Systems and methods for variant classification using proteomics-informed large language models

Abstract

Systems and methods for predicting the effects of genetic variants are disclosed. A system in accordance with the present disclosure comprises a processor configured to train and refine a machine-learning model using proteomic-labeled data. The processor is further configured to receive an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence and apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to the machine-learning model. The processor is further configured to receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant. The processor is further configured to, in response to the numerical value being above a predetermined threshold, associate the variant with a trait.

Description

[0001] DOCKET NO.: 38120-4362 (11652WO01) PCT Systems And Methods For Variant Classification Using Proteomics-Informed Large Language Models Field The present disclosure generally relates to predictive modeling and, more particularly, to systems and methods for predicting variant classifications using large language models. Background As sequencing efforts have scaled to hundreds of thousands of individuals, rare- variant focused genetic analyses have helped inform human biology. These studies use naturally occurring variation to understand the consequences of disrupting the function of each gene. For a small minority of genes, protein truncation and frameshift variants known as putative loss-of-function (pLoF) variants can be observed in sufficient numbers to enable genetic association studies. For most genes, these pLoF variants are rare, and genetic association studies must aggregate information across rare missense variants of uncertain functional impact to have sufficient power to discover new targets. Effectively aggregating evidence across variants relies on distinguishing likely deleterious or damaging variants from other missense variants, and many approaches have been proposed for this task. One popular strategy involves using an ensemble of expertly tuned functional prediction scores. More recently, large language models (LLMs) have demonstrated promise at distinguishing between deleterious and benign coding variants. Summary The present disclosure provides systems and methods for predicting variant features. The system can include at least one memory storing computer-executable instructions and at least one processor in communication with the at least one memory. The at least one processor is configured to execute the computer-executable instructions to: (1) receive an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; (2) apply the amino acid sequence, the DOCKET NO.: 38120-4362 (11652WO01) PCT amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; (3) receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and (4) in response to the numerical value being above a predetermined threshold, associate the variant with a trait. The present disclosure also provides a computer-implemented method for predicting one or more variant features. The method can be implemented using a system including a computing device including a processor communicatively coupled to a memory device. Additionally, or alternatively, the computer-implemented method can be implemented via one or more local or remote processors, servers, transceivers, memory units, mobile devices, wearables, smart watches, smart contact lenses, smart glasses, augmented reality glasses, virtual reality headsets, mixed or extended reality glasses or headsets, voice or chat bots, Artificial intelligence (AI) bots (including generative AI bots), and / or other electronic or electrical components, which may be in wired or wireless communication with one another. The method comprises (1) receiving an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; (2) applying the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; (3) receiving one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and (4) in response to the numerical value being above a predetermined threshold, associating the variant with a trait. The present disclosure provides at least one non-transitory computer-readable storage media having computer-executable instructions embodied thereon. The computer- executable instructions, when executed by at least one processor, cause the at least one DOCKET NO.: 38120-4362 (11652WO01) PCT processor to: (1) receive an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; (2) apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; (3) receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and (4) in response to the numerical value being above a predetermined threshold, associate the variant with a trait. The present disclosure provides systems and methods for predicting variant features. In some embodiments, a system for predicting one or more variant features is provided. The system can include at least one memory storing computer-executable instructions and at least one processor in communication with the at least one memory. The at least one processor is configured to execute the computer-executable instructions to: (1) receive first variant type data and second variant type data, wherein the first variant type data includes proteomic data for a first variant type and the second variant type data includes proteomic data for a second variant type; (2) label first variant type data with a first numerical value and second variant type data with a second numerical value; (3) train a first machine learning model using the labeled first variant type data and labeled second variant type data; (4) apply third variant type data associated with a third variant type to the first machine learning model previously trained using the labeled first variant type data and the labeled second variant type data; (5) receive one or more first outputs from the first machine learning model, the one or more first outputs comprising a third numerical value, the third numerical value being between the first numerical value and the second numerical value; (6) receive one or more input datasets, each of the plurality of input datasets comprising reference sequence data and variant sequence data; (7) label the one or more input datasets with the one or more first outputs from the first machine learning model; and (8) train a second machine learning model using the labeled one or more input datasets, the second machine learning model configured to predict a variant score associated with one or more effects of a DOCKET NO.: 38120-4362 (11652WO01) PCT variant. The system may include additional, less, or alternate functionality, including that discussed elsewhere herein. The present disclosure also provides a computer-implemented method for predicting one or more variant features. The method can be implemented using a system including a computing device including a processor communicatively coupled to a memory device. Additionally, or alternatively, the computer-implemented method can be implemented via one or more local or remote processors, servers, transceivers, memory units, mobile devices, wearables, smart watches, smart contact lenses, smart glasses, augmented reality glasses, virtual reality headsets, mixed or extended reality glasses or headsets, voice or chat bots, Artificial intelligence (AI) bots (including generative AI bots), and / or other electronic or electrical components, which may be in wired or wireless communication with one another. The method comprises (1) receiving first variant type data and second variant type data, wherein the first variant type data includes proteomic data for a first variant type and the second variant type data includes proteomic data for a second variant type; (2) labelling first variant type data with a first numerical value and second variant type data with a second numerical value; (3) training a first machine learning model using the labeled first variant type data and labeled second variant type data; (4) applying third variant type data associated with a third variant type to the first machine learning model previously trained using the labeled first variant type data and the labeled second variant type data; (5) receiving one or more first outputs from the first machine learning model, the one or more first outputs comprising a third numerical value, the third numerical value being between the first numerical value and the second numerical value; (6) receiving one or more input datasets, each of the plurality of input datasets comprising reference sequence data and variant sequence data; (7) labelling the one or more input datasets with the one or more first outputs from the first machine learning model; and (8) training a second machine learning model using the labeled one or more input datasets, the second machine learning model configured to predict a variant score associated with one or more effects of a variant. The method may include additional, less, or alternate functionality, including those discussed elsewhere herein. DOCKET NO.: 38120-4362 (11652WO01) PCT The present disclosure provides at least one non-transitory computer-readable storage media having computer-executable instructions embodied thereon. The computer- executable instructions, when executed by at least one processor, cause the at least one processor to: (1) receive first variant type data and second variant type data, wherein the first variant type data includes proteomic data for a first variant type and the second variant type data includes proteomic data for a second variant type; (2) label first variant type data with a first numerical value and second variant type data with a second numerical value; (3) train a first machine learning model using the labeled first variant type data and labeled second variant type data; (4) apply third variant type data associated with a third variant type to the first machine learning model previously trained using the labeled first variant type data and the labeled second variant type data; (5) receive one or more first outputs from the first machine learning model, the one or more first outputs comprising a third numerical value, the third numerical value being between the first numerical value and the second numerical value; (6) receive one or more input datasets, each of the plurality of input datasets comprising reference sequence data and variant sequence data; (7) label the one or more input datasets with the one or more first outputs from the first machine learning model; and (8) train a second machine learning model using the labeled one or more input datasets, the second machine learning model configured to predict a variant score associated with one or more effects of a variant. The storage medium may include additional, less, or alternate actions, including those discussed elsewhere herein. Advantages will become more apparent to those skilled in the art from the following description of the preferred embodiments which have been shown and described by way of illustration. As will be realized, the present embodiments may be capable of other and different embodiments, and their details are capable of modification in various respects. Accordingly, the drawings and description are to be regarded as illustrative in nature and not as restrictive. Brief Description Of The Drawings The Figures described below depict various aspects of the systems and methods disclosed therein. It should be understood that each Figure depicts an embodiment of a DOCKET NO.: 38120-4362 (11652WO01) PCT particular aspect of the disclosed systems and methods, and that each of the Figures is intended to accord with a possible embodiment thereof. Further, wherever possible, the following description refers to the reference numerals included in the following Figures, in which features depicted in multiple Figures are designated with consistent reference numerals. There are shown in the drawings arrangements which are presently discussed, it being understood, however, that the present embodiments are not limited to the precise arrangements and instrumentalities shown, wherein: FIG.1 illustrates a block diagram of an example computer system in accordance with an embodiment of the present disclosure. FIG.2 illustrates a block diagram of an example variant prediction computing device that may be used with the example computer system illustrated in FIG.1. FIG.3 illustrates a block diagram of an artificial intelligence (AI) / deep learning (DL) module that may be used with the variant prediction computing device illustrated in FIG.2. FIG.4 illustrates a process for labelling data in accordance with an embodiment of the present disclosure. FIG.5 illustrates an architecture for a machine-learning model in accordance with an embodiment of the present disclosure. FIG.6 illustrates a flow chart of an example computer-implemented process that may be carried out by the example computer system shown in FIG.1. FIG.7 illustrates a block diagram of another artificial intelligence (AI) / deep learning (DL) module that may be used with the variant prediction computing device illustrated in FIG.2. FIG.8A shows a scatter plot of pLoF variants. FIG.8B shows a scatter plot of synonymous variants. FIG.8C shows a scatter plot of missense variants. FIG. 8D shows predictions from the machine learning (ML) model for missense variants in accordance with an embodiment of the present disclosure. DOCKET NO.: 38120-4362 (11652WO01) PCT FIG. 9 is a diagram illustrating the extraction of features from a protein language model in accordance with an embodiment of the present disclosure. FIG. 10 is a diagram illustrating a structure of a second predictive model in accordance with an embodiment of the present disclosure. FIG. 11 illustrates a flow chart of an example computer-implemented process that may be carried out by the example computer system shown in FIG.1. FIG.12A is a graph comparing ESM-1v+proteomics and 5-method ensemble. FIG. 12B is a graph comparing combined ESM-1v+proteomics and 5-method ensemble –log10(p-values) from omnibus tests. FIG. 13 illustrates an example server computing device that may be used with the computer system shown in FIG.1. FIG. 14 illustrates an example user computing device that may be used with the computer system shown in FIG.1. FIG. 15 is a diagram illustrating large