[0245] The yeast-based growth assay used leverages the fact that cells lacking the yeast ASL ortholog (ARG4) are unable to grow on medium lacking arginine, and that robust growth under this selective condition is restored by expression of the yeast codon-PNDRI.020WO PATENToptimized human ASL protein coding sequence under the transcriptional control of the yeast ARG4 promoter and integrated in single copy into the yeast genome (FIG. 5A and FIG.5B). The assay provides a quantitative readout of variant enz me function, with cells harboring likely benign / benign variants exhibiting growth levels comparable to wildtype and those with known pathogenic ASL variants displaying impaired growth (FIG. 5B). FIG.5B shows the experimental data of the growth of humanized yeast harboring wild-type human ASL, 0286, R12Q, and T131M variants normalized to the growth of wild-type yeast and yeast ortholog ARG4 deletion strains on selective solid agar medium lacking arginine at day 3 (n=10). Error bars indicate standard error.

[0246] Using this assay, the functional impact of 2,431 (88% of all 2,712 possible SNV-derived) missense ASL variants was assessed individually across the length of the protein. FIG. 5C shows the normalized experimental data of the growth of 2,431 SNV-accessible ASL missense variants along the protein sequence on selective solid agar medium lacking arginine at day 3 (n=l-8 per variant). Variants of active and ligand-binding residues are highlighted in black.

[0247] The 1,644 variants (-68%) with activities >85% of wildtype activity include all known benign and likely benign (B / LB) variants (FIG. 6). FIG. 6 illustrates functional impact of clinically-relevant ASL missense variants by showing the normalized experimental data of the grow th of ASL missense variants in ClinVar along the protein sequence on selective solid agar medium lacking arginine at day 3 (n=l-8 per variant). VUS indicates variants of uncertain significance; P / LP indicates pathogenic / likely pathogenic; and B / LB indicates benign / likely benign.

[0248] 419 (17%) amorphic variants with very low' levels of activity (< 5% WT activity), comparable to that of a null allele, were also identified. This range of the assay was enriched for active site residues (one-sided Fisher’s test p-value < 0.001, FIG. 5C). The remaining 368 (1 %) hypomorphic variants showed impaired growth, but above 5% of WT activity7.

[0249] In contrast to the B / LB, known pathogenic and likely pathogenic variants exhibited growth both below and above the threshold of 85%, suggesting that the assay can distinguish many, but not all pathogenic variants from WT. It is possible that the assay can identify strong loss-of-function variants but not those with high residual activity. Therefore, variants displaying growth <85% are likely to be clinically relevant.PNDRI.020WO PATENTExample 2: Intragenic Complementation Paterns Support the Variant Sequestration Model

[0250] The variant sequestration model predicted a restoration of activity between specific combinations of active site variants. Pairwise combinations of 90 complete loss-of-function variants of ASL (having <5% WT ASL activity), located at 20 ligand-binding or other active site residues (FIG. 5C), were tested because these have the potential to show the most dramatic restoration of function. Diploid strains containing two variants in trans, i.e. compound heterozygotes, were generated by mating haploid yeast strains, each harboring one of the 90 variants (FIG. 5C). Pervasive intragenic complementation in ASL active sites was observed.

[0251] FIG. 7A shows the distribution of ASL compound heterozygote diploid yeast growth normalized to homozygous arg4 deletion and homozygous wild-type human ASL complementation diploid strains (n=l-4). FIG. 7B is a heatmap of the average normalized diploid growth of all variant combinations between two amino acid positions. Top left triangle and botom right triangle along the diagonal are mirrored and reflect the same data. Hierarchical clustering of the amino acid positions based on the average normalized diploid growth detects three clusters. FIG. 7C is a 3D surface representation of the ASL monomer (PDB: laos) with active and ligand-binding site residues based on clustering in FIG. 7B. FIG. 7D is a 3D surface representation of the ASL tetramer (PDB: laos), shown 90° clockwise rotation of ASL tetramer in FIG. IB. Inset highlights an active and ligandbinding site formed by three monomers. Active site and ligand-binding residues are based on clustering in FIG. 7B. Grey stick in FIG. 7D represents the ligand, argininosuccinate.

[0252] FIG. 8A shows the distribution of ASL diploid yeast growth normalized to homozy gous arg4A deletion and homozy gous wild-type human ASL complementation diploid strains (n=l-4). FIG. 8B illustrates the additive effect of ASL alleles in 2.219 compound heterozygous ASL diploids with growth > 55% by linear regression analysis (R2=0.88).

[0253] Consistent with the examples illustrated in FIG. 1A, all the homozy gous combinations of loss-of-function active site variants tested showed litle to no growth (<8% of homozygous WT ASL growth, FIG. 8A). In contrast, for the compound heterozygotes, a clear bimodal distribution was observed. 1,380 of the 3,599 combinations tested failed to complement each other, showing growth of less than 10% of WT. How ever, over 60% of the combinations (N=2,219) exhibited high levels ofgrowth(56-110% of homozygous WT ASL growth) consistent with intragenic complementation (FIG. 7A). Because the two distributions were so well separated, there was no ambiguity7as to which pairs arePNDRI.02QWO PATENTcomplementing or non-complementing. Thus, the results indicate that intragenic complementation in ASL occurs frequently and produces a significant restoration of enzyme activity.

[0254] In the results dataset, every variant that was tested complemented at least one other variant, but no variant complemented all other variants (FIG. 9). FIG. 9 summarizes the experimental data of the normalized growth for all tested ASL variant combinations by showing a heatmap of the normalized diploid growth of all variant combinations tested (n=l-4). Variants of the amino acid position are displayed together, grouped in order by amino acid position in FIG. 7B. Top left triangle and bottom right triangle along the diagonal are mirrored and reflect the same data. Variant combinations not tested are indicated by white cells in the heatmap.