scale proteomic assay capture of protein structural alterations. FIG.16A illustrates a two-step machine learning model design in accordance with an embodiment of the present disclosure. FIG. 16B illustrates an evaluation strategy for the two-step model of FIG. 16A, in accordance with an embodiment of the present disclosure. FIG. 17 illustrates how combining the 5-method prediction codings with proteomics model-based ESM-1v prediction methods, in accordance with an embodiment of the present disclosure, led to the highest yield associations. FIG. 18A illustrates the fraction of significant gene-trait pairs versus mean absolute effect size from gene burden analysis. FIG. 18B illustrates the total signal strength versus mean absolute effect size from gene burden analysis. FIG. 18C illustrates the fraction of significant gene-trait pairs versus the fraction of variants excluded from analysis. DOCKET NO.: 38120-4362 (11652WO01) PCT FIG. 18D illustrates the total signal strength versus the fraction of variants excluded from analysis. FIG. 19A illustrates the fraction of significant gene-trait pairs versus mean absolute effect size from gene burden analysis. FIG. 19B illustrates the total signal strength versus mean absolute effect size from gene burden analysis. FIG. 19C illustrates the fraction of significant gene-trait pairs versus the fraction of variants excluded from analysis. FIG. 19D illustrates the total signal strength versus the fraction of variants excluded from analysis. FIGS. 20A-20C illustrate extending burden codings to include both the 5-method ensemble and ESM-1b+proteomics in accordance with an embodiment of the present disclosure. FIG.20A illustrates new missense codings. FIG.20B illustrates adding new codings to the definition file. FIG. 20C illustrates adding new codings to the GENE_P testing framework. FIG. 21A illustrates an assessment of varying thresholds of significance when comparing mask strategies in accordance with embodiments of the present disclosure in replacement of conventional 5 method ensemble strategies. FIG. 21B illustrates an assessment of varying thresholds of significance when comparing mask strategies in accordance with embodiments of the present disclosure in combination with conventional 5 method ensemble strategies. FIG. 22 illustrates a convolutional neural network (CNN) model for ESM-1v feature refinement for a single variant. FIG.23A illustrates a comparison of ESM-1v and 5-method ensemble. FIG.23B illustrates a comparison between combined ESM-1v and 5-method ensemble -log10(p-values) from omnibus tests. DOCKET NO.: 38120-4362 (11652WO01) PCT FIG.24 illustrates the enhanced association strength in combined omnibus tests with ESM-1v+proteomics compared to ESM-1v using 242 positive control gene-traits. FIGS. 25A-25C illustrate plots of the distribution of effect size from pQTL analysis classifying all coding variants from 756 Olink Panel II proteins with a cis-coding variant p < 1x10-8. FIG. 25A illustrates the effect size of all coding variants by predicted type on 756 proteins with a significant coding variant. FIG. 25B illustrates all missense variants stratified by predicted deleteriousness. FIG. 25C illustrates missense variants overlapping ClinVar stratified by likely benign and likely pathogenic variants. FIGS.26A-26D illustrate a correlation analysis of ESM-1b+proteomics predictions with effect size from proteomics analysis compared to state-of-the-art methods. FIG. 26A illustrates effect size per missense variant stratified by variant deleteriousness on 756 validation proteins with a significant coding variant (p < 1x10-8). FIG. 26B illustrates the Spearman Correlation per protein. FIG. 26C illustrates a comparison with AlphaMissense. Each point represents a protein and corresponds to the Spearman correlation between AlphaMissense or ESM-1b+proteomics’s predictions with effect size from proteomic analysis. Among 669 genes with negative correlations from either method, ESM-1b+proteomics shows stronger correlation with effect size 384 times (blue), compared to 285 times for AlphaMissense (green). FIG.26D illustrates a comparison with standard ESM-1b. Among 647 genes with negative correlations from either method, ESM-1b + proteomics shows stronger correlation with effect size 444 times (blue), compared to 203 times for standard ESM-1b (orange). FIGS. 27A-27E illustrate an application of ESM-1b+proteomics on 241 Gold Standard Genes using burden tests. FIG.27A illustrates burden associations for missense only. FIG.27B illustrates the Winsorized mean of -log10(p-value) across 241 gold standard genes using burden tests. FIG.27C illustrates ESM-1b+proteomics versus AlphaMissense -log10(p-value) from 241 gold standard singleton gene-trait association. FIG. 27D illustrates ESM- 1b+proteomics versus standard ESM-1b -log10(p-value) from 241 gold standard singleton gene-trait associations. FIG.27E illustrates the Winsorized mean -log10(p-value) versus mean absolute effect size for all methods. DOCKET NO.: 38120-4362 (11652WO01) PCT FIG.28 illustrates an evaluation of LLMs’ ability to discover rare damaging variants. FIG.29A illustrates a comparison of ESM-1b+proteomics to standard ESM-1b at more stringent p-value thresholds of p-value < 0.5, p-value < 1x10-3, and p-value < 1x10-4. FIG.29B illustrates a comparison of ESM-1b+proteomics to standard ESM-1b at more stringent p-value thresholds of p-value < 0.5, p-value < 1x10-3, and p-value < 1x10-4. FIG.30 illustrates median –log10(p-value) showing a similar trend as Winsorized mean log10(p-value) across 241 gold standard gene-trait pairs using burden tests of missense variants. FIG. 31 illustrates the number of associations (p-value < 0.01) versus mean absolute effect size, wherein each point represents a different deleteriousness threshold for each model. FIGS. 32A-32C illustrate two gene-trait associations identified with ESM-1b+ proteomics as compared to standard ESM-1b and AlphaMissense. FIG. 32A illustrates the number of significant associations p-value < 5x10-9using burden tests across 10 example UK Biobank traits. FIG.32B illustrates PCSK6 with triglyceride-levels association. FIG.32B shows summary statistics and p-values for burden codings included in the gene-based testing framework (AF < 0.1%). The “Sample Counts” column shows genotype counts for homozygous reference, heterozygotes and homozygous alterative counts respectively. The "Variants Included" column displays each variant as a vertical line, with the height corresponding to the magnitude of Z-statistics. Red lines indicate variants with a negative effect size, and blue lines indicate variants with a positive effect size. Exons are shown with different background shades. FIG.32C illustrates SIX1 association with hearing loss. FIG. 33 illustrates a comparison of burden tests across allele frequency and p-value thresholds using 10 UKB traits. Four panels are presented corresponding to different variant inclusion allele frequencies for the burden tests across all genes from 10 UKB traits. The x- axis represents the -log10(p-value) threshold used to determine if an association is significant. The y-axis indicates the total number of significant associations. Each color corresponds to a different method. DOCKET NO.: 38120-4362 (11652WO01) PCT FIGS. 34A-34C illustrate six additional novel associations. FIG. 34A illustrates SPHK2 with mean reticulocyte volume association. FIG. 34B illustrates PPARG with triglycerides association. FIG.34C illustrates RRBP1 with LDL association. FIGS. 35A-35B illustrate a comparison of ESM-1v with proteomics fine tuning over standard ESM-1v. FIG. 35A illustrates significant burden associations (p-value < 0.01) across 241 gold standard genes plotted using only missense variants from different allele frequency bins, where each bin includes all variants up to the specified maximum frequency. FIG. 35B illustrates winsorized mean of -log10(p-value) across 241 gold standard genes. FIG.36 illustrates an ESM model comparison on 20001+ plus star ClinVar variants. The Figures depict preferred embodiments for purposes of illustration only. One skilled in the art will readily recognize from the following discussion that alternative embodiments of the systems and methods illustrated herein may be employed without departing from the principles of the aspects described herein. Description Of Embodiments Large language models (LLMs) have demonstrated great promise in the task of distinguishing deleterious and benign coding variants. Here, large scale proteomics data is used as a tool to refine these LLMs to enhance variant classification and burden coding construction, such as for subsequent rare variant analysis. Combining genetic variation with large-scale proteomics data accounts for information about protein structure alterations and this information can be used to inform LLMs. As an intermediate phenotype, proteomics captures a view of an organism’s state and it has been successfully combined with disease phenotypes. Large scale proteomics data, when combined with DNA variation data on the same individuals, enables characterization of the association between sequence variation and proteomic assay results at scale. Here, proteomics data is used to improve the power of burden association and model refinement to lead to further gains in power. More particularly, in some embodiments, a machine learning (ML) framework for refining a LLM is disclosed. Proteomic training labels are used to refine the LLM, and a neural network, such as a convolutional neural network head, is added to the LLM to pool hidden features from its final transformation layer. Once DOCKET NO.: 38120-4362 (11652WO01) PCT the LLM is trained, it can be used to make predictions on various missense variants. The LLMs trained on proteomic data demonstrated improved power over conventional approaches, as discussed in more detail below. In other embodiments, a ML framework for refining a LLM using variant proteomic-trait-loci (pQTL) data is disclosed. The present embodiments may relate to, inter alia, systems and methods for predicting genetic variants. For example, in some embodiments, the system for predicting genetic variants comprises at least one memory storing computer-executable instructions and at least one processor in communication with the at least one memory. The at least one processor is configured to: (1) receive an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; (2) apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; (3) receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and (4) in response to the numerical value being above a predetermined threshold, associate the variant with a trait. The present embodiments may relate to, inter alia, systems and methods for predicting genetic variants. For example, in some embodiments, the system for predicting genetic variants comprises at least one memory storing computer-executable instructions and at least one processor in communication with the at least one memory. The at least one processor is configured to: (1) receive first variant type data and second variant type data, wherein the first variant type data includes proteomic data for a first variant type and the second variant type data includes proteomic data for a second variant type; (2) label first variant type data with a first numerical value and second variant type data with a second numerical value; (3) train a first machine learning model using the labeled first variant type data and labeled second variant type data; (4) apply third variant type data associated with a third variant type to the first machine learning model previously trained using the labeled first variant type data and the labeled second variant type data; (5) receive one or more first outputs from the first machine learning model, the one or more first outputs comprising a DOCKET NO.: 38120-4362 (11652WO01) PCT third numerical value, the third numerical value being between the first numerical value and the second numerical value; (6) receive one or more input datasets, each of the plurality of input datasets comprising reference sequence data and variant sequence data; (7) label the one or more input datasets with the one or more first outputs from the first machine learning model; and (8) train a second machine learning model using the labeled one or more input datasets, the second machine learning model configured to predict a variant score associated with one or more effects of a variant. In other embodiments, the system for genetic variants comprises at least one memory storing computer-executable instructions and at least one processor in communication with the at least one memory. The at least one processor is configured to: (1) receive one or more inputs, the one or more inputs comprising sequence data associated with a first variant; (2) apply the one or more inputs to a machine learning model, the machine learning model being previously trained on one or more statistical measures of proteomic- trait-loci (pQTL) of variants, the machine learning model configured to predict a variant score associated with one or more effects of genetic variants; and (3) receive one or more outputs from the machine learning model, the one or more outputs associated with one or more effects of the first variant. The systems and methods described herein may include additional, less, or alternate functionality, including that discussed elsewhere herein. “Amino acid sequence,” as used herein, may refer generally to the order of amino acids as they occur in a polypeptide chain. A “primary structure” or a “primary amino acid sequence” may refer to the linear sequence of amino acids beginning at the amino-terminal (N) end and ending at the carboxyl-terminal (C). The sequence of a protein is usually notated as a string of letters, according to the order of the amino acids from the amino-terminal to the carboxyl-terminal of the protein. Either a single or three-letter code may be used to represent each amino acid in the sequence. There are 20 amino acids that occur naturally in nature, which can be represented by a three or single letter code as follows: Alanine (Ala, A); Arginine (Arg, R); Asparagine (Asn, N); Aspartic acid (Asp, D); Cysteine (Cys, C); Glutamic acid (Glu, E); Glutamine (Gln, Q); Glycine (Gly, G); Histidine (His, H); Isoleucine (Ile, I); Leucine (Leu, L); Lysine (Lys, K); Methionine (Met, M); Phenylalanine (Phe, F); Proline (Pro, P); Serine (Ser, S); Threonine (Thr, T); Tryptophan (Trp, W); Tyrosine (Tyr, Y); Valine (Val, V). “Amino acid DOCKET NO.: 38120-4362 (11652WO01) PCT sequence” as used herein also includes nonstandard amino acids, e.g., amino acids that have been chemically modified after they have been incorporated into a protein and those amino acids that occur in living organisms but are not found in proteins. Examples of nonstandard amino acids include Selenocysteine (Sec, U) and Pyrrolysine (Pyl, O). “Amino acid sequence masked out,” as used herein, may refer generally to an amino acid sequence in which repetitive and low-complexity sequences within the amino acid sequence are masked using soft masking and / or hard masking methods. Soft masking indicates masked regions by using lower-case letters. Hard masking overwrites masked regions with a wildcard letter (e.g., N for nucleotides or X for proteins). “Variants,” as used herein, may refer generally to differences in DNA sequence that occur naturally between individuals in a given population. “Synonymous variant,” as used herein, may refer generally to a type of variant which does not change the amino acid in the coded protein. Synonymous variants, in general, do not have a significant, measurable effect on protein levels. “Putative loss of function (pLoF) variant,” as used herein, may refer generally to a type of variant which measurably reduces protein levels. “Missense variant,” as used herein, may refer generally to a type of variant that results in different amino acids being encoded at a particular position in the resulting protein. “Proteomic data” or “proteomics data”, as used herein, may refer to information related to the study of proteins within a biological system. “Proteomics”, as used herein, may refer to the large-scale analysis of proteins, including their structure, function and expression. Proteomics data may encompass a wide range of information, including identification, quantification, and characterization of proteins present in cells, tissues, or organisms. Proteomics data provides a readout of the effect of a sequence variant on a protein. For example, this type of data may include data from protein expression profiling, for example protein abundance data. “Protein abundance data” may refer to a normalized quantitative measurement of expression levels based on protein sequence variation. Features of the proteomics data that may be used as input for training the first model may include, but is not limited to: allele frequency and / or standard error, p-value, and / or beta value from a protein quantitative trait loci (pQTL) analysis on coding variants. Allele frequency is the relative DOCKET NO.: 38120-4362 (11652WO01) PCT frequency of an allele at a particular locus in a population, which may be expressed as a fraction or percentage. The standard error is a standard deviation of its sampling distribution or an estimate of that standard deviation. The p-value is the probability under the assumption of no effect or no difference (null hypothesis), of obtaining a result equal to or more extreme than what was actually observed. The beta value is the probability of making a Type II error (accepting the null hypothesis when the null hypothesis is false). “Variant association analysis”, as used herein, may refer to an assessment of the aggregate effect of rare coding deleterious variants within a gene. “Gene burden association yield”, as used herein, may refer to collapsing all information in a genetic region (e.g. a gene) into one genetic burden score that can be used for association analysis. The association is then analyzed between the burden score and a certain phenotype. “Gene burden coding”, as used herein, may refer to a single collapsed coding that combines genotypic codings across individuals. “App,” as used herein, may refer generally to a software application installed and downloaded on a user computing device and executed to provide an interactive graphical user interface at the user computing device. An app associated with the computer system, as described herein, may be understood to be maintained by the computer system and / or one or more components thereof. Accordingly, a “maintaining party” of the app may be understood to be responsible for any functionality of the app and may be considered to instruct other parties / components to perform such functions via the app. By numerical value is meant a value representing the type of variant. This value is based on a classifier. The classifier may be built using the proteomics data. For example, the classifier may be built using summary statistics from pQTL analysis. The proteomics data may indicate variant results in a complete loss of function (pLoF). This variant may be given a numerical value indicating this, for example, a numerical value of 1 indicating pLoF. If the variant does not have any effect as seen from the proteomics data (i.e. it is a synonymous variant), this may be given a different numerical value, e.g. 0. The classifier may be used to generate continuous labels for fine-tuning a protein language model. The protein language model may have a convolutional network head added to it. Reference, alternative, and DOCKET NO.: 38120-4362 (11652WO01) PCT masked protein sequences are input to the protein language model and latent features are used with the convolutional network head to predict the impact of changes in coding sequences. In some embodiments, a numerical value is calculated from a first predictive model trained for example with data labeled either as pLoF or synonymous. This numerical value may be calculated for further missense variants. This numerical value may indicate the pLoF on a scale depending on the effect of the missense variant as predicted from the first predictive model. The scale may be between 0 and 1. With reference to the one or more input datasets comprising reference sequence data and variant sequence data, these input datasets may be output from a protein language model. The protein language model may be a transformer protein language model. The input data sets may comprise representations of the protein sequences, for example vector representation, obtained from a protein language model. These vectors may be used as input features to the second model. The input data sets may then be labeled with the output from the first machine learning model. That is, the input data sets may be labeled with a variant score, e.g. pLoF score, from proteomics data. Therefore, each sequence in the input data set will be represented by vectors (input features of training data set) and labeled with a variant score. In the example embodiment, the computer system may capture and synthesize data, leveraging machine learning and / or artificial intelligence, to predict the effects of genetic variants, and in particular missense variants. Example, non-limiting applications of the prediction include determining the deleteriousness of the variant, determining novel rare variant associations, conducting burden analyses using positive gene-trait pairs, determining whether to include missense variants in gene burden codings and association tests. The computer system may include any suitable data storage capabilities, such as cloud storage, to access and / or store any of the above data. In that way, the computer system may access and analyze historical and / or current (e.g., real-time or near real-time) data. In some embodiments, the computer system may include or be implemented via one or more local or remote processors, servers, transceivers, memory units, mobile devices, wearables, smart watches, smart contact lenses, smart glasses, augmented reality glasses, virtual reality DOCKET NO.: 38120-4362 (11652WO01) PCT headsets, mixed or extended reality glasses or headsets, voice or chat bots, AI bots (including generative AI bots), and / or other electronic or electrical components, which may be in wired or wireless communication with one another. In the example embodiment, the computer system includes at least one computing device. The computing device is configured to perform the functions that may be more generally described herein as being performed by and / or attributed to the overall computer system. In particular, in the example embodiment, the computing device may be in communication with one or more computing devices associated with a user. These computing devices may include a personal mobile computing device, such as a smart phone, tablet, and the like. The computing device may receive data from the computing device(s), including, first variant type data (e.g., synonymous variant data), second variant type data (e.g., putative loss of function (pLoF) data), third variant type data (e.g., missense variant data), proteomics data, reference sequence data (e.g., amino acid sequence data), variant sequence data (e.g., amino acid sequences with at least one altered amino acid), and masked amino acid sequence data (e.g., amino acid sequences masked out), and / or any other data relevant to predicting the effects of a variant. In some embodiments, the proteomics data comprises high-throughput proteomics data derived from 54,219 human blood samples collected by the UK Biobank (Sun, B. B. et al. Plasma proteomic associations with genetics and health in the UK Biobank. Nature 1–10 (2023)). The computing device may receive portions of first variant type data (e.g., synonymous variant data), second variant type data (e.g., pLoF data), third variant type data (e.g., missense variant data), proteomics data, reference sequence data (e.g., amino acid sequence data), variant sequence data (e.g., amino acid sequences with at least one altered amino acid), and masked amino acid sequence data (e.g., amino acid sequences masked out), and / or any other data from other computing devices. Additionally, or alternatively, the computing device may access portions of such data from one or more databases or other memory devices. The computing device may be configured to aggregate, combine, DOCKET NO.: 38120-4362 (11652WO01) PCT synthesize, parse, compare, and / or otherwise process this data, as described in more detail herein, in order to determine a variant prediction score. The computing device may store any received, retrieved, and / or accessed data in one or more databases, and may store the variant prediction score and / or other generated data in the one or more databases. A database may be any suitable storage location, and may in some embodiments include a cloud storage device such that the database may be accessed by a plurality of computing devices (e.g., a plurality of user analytics computing devices, third- party computing devices, etc.). The database may be integral to the computing device or may be remotely located with respect thereto. The technical effect of the systems and processes described herein may be achieved by performing one or more of the following steps: (1) receive an amino acid sequence, the amino acid sequence with a at least one altered amino acid, and the amino acid sequence masked out; (2) apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; (3) receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and (4) in response to the numerical value being above a predetermined threshold, associate the variant with a trait. Additionally, or alternatively, the technical effect of the systems and processes described herein may be achieved by performing one or more of the following steps: (1) receive first variant type data and second variant type data, the first variant type data includes proteomic data for a first variant type and the second variant type data includes proteomic data for a second variant type; (2) label first variant type data with a first numerical value and second variant type data with a second numerical value; (3) train a first machine learning model using the labeled first variant type data and labeled second variant type data; (4) apply third variant type data associated with a third variant type to the first machine learning model previously trained using the labeled first variant type data and the labeled second variant type data; (5) receive one or more first outputs from the first machine learning model, the one or more first outputs comprising a third numerical value, the third numerical value being DOCKET NO.: 38120-4362 (11652WO01) PCT between the first numerical value and the second numerical value; (6) receive one or more input datasets, each of the plurality of input datasets comprising reference sequence data and variant sequence data; (7) label the one or more input datasets with the one or more first outputs from the first machine learning model; and (8) train a second machine learning model using the labeled one or more input datasets, the second machine learning model configured to predict a variant score associated with one or more effects of a variant. Additionally, or alternatively, the technical effect of the systems and processes described herein may be achieved by performing one or more of the following steps: (1) receiving one or more inputs, the one or more input inputs comprising sequence data associated with a first variant; (2) applying the one or more inputs to a machine learning model, the machine learning model being previously trained on one or more statistical measures of proteomic-trait-loci (pQTL) of variants, the machine learning model configured to predict a variant score associated with one or more effects of genetic variants; and (3) receiving one or more outputs from the machine learning model, the one or more outputs associated with one or more effects of the first variant. At least one of the technical problems addressed by the systems and methods disclosed herein may include: (i) time-consuming, labor-intensive, and costly process for predicting the effects of variants; (ii) limited data being leveraged in the predicting of the effects of variants; (iii) limited qualitative and quantitative analysis of proteins in the candidate protein selection process; and (iv) significant processing power and memory usage required to train machine-learning models. The resulting technical effects may include, for example: (i) the stratification of