[0255] Examining the pattern of which pairs were able or unable to complement demonstrated that complementation is primarily a function of amino acid positions and almost independent of the identity of the amino acid (FIG. 7B and FIG. 9). If two amino acid positions, A and B, are considered, then either all the variants of A will complement all the variants of B, or none of the variants of A will complement any of the variants of B. This allowed for the analysis of the patterns of intragenic complementation on the basis of amino acid positions. Clustering of the complementation data by amino acid positions produced three discrete clusters. Pairs of variants from amino acid positions in the same cluster always failed to complement (FIG. 7B). while pairs of variants from different clusters always displayed complementation (FIG. 7B). These clusters mapped to three distinct local regions that correspond to the three active site regions on the monomer structure (FIG. 7B and FIG. 7C) that assemble into the active sites of the ASL tetramer (FIG. 7D). This matched the pattern shown in FIG. 1A. with variants from the same local region failing to complement, and complementation only occurring between variants from different local regions, which can be sequestered into a subset of the multiple active sites. Therefore, the results of the analysis recapitulate the variant sequestration model of intragenic complementation.

[0256] Although the results of the analysis demonstrated that qualitatively, intragenic complementation is driven by amino acid position, amino acid identity does contribute to the quantitative phenotype. Some variation in growth was observed among the complementing pairs, and additive allelic effects explain most of this variation (FIG. 8B, R2=0.88). Therefore, knowledge of which position pairs complement (FIG. 7B) togetherPNDRI.02QWO PATENTwith the additive modifier effect of the alleles (FIG. 8B) explain essentially all the phenotypic variance (R2=0.997).Example 3: Plasmid Construction

[0257] For the ASL variant library, template plasmid AB635 contained the wild-type ASL coding sequence (NM_000048.4) codon-optimized and synthesized by IDT (hereafter yASL) and assembled into a pUC 19-based vector downstream of the yeast ARG4 promoter (439 bp), and upstream of yeast TRP5 terminator and natMX drug marker cassette, using the NEB HiFi DNA Assembly Cloning Kit (New England Biolabs). Plasmid sequence was verified by whole plasmid sequencing (Plasmidsaurus).

[0258] To facilitate mating type switching in the yASL variant strains, pYML104H plasmid (GenBank PV147142) was generated. First, pML104 was modified via DNA assembly (NEB HiFi DNA Assembly) to contain the yeast codon-optimized Cas9 and the hphNTl marker cassette in place of the URA3 selectable marker. Then MATa protospacer sequence GTTCTAAAAATGCCCGTGCT (SEQ ID NO: 1) was added downstream of the SNR52 promoter, and a second pSNR52-gRNA-tSUP cassette was inserted into the plasmid that expresses sgRNA GCTGACCGTCGAGGACATCG (SEQ ID NO: 2) targeting the nourseothricin N-acetyl transferase gene in the natMX marker.TABLE 2. PRIMERS FOR DNA ASSEMBLY.pML 104Hph_fwd gcgaattgggtaccgggc SEQ ID NO: 3 25nm STD make pML 104Hph_rev cctatagtgagtcgtattacg SEQ ID NO: 4 25nm STD double ege guides gRN A_cas s ette_fwd gtaatacgactcactataggg SEQ ID NO: 5 25nm STD aacaaaagctggagctcgRN A_cas s ette_r e v gggcccggtacccaattcgc SEQ ID NO: 6 25nm STDcgtggtatcgtttagattg

[0259] Expression plasmids of FH variants were generated via site-directed mutagenesis of modified p413TEF-MCS plasmids containing the yeast codon-harmonized FH coding sequence and the kanMX or natMX selectable marker. The p413TEF plasmid was first modified by inserting the kanMX or the natMX selectable marker EagI site after the CYC1 terminator. Whole plasmid sequencing (Plasmidsaurus) confirms the correct assembly of p413TEF-MCS-tCYCl-kanMX and p413TEF-MCS-tCYCl -natMX. The yeast codon-optimized FH coding sequence was synthesized (Twist Bioscience) and inserted into the MCS of modified p413TEF using NEB HiFi DNA Assembly Cloning Kit (New EnglandPNDRI.020WO PATENTBiolabs). Codon-harmonization was conducted. Site-directed mutagenesis was performed using Platinum SuperFi II DNA Polymerase and overlapping mutagenic primers. Expression plasmids of FH variants were confirmed by whole plasmid sequencing (Plasmidsaurus).Example 4: Yeast Media

[0260] Yeast strains were grown in YPD or YPG (1% yeast extract, 2% peptone, 2% glucose or 3% glycerol, respectively) unless otherwise noted. When selection was required, antibiotics were added to the following final concentrations: 100 mg / L nourseothricin sulfate (NAT, AG Scientific), 400 mg / L geneticin (G-418 sulfate, Gibco), and 250 mg / L hygromycin B (HYG, Goldbio). Yeast minimal (SD) media without amino acid was composed of 0.17% yeast nitrogen base, 0.5% ammonium sulfate and 2% glucose. Solid media plates were prepared as 2% agar.Example 5: Yeast Strain Construction

[0261] The yeast strain, YAD1255, containing the cassette KAN-tTEF in place of the yeast ARG4 gene was used as the transformation strain for the variant library. Briefly, a pARG4-KAN-tTEF-tARG4 amplicon was amplified from YKO981 with RL153 and RL154 primers and transformed into prototrophic MATa FY4 using a standard yeast transformation protocol. Deletion of yeast ARG4 was confirmed by Sanger sequencing. The strain was auxotrophic for arginine and resistant to G418.

[0262] The deletion strain Y AD 1257, containing the NATMX selectable marker in place of the yeast ARG4 was used as the negative control strain for the ASL variant assay experiment. An amplicon containing pARG4-pTEF-NAT-tTEF-tARG4 was amplified from AB635 with AO628: AGAATATTCGCAATTGAAGAGCTCAAAAGCAGGTAACTATATAACAAGACTA AGGCAAACGACATGGAGGCCCAGAATAC (SEQ ID NO: 7) and RL154 primers and transformed into FY4. The strain was auxotrophic for arginine and resistant to NAT.

[0263] The positive control yeast strain for the ASL variant assay experiment, YAD1260, contained the expression cassette pARG4-yASL-tTRP5 -NATMX from AB635 in place of the yeast ARG4 gene. The expression construct was amplified with ARG4_Frag_F and RL154 primers and transformed into YAD1255.

[0264] Variant yeast strains harboring FH variants were generated by transforming variant FH expression vectors into fumlA0;:hphMX Mata and Matalpha strains. First, thePNDRL02QWO PATENTselectable marker in fumlAO: : KanMX was swapped with HphMX amplified from pFA6a-HphMX with AO677 TEF_prom_fwd GACATGGAGGCCCAGAATAC (SEQ ID NO: 8) and AO471 TEF term R: CGCACTTAACTTCGCATCTG (SEQ ID NO: 9). The fuml O:: HphMX expression construct was amplified with FUM1_A and FUM1_D and integrated into FY4 and FY5.Example 6: ASL Variant Library Construction

[0265] When constructing the ASL variant library, variants of the ASL open reading frame were synthesized by Twist Biosciences in which variants were returned a pool of 4-7 missense variants by amino acid position. The number of variants per amino acid position was determined by what missense variants are accessible by a single nucleotide variation in the native human ASL coding sequence. The AB635 served as the template for site saturation variant library synthesis (Twist Biosciences). The variant gene cassettes were amplified from each pool with Fwd (5’-3’) primer: GGTGGTTACTCATTGGCAG (SEQ ID NO: 10) and Rev (5’-3’) primer: GCATAACGTCGCCATCTG (SEQ ID NO: 11) using a high-fidelity polymerase and then transformed into YAD1255 strain. Since each amino acid position is represented by 4-7 variants, 18 to 21 colonies from each transformation plate were picked to obtain at least 3 independent transformants per variant. Individual transformants were arrayed into 96-well flat-bottom microplates. Each microplate contained the negative control deletion strain YAD1257 in positions Al and H12, and the wild-type ASL control strain YAD1260 in C3, CIO, F3 and F10 positions. Microcultures of the variant library in 96-well microplates were saved as glycerol stocks as well as pinned onto YPG+NAT omnitrays to select against transformants with respiratory chain defects and petite phenotypes that can misconstrue downstream analysis. The variant library' was maintained on YPG+NAT at 4°C until phenotyping.Example 7: ASL Variant Library Sequencing

[0266] After phenotype pinning, a 45 pl. microculture aliquot from each well of the source plates were frozen and stored in a 96-well PCR plate at -20°C. The variant library was constructed such that the target amino acid position is known for each transformant in a given well of a 96-well plate, but not the specific amino acid substitution. To determine the specific variant, the variant library was sequence-confirmed using a custom long-read sequencing pipeline. Yeast aliquots of 16 source plates were pooled by well positions such that no target amino acid position w as repeated in a single pool (well position). VariantPNDRI.020WO PATENTORFs were amplified from each pool using Phusion High-Fidelity DNA Polymerase (Thermo Scientific) with in-line barcoded primers for identifying well positions in downstream analysis. Amplicon concentration was quantified using Sybr Green. Approximately 20 ng of barcoded amplicons from each well (minus control wells) were combined and purified (Zymo DNA Clean and Concentrator) for long-read sequencing library preparation according to Oxford Nanopore protocol for SQK_LSK109 ligation kit. Fifteen fmol of library were loaded onto the MinlON Flongle (R9.4.1) and sequenced for 24 hours. Raw FAST5 files were processed with Guppy 6.0.1 basecaller using the super high accuracy model, and the output FASTQ files were demultiplexed using Minibar (https: / / github.com / calacademy-research / minibar) with the following parameters: method (M)=l: search length (1) =150 and percentage match (p)=0.75. Variants were determined with custom-R scripts.Example 8: ASL Variant Effect Assay

[0267] Source plates were prepared by inoculating yeast from YPG+NAT into 96-well flat bottom microplates containing 150 pL YPD in each well. The source plates were then sealed with adhesive film and incubated at 30°C overnight until the microcultures reached a stationary phase. The next day, each source plate was pinned onto three SD agar destination OmniTrays (ThermoFisher cat. no. 165218) using the Biomek i7 automated workstation outfitted with a 96 floating-pin replicator tool (V&P Scientific). Destination plates were grown at 30°C for 72 h and then imaged under consistent lighting with a mounted Canon PowerShot SX10 IS compact digital camera (IS0200, £4.5 aperture, 1 / 40 s exposure). Growth measurements were extracted using PyP18 (pypi.org / project / PyP18) with s=200 and pin size=0 or 25. Briefly, yeast colonies were detected and segmented from grayscale images with Otsu thresholding with optional circle detection for null and low-growth colonies when pin size was set at 25. A mask differentiating pixels of the yeast colonies from the background was produced, and pseudo-volumes, the sum of grayscale values of pixels corresponding to the yeast colonies, were used for downstream analysis.Example 9: Growth Estimate Data Analysis

[0268] To prepare for phenotyping, each isolate was picked and grown in rich medium supplemented with the selectable drug and then pinned onto solid rich medium containing 3% glycerol as the carbon source (YPG), to select against cells that were slow growing as a result of loss of respirator}7function. Cells were then pinned from YPG solid mediumPNDRL02QWO PATENTback into non-selective rich liquid medium (YPD) and grown to saturation overnight at 30°C. Next, cultures were pinned onto solid omniplates containing minimal (SD) medium (lacking arginine) with a Biomek i7 robot outfitted with a V&P Scientific 96 floating-pin replicator head. Pinning onto SD plates was done in triplicate, followed by growth at 30°C for 72 h. Plates were photographed with a mounted Canon PowerShot SX10 IS compact digital camera under consistent lighting, camera to subject distance, and zoom. Images (ISO200, f4.5, 1 / 40 s exposure) were acquired as .jpg files. Growth phenotypes were extracted from the images via a custom pipeline, PyPL8 (github.com / lacyk3 / PyP18). Briefly, a pseudo patch “volume” for each pinned isolate w as generated, consisting of the product (“IntDen”) of the patch area (obtained via thresholding with Otsu’s method28 or circle detection29) and the mean patch pixel intensity (in grayscale).Example 10: Data Normalization and Relative Growth Estimation

[0269] Normalization steps to account for the effects of plate-to-plate variation, relative grow th of neighboring patches, and plate edge effects w ere applied to the haploid grow th data using a custom script. Isolates with (non-synonymous) secondary mutations were removed from the dataset and any variants that had discordant growth estimates between isolates were also removed. Finally, a linear model was fit with weighted least-squares regression to estimate the relative growth of each y ASL variant. Similar normalization steps were applied to the diploid growth data.Example 11: Mating Type and Marker Switching for ASL Variant Strains