variants for use in gene-based genetic association testing; (ii) more accurate prediction of the effect of missense variants; (iii) improved qualitative and quantitative analysis of candidate proteins; (iv) ability to receive data from a plurality of different data sources and translating the data into a format for further analysis; (v) increased power of genetic association; and (v) reduced processing time, memory usage, and costs associated with the candidate protein selection process. Example computer systems for predicting the effects of variants are disclosed herein. For example, FIG. 1 depicts a schematic diagram of an example computer system 100. The DOCKET NO.: 38120-4362 (11652WO01) PCT computer system 100 is configured to predict the effects of variants, for example, by calculating a variant prediction score. In one embodiment, the computer system 100 may include and / or facilitate communication between a variant prediction computing device 110 and one or more user computing devices 130 (which may also be referred to as “mobile devices”) and / or between the variant prediction computing device 110 and one or more of third-party devices 140 and / or variant prediction servers 150. The variant prediction computing device 110 may be implemented as a server computing device with artificial intelligence and deep learning functionality. Alternatively, the variant prediction computing device 110 (and / or the user computing devices 130) may be implemented as any device capable of interconnecting to the Internet, including mobile computing device or “mobile device,” such as a smartphone, a “phablet,” or other web- connectable equipment or mobile devices (such as one or more local or remote processors, servers, transceivers, memory units, mobile devices, wearables, smart watches, smart contact lenses, smart glasses, augmented reality glasses, virtual reality headsets, mixed or extended reality glasses or headsets, voice or chat bots, AI bots (including generative AI bots), and / or other electronic or electrical components, which may be in wired or wireless communication with one another). The variant prediction computing device 110 may be in communication with the one or more user computing devices 130, third party devices 140, and / or variant prediction servers 150, such as via wireless communication or data transmission over one or more radio frequency links or wireless communication channels. In the example embodiment, components of the computer system 100 may be communicatively coupled to the Internet through many interfaces including, but not limited to, at least one of a network, such as the Internet, a local area network (LAN), a wide area network (WAN), or an integrated services digital network (ISDN), a dial-up-connection, a digital subscriber line (DSL), a cellular telecommunications connection (e.g., a 3G, 4G, 5G, etc., connection), a cable modem, and a BLUETOOTH connection. The computer system 100 also includes one or more database(s) 120 containing information on a variety of matters. For example, the database 120 may include such information as reference sequence data (e.g., amino acid sequence data), variant sequence DOCKET NO.: 38120-4362 (11652WO01) PCT data (e.g., amino acid sequences with at least one altered amino acid), and masked amino acid sequence data (e.g., amino acid sequences masked out), first variant type data (e.g., synonymous variant data), second variant type data (e.g., pLoF data), third variant type data (e.g., missense variant data), proteomic data and / or any other information used, received, and / or generated by the computer system 100 and / or any component thereof, including such information as described herein. In one embodiment, the database 120 may include a cloud storage device, such that information stored thereon may be securely stored but still accessed by one or more components of the computer system 100, such as, for example, the variant prediction computing device 110, the user computing devices 130, and / or the variant prediction servers 150. In some embodiments, the database 120 may be stored on the variant prediction computing device 110. In an alternative embodiment, the database 120 may be stored remotely from the variant prediction computing device 110 and may be non- centralized. In some embodiments, the user computing devices 130 may be computers that include a web browser or a software application to enable the user computing devices 130 to access the functionality of the variant prediction computing device 110 using the Internet or a direct connection, such as a cellular network connection. The user computing devices 130 may be any device capable of accessing the Internet including, but not limited to, a desktop computer, a mobile device (e.g., a laptop computer, a personal digital assistant (PDA), a cellular phone, a smartphone, a tablet, a phablet, netbook, notebook, smart watches or bracelets, smart glasses, wearable electronics, pagers, virtual reality headsets, augmented reality glasses, voice or chat bots, wearables, etc.), or other web-based connectable equipment. The user computing devices 130 may be used to access a data management app 112 maintained by the variant prediction computing device 110, for example, via a user interface 132 when the data management app 112 is executed on the user computing device 130. A user may use the data management app 112 to provide inputs to the variant prediction computing device 110, view predictions generated by the variant prediction computing device 110, and perform other actions, including those described elsewhere herein. DOCKET NO.: 38120-4362 (11652WO01) PCT The third party devices 140 may be computing devices associated with external sources of data. The variant prediction computing device 110 may request, receive, and / or otherwise access data from the third party devices 140. The third party devices 140 may be any devices capable of interconnecting to the Internet, including a server computing device, a mobile computing device or “mobile device,” such as a smartphone, or other web- connectable equipment or mobile devices. Example user analytics computing devices are disclosed herein. For example, FIG.2 depicts the variant prediction computing device 110 (as shown in FIG. 1), according to an embodiment. In some embodiments, the variant prediction computing device 110 may include a processor 202, a memory 204 (which may be similar to the database 120, also shown in FIG.1), a communication interface 206, and a storage interface 208. The processor 202 is configured to execute instructions, which may be stored in the memory 204. The processor 202 includes one or more processing units (e.g., in a multi-core configuration) and may be configured to execute a plurality of modules. In some embodiments, the processor 202 is operable to execute an artificial intelligence / deep learning (AI / DL) module 210, a variant prediction module 212, and a module 214 that maintains functionality for the data management app 112 (shown in FIG.1). The AI / DL module 210, the variant prediction module 212, and the data management module 214 may include specialized instruction sets, and / or coprocessors. The database 120 and / or the memory 204 may store any data and / or instructions necessary for the modules 210, 212, and 214 to function as described herein. In some embodiments, the database 120 may store first variant type data (e.g., synonymous variant data) 220, second variant type data (e.g., pLoF data) 222, third variant type data (e.g., missense variant data) 224, proteomic data 226, reference amino acid sequence data 230, amino acid sequence with at least one altered amino acid data 232, and amino acid sequence masked out data 234, and / or any other information used, received, and / or generated by the variant prediction computing device 110. In some embodiments, first variant type data (e.g., synonymous variant data) 220, second variant type data (e.g., pLoF data) 222, third variant type data (e.g., missense variant data) 224 includes proteomic data. DOCKET NO.: 38120-4362 (11652WO01) PCT The AI / DL module 210 may execute artificial intelligence and / or deep learning functionality on behalf of the variant prediction module 212. Specifically, the AI / DL module 210 may include any rules, algorithms, training data sets / programs, and / or any other suitable data and / or executable instructions that enable the variant prediction computing device 110 to employ artificial intelligence and / or deep learning to predict variant effects. FIG.3 depicts AI / DL module 210 (as shown in FIG.2), according to an embodiment. In some embodiments, the AI / DL module 210 includes a training set builder module 352 configured to submit one or more queries to the database 120 (shown in FIGS.1 and 2) to retrieve data and / or subsets of data, and to use those subsets of data to build training data sets for generating and / or training a predictive model 358. In example embodiments, the training set builder module 352 is configured to retrieve training data sets from the retrieved subsets of data. Each training data set corresponds to historical data, which may include reference amino acid sequence data 230, amino acid sequence with at least one altered amino acid data 232, and amino acid sequence masked out data 234, and the corresponding deleteriousness score. After the training set builder module 352 generates training data sets, it passes the training data sets to a model trainer module 354, which is configured to apply the model input data fields of each training data set as inputs to the predictive model 358, which may comprise one or more machine learning models. Each of the one or more machine learning models is configured to produce, for each training data set, at least one output intended to correspond to, or “predict,” a value of the at least one result data field of the training data set. Machine learning techniques may be used to train the model to identify and recognize patterns in existing data in order to facilitate making predictions for subsequent new input data. For example, neural networks and / or extreme gradient boosted trees may be used to train and optimize the model. In some embodiments, a classifier is used to generate continuous labels for refining the predictive model 358. For example, in FIG.4, a classifier is built for synonymous and / or variant data using proteomic data 226 (e.g., summary statistics from pQTL analysis). In some embodiments, the classifier is built using first variant type data 220 (e.g., synonymous variant data) and / or second variant type data 222 (e.g., pLoF data). The classifier may then DOCKET NO.: 38120-4362 (11652WO01) PCT be applied to third variant type data 224 (e.g., missense variant data) to generate continuous labels (yassay) for fine-tuning the predictive model 358 (shown in FIG.3). Returning to FIG.3, model trainer module 354 is configured to compare, for each training data set, the at least one output of the model to the at least one result data field of the training data set, and apply a machine learning algorithm to adjust parameters of the model in order to reduce the difference or “error” between the at least one output and the corresponding at least one result data field. In this way, the model trainer module 354 trains the machine learning model to accurately predict a deleteriousness score. In other words, the model trainer module 354 cycles the one or more machine learning models through the training data sets, causing adjustments in the model parameters, until the error between the at least one output and the deleteriousness score falls below a suitable threshold. For example, the model trainer module 354 may be configured to compare, for each training data set, a deleteriousness score as determined by the model to a deleteriousness score as previously determined using traditional testing methods. The model trainer module 354 may then adjust one or more weight values using a machine learning algorithm, such as a backpropagation algorithm, in order to reduce the difference between the deleteriousness score as determined by the model and the deleteriousness score as determined using traditional testing methods. Once the error falls below a suitable threshold, the model trainer module 354 may then upload at least one trained machine learning model to the predictive model 358 for application to new variant data. For example, in some embodiments, the predictive model 358 is used for application for a reference amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out. For example, in some embodiments, the predictive model 358 is used to predict a deleteriousness score for missense variants, the deleteriousness score being on a scale from 0 to 1. In some embodiments, the model trainer module 354 uses low-rank adaptation (LoRA). LoRa is an efficient finetuning technique to adapt large models to specific tasks and datasets. Unlike traditional fine-tuning that requires adjusting the entire model, LoRA focuses on modifying a smaller subset of parameters (lower-rank matrices), thereby reducing computational and memory overhead. LoRA is built on the understanding that DOCKET NO.: 38120-4362 (11652WO01) PCT large models inherently possess a low-dimensional structure. By leveraging low-rank matrices, LoRA adapts these models effectively. This method focuses on the core concept that significant model changes can be represented with fewer parameters, thus making the adaptation process more efficient. This significantly reduces the trainable parameters of the predictive model, thereby increasing the speed at which the predictive model may be trained, decreases the computing resources required for the training process, and improves memory efficiency for the training process. In some embodiments, the predictive model 358 may include one or more neural networks, such as a convolutional neural network, a recurrent neural network (Long Short- Term Memory (LSTM), Gated Recurrent Unit (GRU), etc.), a deep learning neural network, a transformer network, or the like. The neural network may have one or more layers of nodes, and the model parameters adjusted during training may be respective weight values applied to one or more inputs to each node to produce a node output. In other words, the nodes in each layer