[0270] For intragenic complementation assays, MATalpha ASL variant yeast strains were generated with the KANMX selectable marker using yeast strains from the MATa ASL variant natMX-marked library'. For variants with multiple replicates in the variant library, strains were selected with the highest variant to reference read count ratio to swap mating type and selectable marker via yeast co-transformation with kanMX dsDNA and AB753 plasmid DNA. The kanMX cassette was PCR-amplified from pFA6a-kanMX plasmid withFwd primer: GACATGGAGGCCCAGAATAC-3' (SEQ IDNO: 12) and Rev primer: CGCACTTAACTTCGCATCTG (SEQ ID NO: 13). Fifty microliters of kanMX PCR product (AO677 and AO471 primers) and 1 ug AB753 plasmid DNA were used for yeast transformation. Transformants were selected on YPD+G418+HYG, followed by three passages in YPD to cure the plasmid. Marker sw apping was confirmed to be sensitive to NAT and resistant to G418. Mating ty pe was determined by PCR. Two MATalpha,PNDRI.020WO PATENTG418-resistant transformants of each variant were cherry-picked and arrayed in 96-well plates.Example 12: Yeast Mating

[0271] MATa strains were grown in 5 mL YPD to saturation in culture tubes overnight at 30°C. MATalpha strains were cultured in 1 mL YPD to saturation in 96-deep well plate overnight at 30°C. Yeast mating was conducted by adding 10 pL of stationary phase cultures of opposite yeast mating type strains to 96-well F-bottom microplates containing 120 pL YPD and grown overnight at 30°C. Diploids were selected by pinning the overnight mating culture onto YPG+G418+NAT solid OmniTrays.Example 13: Intragenic Complementation Assay

[0272] Source plates were prepared by transferring yeast from YPG+G418+NAT diploid selection OmniTrays into 96-well F-bottom microplates containing 150 pL YPD. Source plates were cultured to saturation for 24 hours at 30°C and then robotically pinned onto SD solid OmniTrays using the same method used for variant library assay. Images of phenotype plates were captured at 72 hours with consistent lighting and a Canon EOS 80D DSLR camera (ISO200, f / 8 aperture, 1 / 20 s exposure) mounted to maintain a consistent distance from the OmniTrays. Pseudo-volumes of yeast colonies were extracted from images using PyP18 with s=328 and pin size=0 or 30.Example 14: Identify ing Candidate Proteins for Intragenic Complementation

[0273] Candidate proteins for intragenic complementation were identified from PDB searches (RCSB December 2024) for oligomeric assemblies that a) contain at least two protein chains with the same identity and b) a single ligand with a contact distance of < 6 A with at least tw o chains. PDB IDs were filtered for homo-oligomeric assemblies and their corresponding UniProt IDs, gene names and systematic IDs from the Alliance of Genome Resources (AGR, www.alliancegenome.org) were retrieved using the ID mapping tool on UniProt (uniprot.org / id-mapping). Since not all human protein structures have been structurally determined by x-ray crystallography, the number of human protein candidates were extended by identifying the human orthologs of proteins from model organisms that are a part of the AGR. Additionally, since not all human proteins were solved with their ligands, a strict similar structure search was conducted.PNDRI.020WO PATENTExample 15: Identifying Human Genes Impacted by Variant Sequestration

[0274] Given that intragenic complementation via variant sequestration occurs frequently in one homomultimer with multi-chain active sites, how many other proteins might experience this form of genetic interaction was estimated. First, the Protein Data Bank was queried for homo-multimeric proteins where each ligand interacts with more than one monomer. Then a distance threshold was applied between each chain and ligand of 6 Angstroms or less, which captures direct hydrogen bonds or interactions via a bridging water molecule. The search was limited to crystal structures solved experimentally in the presence of their ligand(s) and with resolution sufficient to calculate the distance between the chains and the ligand(s). This search identified proteins that correspond to 662 human genes. Because the search parameters exclude human proteins solved in the absence of their ligands, the set of proteins corresponding to the 662 human genes was used to conduct a strict structure similarity search in PDB. This search identified proteins corresponding to 460 additional human genes. The search identified 125 additional human genes where solved structures for orthologous proteins from model organisms passed the search criteria. Taken together, it was estimated that as many as 1,247 human genes, ~6% of the genes in the human genome, could participate in this form of genetic interaction.Example 16: Predicting Complementation andNon-Complementation Between Variants of a Homomultimeric Enzyme from Amino Acid Sequence

[0275] A candidate multimeric protein is identified by querying a protein structure database, such as the Protein Data Bank (PDB), for homomultimeric protein assemblies in which at least one ligand is in contact with more than one subunit. Specifically, the query retrieves all homomultimeric assemblies in which a ligand of more than 5 atoms has more than 50% of its atoms within 3.6 A of at least two protein chains of the assembly, and in which the ligand contacts at least two identical subunits within a heavy atom-to-heavy atom distance of 6 A or less. From this query, a homomultimeric enzyme is identified, herein referred to as Enz me X, which is a homodimer in which the active site is formed at the interface of the two identical subunits, with catalytic residues contributed by both chains.

[0276] To identify additional candidate proteins, a structural similarity search is performed against the protein structure database using the three-dimensional structure of Enzyme X as a query. Structurally similar candidate proteins with a structure similarity score of at least 0.5 are retrieved. Human orthologs of candidate proteins identified from model organisms are further identified using an ortholog database. Enzyme X is confirmedPNDRI.020WO PATENTas the primary candidate for complementation analysis, as its multichain active site architecture is consistent with the potential for intragenic complementation between loss-of-function variants residing at different structural regions of the active site.

[0277] The amino acid sequence of Enzyme X is retrieved from a sequence database. The full-length amino acid sequence of Enzy me X comprises 480 residues. The sequence is preprocessed to identify active site residues, ligand-binding residues, and structurally important positions, based on the available three-dimensional structure of Enzyme X deposited in the protein structure database. This preprocessing identifies 24 residue positions that form or directly contact the multichain active site, including residues contributed by both subunits of the homodimer. These 24 residue positions are designated as candidate positions for loss-of-function variant analysis.