may receive one or more inputs and apply a weight to each input to generate a node output. The node inputs to the first layer may correspond to the model input data fields and the node outputs of the final layer may correspond to the at least one output of the model, intended to predict the at least one result data field. The node outputs of the final layer may correspond to a deleteriousness score for the variant. One or more intermediate layers of nodes may be connected between the nodes of the first layer and the nodes of the final layer. In some embodiments, the model trainer module 354 provides an advantage by automatically discovering and properly weighting complex, second- or third- order, and / or otherwise nonlinear interconnections between the model input data fields and the at least one output. Absent the machine learning model, such connections are unexpected and / or undiscoverable by human analysts. In some embodiments, the predictive model 358 includes a convolutional neural network head. For example, in FIG.5, the last hidden layer was extracted and concatenated into separate channels. A 1x1 convolution kernel was applied to merge the channels into a single matrix. After this, one or more kernels (e.g., seven kernels) of a predetermined hidden size (e.g., seven x hidden size) may be applied to the resulting matrix with adaptive pooling. In some embodiments, a Gaussian Error Linear Units (GELU) nonlinear activation DOCKET NO.: 38120-4362 (11652WO01) PCT may be used. In some embodiments, a drop out operator after this layer may be included. This output may be passed into a final fully connected layer. A skip connection may be added from the final hidden layer (selecting out the embedding at the amino acid position for the variant) to the fully connected layer. To reduce dimensionality, a 1x1 convolution may be applied to merge across channels before passing into the fully connected layer. In some embodiments, the predictive model 358 compares the known deleteriousness score with the output from the trained model, and routes the comparison result to a model updater module 356 of AI / DL module 210. The model updater module 356 is configured to derive a correction signal from the comparison results, and to provide the correction signal to the model trainer module 354 to enable updating or “re-training” of the at least one machine learning model to improve performance. For example, one or more new weight values may be derived from the comparison results, and the correction signal may adjust weight values applied to one or more inputs. A re-trained machine learning model may be periodically re-uploaded to the predictive model 358. In some embodiments, the first model trainer module 704 may update the training dataset by creating one or more new historical records which includes new data and re- training the operator model using the updated training dataset, further improving the accuracy of the operator model. The predictions and scores determined by the predictive model 358 and other data may be viewable via the data management app 112. The app module 214 is configured to facilitate maintaining data for the management app 112 and providing the functionality thereof to users. The app module 214 may store instructions that enable the download and / or execution of the data management app 112 at the user computing devices 130. The app module 214 may store instructions regarding user interfaces, controls, commands, settings, and the like, and may format data into a format suitable for transmitting to user computing devices 130 for display thereof. In some embodiments, the processor 202 is operatively coupled to the communication interface 206 such that the variant prediction computing device 110 is capable of communicating with remote device(s) such as the user computing devices 130, DOCKET NO.: 38120-4362 (11652WO01) PCT third party devices 140, and / or variant prediction servers 150 (all shown in FIG. 1) over a wired or wireless connection. For example, the communication interface 206 may receive first variant type data (e.g., synonymous variant data), second variant type data (e.g., pLoF data), third variant type data (e.g., missense variant data), reference amino acid sequence data 230, amino acid sequence with substituted amino acid data 232, and amino acid sequence masked out data 234, and the like, from the user computing devices 130, and / or third-party devices 140. The communication interface 206 may include, for example, a wired or wireless network adapter and / or a wireless data transceiver for use with a mobile telecommunications network. The processor 202 may also be operatively coupled to the database 120 (and / or any other storage device) via the storage interface 208. The database 120 may be any computer- operated hardware suitable for storing and / or retrieving data. In some embodiments, the database 120 may be integrated in the variant prediction computing device 110. For example, the variant prediction computing device 110 may include one or more hard disk drives as the database 120. In other embodiments, the database 120 is external to the variant prediction computing device 110 and is accessed by a plurality of computer devices. For example, the database 120 may include a storage area network (SAN), a network attached storage (NAS) system, multiple storage units such as hard disks and / or solid-state disks in a redundant array of inexpensive disks (RAID) configuration, cloud storage devices, and / or any other suitable storage device. The storage interface 208 may be any component capable of providing the processor 202 with access to the database 120. The storage interface 208 may include, for example, an Advanced Technology Attachment (ATA) adapter, a Serial ATA (SATA) adapter, a Small Computer System Interface (SCSI) adapter, a RAID controller, a SAN adapter, a network adapter, and / or any component providing the processor 202 with access to the database 120. The processor 202 may execute computer-executable instructions for implementing aspects of the disclosure. In some embodiments, the processor 202 may be transformed into a special purpose microprocessor by executing computer-executable instructions or by otherwise being programmed. For example, the processor 202 may be programmed with the instructions such as those illustrated in FIGS.6 and 11. DOCKET NO.: 38120-4362 (11652WO01) PCT The memory 204 may include, but is not limited to, random access memory (RAM) such as dynamic RAM (DRAM) or static RAM (SRAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), and non-volatile RAM (NVRAM). The above memory types are example only, and are thus not limiting as to the types of memory usable for storage of a computer program. Example computer-implemented methods for predicting a variant score is disclosed herein. For example, FIG.6 depicts a process 600 for predicting a score in real-time. The process 600 may be executed by one or more processors, such as the processor 202 (shown in FIG.2) of the variant prediction computing device 110 (shown in FIGS.1 and 2). At 602, a machine-learning model is trained and / or refined using proteomic labelled data. Once the machine-learning model is trained and / or refined, an amino acid sequence, the amino acid sequence with a substituted amino acid, and the amino acid sequence is received at 604. The amino acid sequence, the amino acid sequence with the substituted amino acid, and the amino acid sequence masked out is applied to the machine-learning model at 606. Next, at 608, one or more outputs are received from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant. If the numerical value is above a predetermined threshold, the variant is associated with a trait at 610. FIG. 7 depicts an AI / DL module 210 (as shown in FIG. 2), according to an embodiment. In some embodiments, the AI / DL module 210 includes a first training set builder module 702 configured to submit one or more queries to the database 120 (shown in FIGS.1 and 2) to retrieve data and / or subsets of data, and to use those subsets of data to build training data sets for generating a first predictive model 708. In example embodiments, the first training set builder module 702 is configured to retrieve training data sets from the retrieved subsets of data. Each training data set corresponds to historical data, which may include one or more of first variant type data 220, second variant type data 222, proteomic data 226, and the corresponding pLoF score. For example, in some embodiments, first variant type data 220 comprises synonymous variant DOCKET NO.: 38120-4362 (11652WO01) PCT data and second variant type variant data 222 comprises pLoF data, and the output comprises a pLoF score, in which synonymous variants are labeled 0 and pLoF variants are labeled 1. The historical data may comprise data which has been previously classified as either synonymous or pLoF, as opposed to completed in real-time with respect to the time of retrieval by the training set builder module 702. Each training data set can include model input data along with result data representing a pLof score (e.g., a synonymous variant (labeled 0) or a pLoF variant (labeled 1)). The model input data can represent factors that may be expected to, or unexpectedly be found during model training, to have some correlation with the pLoF score. In some embodiments, the model input data may comprise proteomic data. For example, in some embodiments, the input data may comprise one or more of log10 p-value, allele frequency, proteomic-trait-loci (pQTL) effect size, and standard error from proteomic association models. After the training set builder module 702 generates training data sets, it passes the training data sets to a first model trainer module 704, which is configured to apply the model input data fields of each training data set as inputs to the first predictive model 708, which may comprise one or more machine learning models. In some embodiments, the first predictive model 708 comprises an ensemble of two (or more) predictions. Each of the one or more machine learning models is configured to produce, for each training data set, at least one output intended to correspond to, or “predict,” a value of the at least one result data field of the training data set. Machine learning techniques may be used to train the model to identify and recognize patterns in existing data in order to facilitate making predictions for subsequent new input data. For example, neural networks and / or extreme gradient boosted trees may be used to train and optimize the model. The first model trainer module 704 is configured to compare, for each training data set, the at least one output of the model to the at least one result data field of the training data set, and apply a machine learning algorithm to adjust parameters of the model in order to reduce the difference or “error” between the at least one output and the corresponding at least one result data field. In this way, the first model trainer module 704 trains the machine learning model to accurately classify variants. In other words, the first model trainer module 704 cycles the one or more machine learning models through the training data sets, DOCKET NO.: 38120-4362 (11652WO01) PCT causing adjustments in the model parameters, until the error between the at least one output and the pLof score falls below a suitable threshold. For example, the first model trainer module 704 may be configured to compare, for each training data set, a pLof score for variant data as determined by the model to a pLof score for the variant data as previously determined using traditional testing methods. The first model trainer module 704 may then adjust one or more weight values using a machine learning algorithm, such as a backpropagation algorithm, in order to reduce the difference between the pLof score as determined by the model and the pLof score as determined using traditional testing methods. Once the error falls below a suitable threshold, the first model trainer module 704 may then upload at least one trained machine learning model to the first predictive model 708 for application to new variant data. For example, in some embodiments, the first predictive model 708 is used for application for third variant type data, such as missense variant data. For example, in some embodiments, the first predictive model 708 is used to predict a pLoF score for missense variants, the pLoF score being on a scale from 0 to 1. FIGS. 8A-8D are scatter plots showing cis-pQTL effect size and p-value stratified by variant annotation type. More particularly, in FIGS.8A-8D, the x-axis corresponds to the effect size from pQTL and the y-axis corresponds to the p-value. FIG.8A is a scatter plot showing the pQTL p-value of pLof variants. The x-axis corresponds to the effect size from pQTL and the y- axis corresponds to the p-value. FIG. 8B is a scatter plot showing the pQTL p-value of synonymous variants. FIG.8C is a scatter plot showing the pQTL p-value of missense variants. FIG.8D is a scatter plot showing the pQTL p-value of predictions from the machine learning model for missense variants, the machine learning model previously trained on pLoF (labeled 1) and synonymous variants (labeled 0) using four features (allele frequency, and beta, p- value, and standard error from the pQTL analysis) for Olink Panel I proteins. To generate the plot of FIG.8D, the machine learning model was then applied to missense variants to get a probabilistic prediction for how much a missense variant looks like a pLoF. The pLoF score increases along with pQTL effect size, as demonstrated in the graphs shown in FIGS. 8A-8D, which suggests that the pLoF score is capturing biological signal related to protein function. In some embodiments, the first predictive model 308 may include one or more neural networks, such as a convolutional neural network, a deep learning neural network, a DOCKET NO.: 38120-4362 (11652WO01) PCT transformer, or the like. The neural network may have one or more layers of nodes, and the model parameters adjusted during training may be respective weight values applied to one or more inputs to each node to produce a node output. In other words, the nodes in each layer may receive one or more inputs and apply a weight to each input to generate