[0278] A set of loss-of-function variants is identified at the 24 candidate active site residue positions. Specifically, for each of the 24 candidate positions, all possible single amino acid substitutions relative to the wild-type sequence of Enzyme X are computationally generated, yielding a total of 24 x 19 = 456 candidate single amino acid substitution variants. From this set, variants exhibiting no more than 10% of wild-type protein activity when present alone are identified as loss-of-function variants based on prior experimental deep mutational scanning data obtained using a high-throughput yeast-based growth assay performed in accordance with the methods described in International Publication Nos. WO2019204295, WO2021076654. WO2024081592. and WO2024097899, each of which is incorporated herein by reference in its entirety. A total of 82 loss-of-function variants are identified across the 24 candidate positions.

[0279] All pairwise combinations of the 82 loss-of-function variants are computationally generated, yielding a total of (82 x 81) / 2 = 3,321 candidate pairwise variant combinations. The first variant is designated as the first variant at the first residue position, and the second variant is designated as the second variant at the second residue position, for each pairwise combination. All 3,321 pairwise combinations are retained for complementation prediction.

[0280] For each of the 82 loss-of-function variants, a local structural context representation is generated as follows. A sequence window of 21 amino acids centered on the respective variant residue position is extracted from the amino acid sequence of Enzy me X, with the variant amino acid substituted at the central position of the window; For example, for a variant at residue position 142 of Enzyme X, a 21-amino acid window comprising residues 132 through 152 of the amino acid sequence is extracted, with the variant amino acid substituted at position 142.PNDRI.020WO PATENT

[0281] Each 21 -amino acid sequence window is input into the ESM-2 pretrained protein language model, which computes an embedding of the sequence window. The embedding comprises a vector of 1,280 parameters that encodes local structural context information from the amino acid sequence surrounding the variant position, without requiring three-dimensional protein structure coordinates as input. The local structural context representations for all 82 loss-of-function variants are computed independently of one another and are stored for reuse across all 3,321 pairwise complementation prediction queries, thereby avoiding redundant computation.

[0282] Prior to performing complementation predictions for all 3,321 pairwise combinations, a vector of pairwise complementation probabilities is generated to reduce the computational search space. Each element of the vector represents the probability that a given pairwise variant combination exhibits complementation, calculated by combining the local structural context representations of the first variant and the second variant with prior information derived from structural clustering of the 24 candidate residue positions.

[0283] Specifically, the 24 candidate residue positions are clustered into two or more spatially defined structural regions based on their three-dimensional positions in the active site of Enzyme X, as determined by hierarchical clustering of the positions using their Cartesian coordinates in the deposited cry stal structure. This clustering identifies two distinct structural regions of the multichain active site: Region A, comprising residues contributed by one face of the homodimer interface, and Region B. comprising residues contributed by the opposite face of the homodimer interface. Variant pairs in which both variants reside within the same structural region are assigned a low prior probability' of complementation, consistent with the principle that variants at the same structural region of a multi chain active site are unlikely to complement one another. Variant pairs in which the first variant resides in Region A and the second variant resides in Region B are assigned a high prior probability of complementation, consistent with the principle that variants at different structural regions of a multichain active site are more likely to complement one another.

[0284] Low-probability variant pairs are deprioritized in the evaluation order, and high-probability variant pairs are prioritized for evaluation by the trained computational model, thereby reducing the effective computational search space relative to exhaustive, unordered evaluation of all 3,321 pairwise combinations.

[0285] For each of the 3.321 pairwise variant combinations, a complementation prediction output is determined using a trained computational model. The trainedPNDRI.020WO PATENTcomputational model comprises a transformer-based neural network that receives as input the concatenated local structural context representations of the first variant and the second variant, specifically, the concatenation of the two 1,280-parameter ESM-2 embedding vectors, yielding a 2,560-parameter input vector, and generates as output a continuous predicted functional activity7value representing the predicted level of functional restoration in a compound heterozygous organism expressing the first variant on one allele and the second variant on the other allele.

[0286] The trained computational model has been trained on experimental measurements of functional activity7for pairwise combinations of variants of Enzy me X obtained using a high-throughput yeast-based growth assay, comprising measurements for at least 3,000 pairwise variant combinations performed in accordance with the methods described in International Publication Nos. WO2019204295, WO2021076654, WO2024081592, and WO2024097899, each of which is incorporated herein by reference in its entirety7. The trained computational model has been trained using 20-fold cross-validation and leave-one-amino-acid-position-out cross-validation, and the final complementation prediction output for each pairwise combination is generated as the average prediction across all models in the ensemble. The trained computational model achieves a classification accuracy of at least 0.95 on test data comprising variant combinations at residue positions not included in the training data, an AUC-ROC of at least 0.85, and an R2of at least 0.84 on continuous functional activity predictions.

[0287] Notably, the complementation prediction output is determined without using the numerical positions of the first residue position or the second residue position in the amino acid sequence as inputs to the trained computational model, such that the prediction is based solely on the local structural context information encoded in the embeddings of the two variants.

[0288] The complementation prediction output for each of the 3,321 painvise variant combinations is classified as a complementing pair or a non-complementing pair by applying a predetermined decision threshold to the continuous predicted functional activity value. Specifically, pairwise combinations with a predicted functional activity' value above the decision threshold, corresponding to predicted restoration of at least 50% of wild-type protein activity', are classified as complementing pairs. Pairwise combinations with a predicted functional activity value below7the decision threshold are classified as noncomplementing pairs.PNDRI.020WO PATENT

[0289] Of the 3,321 pairwise combinations evaluated, a subset is classified as complementing pairs and the remainder as non-complementing pairs. Consistent with the structural clustering described above, the large majority of complementing pairs comprise one variant from Region A and one variant from Region B of the multichain active site, while the large majority of non-complementing pairs comprise two variants from the same structural region. This result is consistent with the principle that intragenic complementation between loss-of-function variants of a homomultimeric enzyme is determined primarily by the structural positions of the variants relative to the multichain active site, and is independent of the biochemical identities of the substituted amino acids.

[0290] For each pair classified as a complementing pair, the predicted level of functional restoration is quantified by applying an additive model that accounts for the independent allelic contributions of the first variant and the second variant to the quantitative level of restored protein activity. The predicted functional score for the first variant and the predicted functional score for the second variant are each determined from the individual variant functional activity data obtained in the deep mutational scanning experiment described above. The degree of nonadditivity is quantified by comparing the interaction score for each complementing pair to the expected additive combination of the individual predicted functional scores. Combining the complementation classification output of the trained computational model with the additive model achieves an R2of at least 0.99 for predicting the quantitative level of functional restoration across all complementing pairs.

[0291] A complementation map is generated as aheatmap in which the rows and columns correspond to the 82 loss-of-function variants, ordered by residue position, and each cell of the heatmap displays the complementation prediction output for the corresponding pairw ise variant combination. Complementing pairs are displayed in a first color and noncomplementing pairs are displayed in a second color, enabling rapid visual identification of complementation patterns across all pairwise combinations. The heatmap confirms the two-region structural architecture of the multichain active site of Enzyme X, with complementing pairs clustered in the off-diagonal quadrants corresponding to Region A / Region B combinations, and non-complementing pairs clustered in the diagonal quadrants corresponding to within-region combinations.

[0292] A structured complementation data record is generated comprising a plurality of structured records, one for each of the 3,321 pairwise variant combinations evaluated. Each record encodes the virtual residue coordinates of the first variant and the second variant, the local structural context representations of each variant, the complementation predictionPNDRI.020WO PATENToutput, the complementation class label, the predicted functional activity score, and the structural cluster assignment of each residue position. The structured complementation data record is stored on a non-transitory computer-readable storage medium and is made accessible to downstream analysis pipelines for clinical variant interpretation and therapeutic decision-making.Additional Notes

[0293] The embodiments described herein are exemplary. Modifications, rearrangements, substitute processes, etc. may be made to these embodiments and still be encompassed within the teachings set forth herein. One or more of the steps, processes, or methods described herein may be carried out by one or more processing and / or digital devices, suitably programmed.

[0294] The various illustrative imaging or data processing techniques described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. To illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. The described functionality can be implemented in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the disclosure.

[0295] The various illustrative detection systems described in connection with the embodiments disclosed herein can be implemented or performed by a machine, such as a processor configured with specific instructions, a digital signal processor (DSP), an application specific integrated circuit (ASIC), afield programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A processor can be a microprocessor, but in the alternative, the processor can be a controller, microcontroller, or state machine, combinations of the same, or the like. A processor can also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. For example, systems described herein may be implemented using a discrete memory chip, a portion of memory in a microprocessor, flash, EPROM, or other types of memory.PNDRI.020WO PATENT

[0296] The elements of a method, process, or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory. EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of computer-readable storage medium known in the art. An exemplary storage medium can be coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. A software module can comprise computer-executable instructions which cause a hardware processor to execute the computer-executable instructions.

[0297] While the above detailed description has shown, described, and pointed out novel features as applied to illustrative embodiments, it will be understood that various omissions, substitutions, and changes in the form and details of the devices or algorithms illustrated can be made without departing from the spirit of the disclosure. As will be recognized, certain embodiments described herein can be embodied within a form that does not provide all of the features and benefits set forth herein, as some features can be used or practiced separately from others. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.

[0298] It should be appreciated that all combinations of the foregoing concepts (provided such concepts are not mutually inconsistent) are contemplated as being part of the inventive subject matter disclosed herein. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are contemplated as being part of the inventive subject matter disclosed herein.

Claims

1. PNDRI.020WO PATENTWHAT IS CLAIMED IS:

1. A computer-implemented method for predicting complementation or noncomplementation between at least two variants of a multimeric protein, the method comprising:(a) obtaining, or having obtained, an amino acid sequence of a multimeric protein comprising at least one functional site;(b) identifying a first variant at a first residue position and a second variant at a second residue position in the amino acid sequence;(c) obtaining, for each of the first variant and the second variant, a local structural context representation;(d) determining, from the local structural context representation of both the first variant and the second variant, a complementation prediction output; and (e) classify ing, from the complementation prediction output, the first variant and the second variant as a complementing pair or a non-complementing pair,wherein steps (a) through (e) may be performed in any order, concurrently, or iteratively.

2. The method of claim 1, further comprising, prior to step (a), identifying the multimeric protein as a candidate for complementation by:obtaining from a protein structure database at least one homomultimeric protein assembly comprising at least one ligand in contact with more than one subunit; andselecting at least one homomultimeric protein assembly in which a ligand contacts at least two identical subunits within a heavy atom-to-heavy atom distance of 6 A or less.

3. The method of claim 2, wherein the ligand comprises more than 5 atoms and more than 50% of the atoms of the ligand are within 3.6 A of at least two protein chains of the homomultimeric assembly.

4. The method of claim 2 or 3, further comprising identifying additional candidate multimeric proteins by obtaining, from a protein structure database, a structurally similar candidate multimeric protein, wherein the structurally similar candidate multimeric protein has a structure similarity score of at least 0.5.PNDRI.020WO PATENT5. The method of claim 4, further comprising identifying human orthologs of candidate proteins from model organisms using an ortholog database.

6. The method of any one of claims 1-5, wherein the local structural context representation is generated by extracting a sequence window centered on the respective variant residue position.

7. The method of claim 6, wherein the sequence window comprises from 10 to 50 amino acids centered on the respective variant residue position.

8. The method of claim 7, wherein the sequence window comprises 21 amino acids.

9. The method of any one of claims 6-8, wherein the local structural context representation is generated by computing an embedding of the sequence window using a pretrained protein language model.

10. The method of claim 9, wherein the pretrained protein language model is selected from the group consisting ofESM-2, ESM-lb, ESM-lv, ProtTrans, ProtBERT, TAPE, UniRep, SeqVec, AlphaFold-based language model, and any functionally equivalent pretrained protein language model capable of generating local structural context representations of ammo acid sequences.

11. The method of claim 10, wherein the pretrained protein language model is ESM-2.

12. The method of any one of claims 9-11, wherein the embedding comprises a vector of about 500 to about 5,000 parameters.