a node output. The node inputs to the first layer may correspond to the model input data fields and the node outputs of the final layer may correspond to the at least one output of the model, intended to predict the at least one result data field. For example, the node inputs to the first layer may correspond to one or more features of a variant. For example, in some embodiments, the one or more features comprise allele frequency, and / or beta, p-value, and / or standard error from proteomic association models. The node outputs of the final layer may correspond to a pLoF score for the variant. One or more intermediate layers of nodes may be connected between the nodes of the first layer and the nodes of the final layer. In some embodiments, the first model trainer module 704 provides an advantage by automatically discovering and properly weighting complex, second- or third-order, and / or otherwise nonlinear interconnections between the model input data fields and the at least one output. Absent the machine learning model, such connections are unexpected and / or undiscoverable by human analysts. Additionally, or alternatively, the first predictive model 308 may include extreme gradient boosted trees, which comprise decision trees that utilize a gradient boosting technique. More particularly, extreme gradient boosting builds a predictive model by aggregating the predictions of several weak learner decision trees. It uses a boosting technique to create an extremely accurate ensemble model after it corrects the mistakes of its predecessors. In some embodiments, the first predictive model 708 comprises an ensemble approach in which the output is an average of a plurality of models. For example, in some embodiments, the output of the first predictive model 708 comprises an average of a first model (e.g., a neural network) and a second model (e.g., gradient boosted trees). In some embodiments, the first predictive model 708 compares the known pLof score with the output from the trained model, and routes the comparison result to a first model updater module 706 of AI / DL module 210. The first model updater module 706 is DOCKET NO.: 38120-4362 (11652WO01) PCT configured to derive a correction signal from the comparison results, and to provide correction signal to the first model trainer module 704 to enable updating or “re-training” of the at least one machine learning model to improve performance. For example, one or more new weight values may be derived from the comparison results, and the correction signal may adjust weight values applied to one or more inputs. A re-trained machine learning model may be periodically re-uploaded to the first predictive model 708. In some embodiments, the first model trainer module 704 may update the training dataset by creating one or more new historical records which includes new data and re- training the operator model using the updated training dataset, further improving the accuracy of the operator model. In some embodiments, the first predictive model 708 is used to determine pLoF scores for a plurality of missense variants. This data may be stored in the database 120 (shown in FIGS. 1 and 2). In some embodiments, the pLoF score is on a scale from 0 to 1, and the missense variants and their associated pLoF score is then used as labeled data within a second predictive model 718. Therefore, in some embodiments, the AI / DL module 210 includes a second training set builder module 712 configured to receive data from first predictive model 708 and / or submit one or more queries to the database 120 (shown in FIGS. 1 and 2) to retrieve data and / or subsets of data, and to use those subsets of data to build training data sets for generating the second predictive model 718. In other embodiments, the first predictive model 708 is optional. For example, summary statistics from the pQTL (e.g., effect size, standard error, etc.) are stored in the database 120 (shown in FIGS. 1 and 2), and the second training set builder module 712 is configured to receive summary statistics data from the pQTL and / or submit one or more queries to the database 120 (shown in FIGS.1 and 2) to retrieve data and / or subsets of data and to use those subsets of data to build training data sets for generating the second predictive model 718, which do not include labeled data from the first predictive model 708. In example embodiments, the second training set builder module 712 is configured to retrieve training data sets from the retrieved subsets of data. Each training data set corresponds to historical data. In some embodiments, the historical data comprises proteomic data. In such embodiments, the historical data can also include one or more DOCKET NO.: 38120-4362 (11652WO01) PCT features from dbNSFP. Additionally, or alternatively, the historical data can include previously determined features for missense variants labeled with their associated pLoF scores as determined by the first predictive model 708. In some embodiments, the previously determined features may comprise features derived from a protein language model (e.g., ESM-1v derived features), dbNSFP derived features, or both. In embodiments which exclude the first predictive model 708, the historical data can include summary statistics data from the pQTL. The historical data may comprise data which has been previously determined, as opposed to completed in real-time with respect to the time of retrieval by the training set builder module 712. Each training data set can include model input data along with result data representing one or more features of the missense variant. The model input data can represent factors that may be expected to, or unexpectedly be found during model training to, have some correlation with the pLof score. In some embodiments, the model input data may comprise one or more of features derived from other models (e.g., ESM-1v and / or dbNSFP). In some embodiments, the second training set builder 712 is configured to transform historical data prior to training. For example, in some embodiments, features are extracted from outputs of existing models by capturing vectors related to context and sequence change. In other embodiments, the second predictive model 718 can be trained with the proteomic data directly, i.e., without extracting any features. In such embodiments, the second predictive model 718 can also be augmented with additional features from dbNSFP. FIG.9 is a diagram illustrating the extraction of features from a protein language model (e.g., ESM-1v model). At 902, the existing model is applied to a reference amino acid sequence (e.g., an amino acid sequence without alteration) and an altered amino acid sequence (e.g., an amino acid sequence altered by a missense variant) separately. The internal layers are applied to the reference amino acid sequence and the altered amino acid sequence at 904. At 906, a matrix 920 is extracted from one or more layers (e.g., the final transformer layer) of the existing model. For example, in the embodiment illustrated in FIG.9, features are extracted at 508 to formulate (a) a first vector 921 comprising an average vector of all sequence position vectors, or any other vector summarization, spanning a reference amino acid sequence; (b) a second vector 922 comprising a sequence vector at a variant position for the reference amino acid DOCKET NO.: 38120-4362 (11652WO01) PCT sequence; (c) a third vector 923 comprising a gene context vector for an altered amino acid sequence; (d) a fourth vector 924 comprising a sequence vector at the sequence variant position for the altered amino acid sequence. In some embodiments, first vector 921 , second vector 922, third vector 923, and fourth vector 924 have a length of 1280. First vector 921 , second vector 922, third vector 923, and fourth vector 924 may be concatenated to form matrix 920. A row 926 may be added to matrix 920. Row 926 may comprise protein language model scores for the variant change for first vector 921 and third vector 923, and zeroes for second vector 922 and fourth vector 924. In the original protein language model determination, there is an output layer 910 comprising calculated protein language model scores. Returning to FIG. 7, after the second training set builder module 712 generates training datasets, it passes the training data sets to a first model trainer module 714, which is configured to apply the model input data fields of each training set as inputs to a second predictive model, which may comprise one or more machine learning models. Each of the one or more machine learning models is configured to produce, for each training data set, at least one output predicting a value of the at least one result data field of the training data set. Machine learning techniques may be used to train the model to identify and recognize patterns in existing data in order to facilitate making predictions for subsequent new input data. In some embodiments, missense variants and their associated pLoF scores predicted by first predictive model 708 are used to train and optimize matrix 920 (shown in FIG.9). More particularly, the missense variants and their associated pLoF scores may be used as continuous training labels on the probability scale, as described in more detail below. The second model trainer module 714 is configured to compare, for each training data set, the at least one output of the model to the at least one result data field of the training data set, and apply a machine learning algorithm to adjust parameters of the model in order to reduce the difference or “error” between the at least one output and the corresponding at least one result data field. In this way, the second model trainer module 714 trains the machine learning model to accurately predict a variant score. In other words, the first model trainer module 704 cycles the one or more machine learning models through the training data sets, causing adjustments in the model parameters, until the error between the DOCKET NO.: 38120-4362 (11652WO01) PCT at least one output and the variant score falls below a suitable threshold. For example, the second model trainer module 714 may be configured to compare, for each training data set, a variant for variant data as determined by the model to a pLoF score for the variant data as previously determined using the first predictive model 708. The second model trainer module 714 may then adjust one or more weight values using a machine learning algorithm, such as a backpropagation algorithm, in order to reduce the difference between the classification as determined by the model and the classification as determined using traditional testing methods. Once the error falls below a suitable threshold, the second model trainer module 714 may then upload at least one trained machine learning model to the second predictive model 718 for application to new variant data. In some embodiments, the variant score is on a scale from 0 to 1. In some embodiments, the second predictive model may include one or more neural networks, such as a convolutional neural network, a deep learning neural network, or the like. The neural network comprises one or more layers of nodes, and the model parameters adjusted during training may be respective weight values applied to one or more inputs to each node to produce a node output. In other words, the nodes in each layer may receive one or more inputs and apply a weight to each input to generate a node output. The node inputs to the first layer may correspond to the model input data fields and the node outputs of the final layer may correspond to the at least one output of the model, intended to predict the at least one result data field. For example, the node inputs to the first layer may correspond to one or more features of a variant. The node outputs of the final layer may correspond to a variant score for the variant. One or more intermediate layers of nodes may be connected between the nodes of the first layer and the nodes of the final layer. In some embodiments, the second model trainer module 714 provides an advantage by automatically discovering and properly weighting complex, second- or third-order, and / or otherwise nonlinear interconnections between the model input data fields and the at least one output. Absent the machine learning model, such connections are unexpected and / or undiscoverable by human analysts. In some embodiments, the second predictive model 718 comprises a convolutional neural network. By leveraging a convolutional neural network, each row of the input matrix DOCKET NO.: 38120-4362 (11652WO01) PCT (e.g., matrix 920 shown in FIG.9) could have a kernel applied to it, as shown in FIG.10, thereby dramatically reducing the parameter space as compared to a standard feed forward network. As a non-limiting example, in FIG. 10, 301x4 kernels 1004 are applied to the 1281x4 input matrix 1002 to create a 1281x30 matrix 1006. Next, one 1x30 kernel 1008 is applied to summarize the resulting tensor to a 1281x1 vector. Next, 30 fully connected linear units 1012 with Gaussian Error Linear Unit (GELU) activation are applied to the 1281x1 vector 1010. A final fully connected layer is applied with a sigmoid activation function to give an output 1014 that ranges from 0 to 1, where 1 most closely resembles a pLoF variant. During training, the predicted value is compared to the proteomic derived label (continuous scale). In some embodiments, the second predictive model 718 compares the known variant score with the output from the trained model, and routes the comparison result to a second model updater module 716 of AI / DL module 210. The second model updater module 716 is configured to derive a correction signal from the comparison results, and to provide correction signal to the model trainer module 714 to enable updating or “re-training” of the at least one machine learning model to improve performance. For example, one or more new weight values may be derived from the comparison results, and the correction signal may adjust weight values applied to one or more inputs. A re-trained machine learning model may be periodically re-uploaded to the second predictive model 718. In some embodiments, the second model trainer module 714 may update the training dataset by creating one or more new historical records which includes new