13. The method of claim 12, wherein the embedding comprises a vector of about 1 ,280 parameters.

14. The method of any one of claims 9-13, wherein the embedding encodes local structural context information from the amino acid sequence, without three-dimensional protein structure coordinates as input.

15. The method of any one of claims 1-14, wherein the local structural context representations of the first variant and the second variant are generated independently of one another.

16. The method of any one of claims 1-15, wherein the local structural context representations are precomputed and reused across multiple predictions.

17. The method of any one of claims 1-16, wherein the complementation prediction output is determined by using a trained computational model that receives the local structural context representations of both the first variant and the second variant as input.PNDRI.020WO PATENT18. The method of claim 17, wherein the trained computational model comprises a neural network configured to receive as input the concatenated local structural context representations of the first variant and the second variant.

19. The method of claim 18, wherein the neural network is a transformer-based neural network.

20. The method of any one of claims 17-19, wherein the complementation prediction output is determined without using numerical positions of the first residue position or the second residue position in the amino acid sequence as inputs to the trained computational model.

21. The method of any one of claims 17-20, wherein the trained computational model has been trained on experimental measurements of functional activity for pairwise combinations of variants of a multimeric protein.

22. The method of claim 21, wherein the experimental measurements of functional activity are obtained by a high-throughput yeast-based grow th assay, deep mutational scanning, or other high-throughput functional screening method.

23. The method of claim 21 or 22, wherein the experimental measurements of functional activity comprise measurements for at least 1,000 pairwise variant combinations.

24. The method of claim 23, wherein the experimental measurements of functional activity comprise measurements for at least 3.000 pairwise variant combinations.

25. The method of any one of claims 17-24, wherein the trained computational model has been trained using cross-validation, including 20-fold cross-validation, leave-one-amino-acid-position-out cross-validation, or a combination thereof.

26. The method of any one of claims 17-25. wherein the trained computational model is an ensemble model comprising predictions averaged across a plurality of models trained during cross-validation.

27. The method of any one of claims 1-26, wherein the multimeric protein is a homomultimeric protein comprising two or more identical subunits.

28. The method of claim 27, wherein the homomultimeric protein is a homodimer, homotrimer, homotetramer, homopentamer, or higher-order homomultimer.

29. The method of any one of claims 1-26, wherein the multimeric protein is a heteromultimeric protein comprising at least two non-identical subunits.

30. The method of any one of claims 1-29, wherein the multimeric protein comprises at least two subunits.PNDRI.020WO PATENT31. The method of any one of claims 1-30, wherein the at least one functional site is a multichain active site, ligand-binding site, or catalytic site formed by residues contributed by at least two subunits of the multimeric protein.

32. The method of any one of claims 1-31, wherein the first residue position and the second residue position are in different regions of the amino acid sequence, wherein the regions are defined structurally, empirically through experimental complementation data, or implicitly by a trained computational model.

33. The method of any one of claims 1-32, wherein the first residue position and the second residue position are in different spatial clusters of the protein structure, as determined by hierarchical clustering of complementation data or three-dimensional structural analysis.

34. The method of any one of claims 1-33, wherein the multimeric protein is an enzyme, receptor, transporter, structural protein, or regulatory protein.

35. The method of any one of claims 1-34, wherein the first variant and the second variant are each loss-of-function variants exhibiting no more than 10% of wild-type protein activity when present alone.

36. The method of any one of claims 1-35, wherein the first variant and the second variant are each independently an amino acid substitution, insertion, or deletion relative to a reference sequence.

37. The method of any one of claims 1-36, wherein complementation between the first variant and the second variant is determined by the first residue position and the second residue position in the protein structure and is independent of the biochemical identities of the substituted amino acids.

38. The method of any one of claims 1-37, wherein the first residue position and the second residue position are each located at or near an active site, ligand-binding site, or catalytic residue of the multimeric protein.

39. The method of any one of claims 1-38, wherein the complementation prediction output informs whether the first variant and the second variant produce a nonadditive functional outcome.

40. The method of any one of claims 1-39, wherein the complementation prediction output comprises a continuous predicted growth value, a probability of complementation, a binary' complementation class label, or a direct classification or ranking generated without a separately computed interaction score.PNDRI.020WO PATENT41. The method of any one of claims 1-40, wherein classifying comprises applying a predetermined decision threshold to the complementation prediction output.

42. The method of claim 41, wherein the decision threshold is selected such that predicted complementation prediction output above the threshold is classified as complementing pairs and complementation prediction output below the threshold is classified as non-complementing pairs.

43. The method of any one of claims 1-42, wherein a complementing pair exhibits restoration of at least 50% of wild-tj pe protein activity.

44. The method of claim 43, wherein a complementing pair exhibits restoration of at least 60% of wild-type protein activity.

45. The method of any one of claims 17-44. wherein the trained computational model achieves a classification accuracy of at least 0.90 on test data.

46. The method of claim 45, wherein the trained computational model achieves a classification accuracy of at least 0.95 on test data from the same multimeric protein used to train the trained computational model.

47. The method of claim 46, wherein the trained computational model achieves a classification accuracy of at least 0.95 on test data comprising variant combinations at residue positions not included in the training data.

48. The method of any one of claims 17-47, wherein the trained computational model achieves an area under the receiver operating characteristic curve (AUC-ROC) of at least 0.85.

49. The method of any one of claims 17-48, wherein the trained computational model achieves an R2of at least 0.84.

50. The method of any one of claims 1-49, further comprising, for a pair classified as a complementing pair, quantifying the predicted level of functional restoration by applying an additive model, wherein the additive model accounts for the independent allelic contributions of the first variant and the second variant to the quantitative level of restored protein activity.

51. The method of claim 50, wherein combining the complementation classification with the additive model achieves an R2of at least 0.99.

52. The method of any one of claims 17-51, wherein the trained computational model has been trained on variant combinations from a first multimeric protein and is applied to predict complementation in a second multimeric protein that is structurally similar to but sequence-divergent from the first multimeric protein.PNDRI.020WO PATENT53. The method of claim 52, wherein the first multimeric protein and the second multimeric protein share less than 25% amino acid sequence identity.

54. The method of claim 52 or 53, wherein the first multimeric protein and the second multimeric protein are members of the same structural family or superfamily.

55. The method of any one of claims 52-54, wherein the trained computational model achieves a classification accuracy of at least 0.90 on variant combinations from the second multimeric protein.

56. The method of any one of claims 1-55, further comprising:identifying a plurality of additional variants at additional residue positions in the amino acid sequence; anddetermining complementation prediction outputs for a plurality of pairwise or higher-order combinations of variants.

57. The method of claim 56, further comprising generating a complementation map, heatmap, interaction network, or visual representation identify ing which variant combinations are predicted to complement one another.

58. The method of claim 56 or 7, further comprising clustering the variants by residue position based on predicted or observed complementation patterns, thereby identifying two or more spatially different regions of the amino acid sequence that contribute to multichain active site assembly.

59. The method of claim 58, wherein variants residing within the same local region do not complement each other, and variants residing in different local regions complement each other.

60. The method of any one of claims 56-59, wherein the plurality of pairwise combinations comprises all pairwise combinations of loss-of-function variants at two or more active site or ligand-binding site residue positions.

61. The method of any one of claims 1-60, further comprising:determining a predicted functional score for the first variant; determining a predicted functional score for the second variant; and quantifying the degree of nonadditivity by comparing the interaction score to an expected additive combination of the individual predicted functional scores.

62. The method of any one of claims 1-61, wherein obtaining the amino acid sequence comprises retrieving the amino acid sequence from a sequence database, receiving user-specified input, or extracting the amino acid sequence from sequencing data.PNDRI.020WO PATENT63. The method of any one of claims 1-62, wherein identifying the first variant and the second variant comprises receiving user-specified variants, analyzing sequencing data from a subject, or computationally generating variant combinations from a set of known loss-of-function variants.

64. The method of any one of claims 1-63, wherein the method is performed without experimental screening of variant combinations prior to prediction.

65. The method of any one of claims 1-64, further comprising preprocessing the amino acid sequence to identify active site residues, ligand-binding residues, or structurally important positions prior to identifying the first variant and the second variant.

66. The method of any one of claims 1-65, further comprising using the classification to inform a therapeutic approach selected from the group consisting of gene therapy, cell therapy, enzy me replacement therapy, protein replacement therapy, substrate reduction therapy, and small molecule targeting.

67. The method of any one of claims 1-66, further comprising using the classification to predict the phenotypic severity of a compound heterozygous genotype in a subject, wherein a classification of complementation pair is associated with a predicted milder disease phenotype relative to a homozygous loss-of-function genotype.

68. The method of any one of claims 1-67, further comprising validating the classification by experimentally measuring functional activity of a compound heterozygous organism or cell expressing both the first variant and the second variant.

69. The method of claim 68, wherein the compound heterozygous organism is a yeast diploid strain harboring the first variant on one allele and the second variant on a second allele, and wherein functional activity is measured by a growth assay under selective conditions.

70. A non-transitory computer readable medium comprising instructions that, when executed by a processor, cause the processor to perform the computer-implemented method of any one of claims 1-69.

71. A system comprising:a processor; anda non-transitory' computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the computer- implemented method of any one of claims 1-69.-1 OS-PNDRI.020WO PATENT72. A computer-implemented method of generating a dataset for modeling complementation in multimeric proteins, the method comprising:generating, via a high-throughput functional assay, experimental measurements of functional activity for a plurality of pairwise combinations of variants of a multimeric protein, wherein each variant comprises a modification at a respective residue position in an amino acid sequence;constructing a dataset comprising a plurality of entries, each entry corresponding to a respective pairwise combination of variants and comprising:(i) identifiers of the pairwise combination of variants, (ii) a quantitative measurement of functional activity associated with the pairwise combination, and(iii) a label indicating a degree of non-additive interaction between the variants; and(iv) organizing the dataset such that variant pairs exhibiting complementation are distinguishable from variant pairs that do not exhibit complementation,wherein the dataset encodes relationships between spatially distinct residue positions that contribute to restoration of protein function upon multimer assembly.

73. The method of claim 72, wherein the label indicating the degree of non-additive interaction comprises a deviation from an expected additive functional effect of the individual variants.

74. The method of claim 72, wherein the dataset comprises at least 1,000 pairwise variant combinations.

75. The method of claim 72, wherein the experimental measurements are obtained using a yeast-based growth assay.

76. The method of claim 72, wherein the dataset encodes functional interactions between residue positions independent of explicit three-dimensional structural coordinates.

77. A non-transitory computer-readable medium storing a dataset for modeling complementation in multimeric proteins, the dataset comprising a plurality of entries, each entry corresponding to a pairwise combination of variants of a multimeric protein, each entry' comprising:(i) identifiers of a first variant and a second variant at respective residue positions,PNDRI.020WO PATENT(ii) a quantitative functional activity measurement obtained from experimental evaluation of the pairwise combination, and(iii) a non-additivity metric representing deviation from an expected additive functional outcome;wherein the dataset is structured to encode complementation relationships between variants located in different spatial regions of the protein, such that the dataset enables identification of variant pairs that restore at least partial functional activity upon multimer assembly.

78. The non-transitory computer-readable medium of claim 77, wherein the dataset is configured to train a computational model to predict complementation between variants without requiring structural coordinate input.

79. The non-transitory computer-readable medium of claim 77, wherein the quantitative functional activity measurement is derived from deep mutational scanning.

80. The non-transitory computer-readable medium of claim 77, wherein entries are organized to reflect clustering of residue positions into two or more spatially distinct complementation groups.

81. A computer-implemented method of training a computational model, the method comprising:obtaining a dataset according to any of claims 72-80;training the computational model using the dataset to learn relationships between variant pairs and functional activity outcomes; andgenerating, using the trained computational model, a prediction of whether a pair of variants exhibits complementation,wherein the trained computational model is configured to predict complementation for variant pairs not included in the dataset.

82. The computer-implemented method of claim 81, wherein the trained computational model is capable of predicting complementation in a second multimeric protein that is sequence-divergent from a protein used to generate the dataset.

83. The computer-implemented method of claim 81. wherein the prediction is generated without using three-dimensional structural coordinates of the multimeric protein.

84. The computer-implemented method of claim 81. wherein the trained computational model achieves a classification accuracy of at least 0.90 in predicting complementation.

Images