data and re-training the operator model using the updated training dataset, further improving the accuracy of the operator model. The predictions and scores determined by the first predictive model 708 and the second predictive model 718 and other data may be viewable via the data management app 112. Example computer-implemented methods for predicting a variant score is disclosed herein. For example, FIG.11 depicts a process 1100 for predicting a variant score in real-time. The process 1100 may be executed by one or more processors, such as the processor 202 (shown in FIG.2) of the variant prediction computing device 110 (shown in FIGS.1 and 2). At 1102, pQTL association of coding variants and proteomic binding assays is received. At 1104, DOCKET NO.: 38120-4362 (11652WO01) PCT synonymous variants are labeled 0 and pLoF variants are labeled 1. Next, at 1106 a first predictive model is trained using the labeled synonymous variant data and the pLoF variant data labeled 1. In some embodiments, the first predictive model comprises an ensemble of two predictions. At 1108, missense variant data (e.g., amino acid sequences of missense variants) are applied to the first predictive model. One or more outputs are received from the first predictive model at 1110, the one or more outputs comprising pLoF scores for a plurality of missense variants. Separately, at 1112, reference sequence data (e.g., an amino acid sequence without a variant) and variant sequence data (e.g., the amino acid sequence with the variant) is applied to a protein language model. At 1114, features are extracted from a final transformer layer of the protein language model to create training datasets. The training datasets are then labeled using the pLoF scores determined by the first predictive model at 1116. Next, the second predictive model is trained using the labeled input datasets at 1118. New missense variant data (e.g., an amino acid sequence comprising a variant) is then applied to the second predictive model at 1120. Next, at 1122, one or more outputs are received from the second predictive model, the one or more outputs including a variant score for the missense variant. FIGS. 12A-12B are graphs showing enhanced association strength in combined omnibus tests with ESM-1v+proteomics compared to a 5-method ensemble approach using 242 positive control gene-traits. More particularly, FIG. 12A is a scatter plot showing a comparison of ESM-1v+proteomics and 5 method ensemble. Each point of the scatter plot represents a gene-trait pair association. The x-axis displays the -log10(p-value) of the 5- method ensemble derived from an omnibus test truncated at 300, while the y-axis shows the -log10(p-value) of ESM-1v+proteomics from an omnibus test. FIG. 12B is a scatter plot showing a comparison between combined ESM-1v+proteomics and 5-method ensemble – log10(p-values) from omnibus tests. Example server computing devices are disclosed herein. For example, FIG. 13 is a schematic diagram of an example configuration of a server computing device 1300, in accordance with some embodiments of the present disclosure. Server computing devices having an architecture similar to the server computing device 1300 may be used to implement one or more of the computing systems shown in FIG. 1, such as the variant prediction DOCKET NO.: 38120-4362 (11652WO01) PCT computing device 110. In the example embodiment, the server computing device 1300 includes a processor 1305 for executing instructions (not shown) stored in a memory 1310. In an embodiment, the processor 1305 may include one or more processing units (e.g., in a multi-core configuration). The instructions may be executed within various different operating systems, such as UNIX®, LINUX® (LINUX is a registered trademark of Linus Torvalds), Microsoft Windows®, etc. It should also be appreciated that upon initiation of a computer- based method, various instructions may be executed during initialization. Some operations may be required in order to perform one or more processes described herein, while other operations may be more general or specific to a particular programming language (e.g., C, C#, C++, Java, or other suitable programming languages, etc.). In the example embodiment, the processor 1305 is operatively coupled to a communication interface 1315 such that the server computing device 1300 is capable of communicating with a remote device, such as a user or system administrator computing system (not shown) or another server computing device 1300. In the example embodiment, the processor 1305 is also operatively coupled to a storage device 1330, which may be, for example, a computer-operated hardware unit suitable for storing or retrieving data. In some embodiments, the storage device 1330 is integrated into the server computing device 1300. For example, the device 1300 may include one or more hard disk drives as the storage device 1330. In other embodiments, the storage device 1330 is external to the device 1300 and may be accessed by a plurality of server computing devices 1300. For example, storage the device 1330 may include multiple storage units such as hard disks or solid-state disks in a redundant array of inexpensive disks (RAID) configuration. The storage device 1330 may include a storage area network (SAN) or a network attached storage (NAS) system. The storage device 1330 may be used as a repository for one or more databases or other data structures for storing various data elements received, processed, and / or generated by the variant prediction computing device 110. For example, the storage device 1330 is used to implement the database 120 (shown in FIGS.1 and 2). In some embodiments, the processor 1305 is operatively coupled to the storage device 1330 via an optional storage interface 1320. The storage interface 1320 may include, DOCKET NO.: 38120-4362 (11652WO01) PCT for example, a component capable of providing the processor 1305 with access to the storage device 1330. In one embodiment, the storage interface 1320 further includes one or more of an Advanced Technology Attachment (ATA) adapter, a Serial ATA (SATA) adapter, a Small Computer System Interface (SCSI) adapter, a RAID controller, a SAN adapter, a network adapter, or a similarly capable component providing the processor 1305 with access to the storage device 1330. The memory area 1310 may include, but is not limited to, random-access memory (RAM) such as dynamic RAM (DRAM) or static RAM (SRAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), non-volatile RAM (NVRAM), and magneto-resistive random- access memory (MRAM). The above memory types are for example only, and are thus not limiting as to the types of memory usable for storage of a computer program. Example user computing devices are disclosed herein. For example, FIG. 14 illustrates an example configuration of a user computing device 1402. The user computing device 1402 includes a processor 1404 for executing instructions. In some embodiments, executable instructions are stored in a memory area 1406. The processor 1404 may include one or more processing units (e.g., in a multi-core configuration). The memory area 1406 is any device allowing information such as executable instructions and / or other data to be stored and retrieved. The memory area 1406 may include one or more computer-readable media. The user computing device 1402 also includes at least one media output component 1408 for presenting information to a user. The media output component 1408 is any component capable of conveying information to a user. In some embodiments, the media output component 1408 includes an output adapter such as a video adapter and / or an audio adapter. An output adapter is operatively coupled to the processor 1404 and operatively couplable to an output device such as a display device (e.g., a liquid crystal display (LCD), organic light emitting diode (OLED) display, cathode ray tube (CRT), or “electronic ink” display) or an audio output device (e.g., a speaker or headphones). For example, a user may view variant information via the media output component 1408. DOCKET NO.: 38120-4362 (11652WO01) PCT In some embodiments, the user computing device 1402 includes an input device 1410 for receiving input from user 1401. The input device 1410 may include, for example, a keyboard, a pointing device, a mouse, a stylus, a touch sensitive panel (e.g., a touch pad or a touch screen), a camera, a gyroscope, an accelerometer, a position detector, and / or an audio input device. A single component such as a touch screen may function as both an output device of the media output component 1408 and the input device 1410. The user computing device 1402 may also include a communication interface 1412, which is communicatively couplable to a remote device such as a server system or a web server operated by a bioregistry. The communication interface 1412 may include, for example, a wired or wireless network adapter or a wireless data transceiver for use with a mobile phone network (e.g., Global System for Mobile communications (GSM), 3G, 4G or Bluetooth) or other mobile data network (e.g., Worldwide Interoperability for Microwave Access (WIMAX)). Examples Example 1: Large scale proteomics and LLMs Proteomic data was collected across 46,665 sequenced UK Biobank participants to refine LLM models for protein sequences. Proteomics can be utilized to establish connections between genetic variants, protein abundance, and diseases, thereby providing valuable insights into disease mechanisms and potential therapeutic targets. While proteomic data is expected to measure protein abundance, it also captures protein structure alterations. For example, previous studies suggest both that (1) 86% of missense cis-pQTLs (p<10-8) variants are associated with decreased protein readouts, and that (2) missense cis-pQTLs are enriched in predicted damaging variants. Furthermore, previous studies found that missense variant associations were not confounded by epitope effects. It was thus hypothesized that missense variants that lead to alterations in protein structure would correspond to both altered protein function and altered proteomic assay readouts (See FIG.15). FIG.15 illustrates that proteomic assay readouts are affected by coding variants even when abundance remains the same. In the top row, unaltered proteins show increased quantification due to higher abundance, while in the bottom row, poor antibody binding causes diminished quantification, ranging from high to slight protein alterations. Structural changes, other protein modifications (e.g., DOCKET NO.: 38120-4362 (11652WO01) PCT post-translation modifications) or technical artifacts (slight alterations to antibody binding sites) may result in lower assay readout. It was further hypothesized that measured alterations from missense variants could be used to supervise LLM predictors leading to improved rare variant classification (See FIG. 16A). To assess this, a proteomic-refined LLM model (See FIGS. 4 and 5) was developed, by way of processor 202, and compared with other strategies for identifying damaging missense variants, including an ensemble of conventional approaches, Combined Annotation Dependent Depletion (CADD), and LLMs (e.g., ESM-1v and ESM-1b, and AlphaMissense) (See FIG. 28). The refined LLM model was assessed in four ways: (1) by assessing correlation of proteomic effect size of genetic variants with predicted deleteriousness scores, (2) by evaluating associations between damaging missense variants (identified by the refined LLM model or other classifiers) and specific traits at 241 gene-trait pairs where an association between loss-of-function and the trait is known; (3) by applying the method of the disclosure to gene-level association analysis of ten example traits from UK Biobank; and (4) comparing scores from the refined LLM model for benign and pathogenic clinical variants from ProteinGym (a benchmark for comparing predicted missense variant effects). Protein assay readouts correlate with missense variant deleteriousness Proteomic assays used the Olink Explore 3072 platform on 46,665 UK Biobank participants of European ancestry who had available exome sequence data. These proteins are distributed between Olink Panel I (n=1,454) and Olink Panel II (n=1,444), which were used for training and validation purposes, respectively, as shown in Table 1 below. Table 1 Gene Symbol Olink Panel Training AGRN Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT PARK7 Panel I TRUE ENO1 Panel I TRUECA6 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT RPA2 Panel II FALSE SMPDL3B Panel II FALSEATP5IF1 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT GBP1 Panel II FALSE GBP2 Panel I FALSEGBP4 Panel I FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT CA14 Panel I TRUE ECM1 Panel II FALSEADAMTSL4 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT FCER1A Panel II FALSE APCS Panel II FALSEFCRL6 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT PLA2G4A Panel I TRUE CFH Panel II FALSECFHR4 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT AGT Panel II FALSE EGLN1 Panel I TRUETSNAX Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT MCFD2 Panel I TRUE EPCAM Panel I TRUECHAC2 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT IL1R1 Panel I TRUE IL1RL2 Panel I TRUEIL1RL1 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT GORASP2 Panel II FALSE METAP1D Panel I TRUEITGA6 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT ATG16L1 Panel II FALSE SAG Panel II FALSECOL6A3 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT DAG1 Panel I TRUE MST1 Panel II FALSESEMA3F Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT RUVBL1 Panel I FALSE DNAJB8 Panel I FALSENUDT16 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT RNF4 Panel II FALSE ADD1 Panel II FALSEHGFAC Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT CXCL3 Panel I TRUE EPGN Panel II FALSEEREG Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT FGA Panel II FALSE TDO2 Panel II FALSECTSO Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT IL6ST Panel I TRUE PDE4D Panel II FALSEHTR1A Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT CTNNA1 Panel II FALSE SIL1 Panel II FALSEMZB1 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT ECI2 Panel II FALSE NRN1 Panel II FALSEBMP6 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT PFDN6 Panel II FALSE RGL2 Panel II FALSESYNGAP1 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT CGA Panel I TRUE GRIK2 Panel II FALSERTN4IP1 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT PSMG3 Panel I TRUE MAD1L1 Panel I TRUECYTH3 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT HGF Panel I TRUE ADAM22 Panel I TRUESTEAP4 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT CLEC2L Panel II FALSE SSBP1 Panel II FALSECLEC5A Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT CLU Panel II FALSE PBK Panel II FALSESCARA5 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT RRM2B Panel I TRUE CTHRC1 Panel II FALSEANGPT1 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT RECK Panel II FALSE CLTA Panel I FALSEGNE Panel I FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT ENG Panel I TRUE AK1 Panel I TRUEPTGES2 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT ECHDC3 Panel II FALSE NUDT5 Panel I TRUECDC123 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT PTEN Panel I FALSE LIPF Panel II FALSEACTA2 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT HRAS Panel II FALSE CDHR5 Panel I TRUESCT Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT AGBL2 Panel II FALSE FOLH1 Panel II FALSEPRG3 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT FOLR2 Panel I TRUE INPPL1 Panel I TRUERELT Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT PHLDB1 Panel II FALSE TREH Panel II FALSEHYOU1 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT CLEC4C Panel I TRUE CLEC4A Panel I TRUECLEC6A Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT ITGA5 Panel I TRUE LACRT Panel II FALSECD63 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT ALDH2 Panel II FALSE ERP29 Panel II FALSERBM19 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT PCDH17 Panel I TRUE PCDH9 Panel II FALSEMZT1 Panel I FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT BMP4 Panel I FALSE LGALS3 Panel I TRUEDLGAP5 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT DLK1 Panel I TRUE DYNC1H1 Panel II FALSECINP Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT IGDCC4 Panel II FALSE MEGF11 Panel II FALSEMAP2K1 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT GFER Panel I TRUE SLC9A3R2 Panel II FALSEPKD1 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT ITGAM Panel I TRUE ITGAX Panel II FALSEAHSP Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT VPS53 Panel I TRUE GLOD4 Panel I TRUESCARF1 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT CCL2 Panel I TRUE CCL7 Panel I TRUECCL11 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT CALCOCO2 Panel II FALSE UBE2Z Panel II FALSEGIP Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT ENPP7 Panel I TRUE CBX2 Panel II FALSESGSH Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT AZU1 Panel I TRUE PRTN3 Panel I TRUECFD Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT FDX2 Panel II FALSE ICAM3 Panel I TRUECDC37 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT SBSN Panel II FALSE COX6B1 Panel II FALSEHSPB6 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT ERCC1 Panel II FALSE FOSB Panel I FALSEVASP Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT VSTM1 Panel I FALSE TARM1 Panel II FALSEOSCAR Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT CST5 Panel I TRUE CST7 Panel I TRUEENTPD6 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT USP25 Panel II FALSE CXADR Panel I TRUETMPRSS15 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT SEZ6L Panel I TRUE CRYBB1 Panel II FALSEMN1 Panel II FALSE DOCKET NO.: 38120-4362 (11652WO01) PCT OFD1 Panel II FALSE VEGFD Panel I TRUEACE2 Panel I TRUE DOCKET NO.: 38120-4362 (11652WO01) PCT AMOT Panel II FALSE IL13RA2 Panel II FALSEIL13RA1 Panel I TRUE For the 2,898 proteins measured, 560,295 coding variants were observed among assayed individuals. In an analysis of cis-coding variants, 11,682 variants associated with proteomic assay results (p < 1x10-8, see Example 2) were found across 1,925 proteins (1,168 Panel I, 756 Panel II). The training and validation sets included only proteins that had at least one cis-coding variant significantly associated with assay readouts. Evidence was explored for the hypothesis that proteomic assay results were correlated with variant deleteriousness. FIG.25A shows that when coding variants were grouped into three categories (synonymous, missense, and putative loss of function), the estimated impact of each variant on assay readout correlates strongly with the variant type (as expected). Briefly, synonymous variants had no impact on assay readouts (median effect DOCKET NO.: 38120-4362 (11652WO01) PCT size β = -0.01 +- 0.81 standard deviation units) whereas pLoF variants had the largest magnitude impact on protein levels (median β = -1.40 +- 1.35 standard deviation units), corresponding to an estimated reduction of 40% in estimated binding (on raw scale). The impact of missense variants was intermediate between these two (median β = -0.20 +- 1.04 standard deviation units). FIG.25B shows that when missense variants (regardless of cis- pQTL significance) were grouped according to deleteriousness as quantified by an ensemble of classifiers (number of deleterious predictions from PolyPhen, HVAR / HDIV, MutationTaster, SIFT and LRT), the smallest effects on protein assay results are seen for variants predicted to be benign (median β = -0.11 +- 0.92 standard deviation units) while the largest effects were seen for those predicted to be deleterious (median β = -0.5 +- 1.23 standard deviation units). FIG.25C further shows that proteomic cis-pQTL effect sizes for ClinVar benign variants (median β = -0.09 ± -0.17 SD) and ClinVar pathogenic variants (median β = -1.22 ± - 1.16 SD) are also significantly different (p-value = 6.72x10-48) effect size. ClinVar pathogenic and benign classifications are used to distinguish likely disease-causing variants from other variants within human disease genes. Not only were significantly different associations found with proteomic assay readouts between ClinVar pathogenic and benign variants, but it was also found that proteomic assay results could be used to classify pathogenic and benign variants significantly better (AUROC = 0.76) than a random classifier (AUROC = 0.50) or a classifier using minor allele frequency alone (AUROC= 0.65), consistent with results from a previous study. These results suggest that effect size from pQTL analysis is a more effective classifier of pathogenic and benign variants than minor allele frequency alone, as evidenced by a higher AUROC score. Together, the observations in FIGS.25A-C support the hypothesis that proteomic assays capture not only protein abundance but also provide a quantifiable measure of missense variants deleteriousness. Most likely, this is because these proteomic assay readouts are sensitive to changes in protein conformation and deleterious variants will often alter protein tertiary structure. Construction of proteomic-informed LLM utilizing summary statistics from protein quantitative loci analysis DOCKET NO.: 38120-4362 (11652WO01) PCT The next goal was to use proteomic assays results to improve LLM classifiers of variant deleteriousness. This was conducted in two steps. First, a classifier that used proteomic assay results to classify missense variants as “pLoF like” (generally, with larger impact on assay readouts and smaller allele frequency) or “synonymous like” (generally, with little impact on assay readouts and relatively higher allele frequencies) was built. In a subsequent step, the output of this classifier was used to refine LLM predictors of deleterious variants. Thus, in the first step, each variant was annotated, by way of processor 202, with allele frequency, proteomics effect size, standard error and p-value (See FIGS.8A-8D). An ensemble classifier (See FIG.4, Example 2) was built, by way of processor 202, that used these four features to distinguish benign variants (synonymous variants in Panel I proteins as negative examples) from deleterious variants (pLoF variants in Panel I proteins as positive examples). Once trained, the classifier allowed for scoring of all 118,647 observed missense variants with allele frequency < 1% observed in Panel I proteins. The output of this classifier was better at distinguishing ClinVar pathogenic and benign variants (AUROC 0.83 in Panel I proteins, 0.79 in Panel II) then classifiers using only effect sizes (AUROC 0.78 in Panel I, 0.76 in Panel II) and those using only allele frequencies (AUROC 0.72 in Panel I, 0.65 in Panel II). (See Table 2 below). Table 2 Possible Proteomic Labels for AUROC AUPRC Refining ESM-1b In the second step, the ESM-1b proteomics LLM was fine-tuned using results of the classifier for 118,647 missense variants (ESM-1b+proteomics (ESM-1b+p) method). The DOCKET NO.: 38120-4362 (11652WO01) PCT training variants were used as continuous supervised labels, where a convolutional neural network head was added to ESM-1b to pool hidden features from its final transformer layer. A small sample of 2,000 random ClinVar pathogenic and benign variants was used to determine the optimal point for the model to stop training, thereby reducing overfitting to the Olink proteomic platform-specific effects. The optimal point was determined by comparing Spearman correlation of pathogenic and benign variants with the predicted deleteriousness. The models achieved a Spearman correlation above 0.70 during training compared to 0.67 for standard ESM-1b and 0.62 for ESM-1v on these 2,000 variants. The ESM-1b+p model allowed for generation of proteomic informed missense deleteriousness scores for 73,490,006 potential missense variants across 19,596 proteins in the human genome. Missense variant predictions from proteomics-guided LLM are more correlated with validation proteomic assay readouts than all other tested approaches Several experiments to evaluate the performance of the refined LLM model, compared to other approaches using the 756 Panel II validation proteins were performed. Panel II proteomic assay readouts for 99,730 missense variants not included in training were compared with the ESM-1b+p score and with several other methods for predicting variant deleteriousness. In each evaluation, predictions for each method were grouped into three categories (Likely Benign, Default, Likely Deleterious) matching the category sizes to those from the 5-method conventional ensemble (FIG.25B). The LLM was compared with five other approaches: ESM-1b, ESM-1v, AlphaMissense, CADD and the 5- method conventional ensemble (the ensemble scores the number of deleterious predictions from PolyPhen, HVAR / HDIV, MutationTaster, SIFT, and LRT in a scale of 0 – 5 and assigns variants with a score of zero as Likely Benign and those with a score of 5 as Likely Deleterious). FIG.26A shows that likely deleterious variants based on the protein-guided LLM had the highest impact on protein assay results among the approaches compared. In principle, a method for predicting variant deleteriousness might work well by distinguishing deleterious variants from benign variants within a gene, or by distinguishing genes that are tolerant of variation and mostly contain benign variants from those that are not and might be enriched for deleterious variants. To explore the value of the within gene DOCKET NO.: 38120-4362 (11652WO01) PCT predictions, the correlation between ESM1b+p scores and proteomic assay readouts within the 756 Panel II genes was evaluated. FIG.26B shows a box plot of the resulting within gene correlation coefficients, where larger magnitude negative values are better. For this metric, the proteomics-guided LLM outperformed the other methods with a median Spearman correlation of -0.199, followed by ESM-1b with -0.180 and AlphaMissense with -0.172. Among proteins with a negative correlation for either ESM-1b+proteomics or AlphaMissense, ESM-1b+proteomics had a stronger correlation 57.4% of the time (p=7.4x10-5, FIG.26C). For standard ESM-1b, ESM-1b+proteomics had a stronger correlation 68.6% of the time (p=6.6x10-22, FIG.26D). Finally, it was examined whether the within-protein Spearman correlation extended across all missense variants and validation proteins. The overall Spearman correlation between effect size and predicted scores across all missense variants and validation proteins with scores for all methods were compared in Table 3, shown below. Table 3 Spearman correlation of Score Spearman correlation of Score and Effect Size and Absolute Effect Size ESM-1b+p had a Spearman correlation of -0.212 compared to -0.16 for ESM-1v, -0.141 for standard ESM-1b, and -0.137 for AlphaMissense. Proteomics-guided LLM recapitulates the most pLoF-defined gold standard gene-trait pairs using only missense variants in burden tests While the results showed that the proteomic-refined model outperformed other methods when predicting proteomic assay readouts, the primary goal was to determine whether the proteomic-guided model could improve power for genetic association studies. To evaluate this, 241 positive control autosomal gene-trait pairs, as listed in Table 4 below, were selected where pLoF burden tests had previously established associations with a health-related trait. These 241 genes did not overlap Olink Panel I proteins used for training. DOCKET NO.: 38120-4362 (11652WO01) PCT Table 4: 241 Positive Control Gene-Trait Pairs Gene TraitSLC4A1 Reticulocyte percentage ABCB4 Alanine aminotransferase DOCKET NO.: 38120-4362 (11652WO01) PCT SEC23B Red blood cell erythrocyte distribution width CST9L Cystatin CLIPG Apolipoprotein A DOCKET NO.: 38120-4362 (11652WO01) PCT HNF1A Cross-cohort ICD10-based definition CD36 Apolipoprotein AMIP Other cataract f DOCKET NO.: 38120-4362 (11652WO01) PCT ALPL Alkaline phosphatase TRIM58 Mean platelet thrombocyte volumeDDR2 Standing height DOCKET NO.: 38120-4362 (11652WO01) PCT IFT140 Cyst of kidney, acquired FANCA Natural hair color before greyingINSC Vitamin D It was reasoned that it should be possible to recapitulate some of these pLoF burden signals using deleterious missense variants and that improved classification of missense variants would increase the proportion of signals that could be recapitulated. Thus, singleton missense variants in the full UK Biobank cohort were classified with the ESM-1b+p classifier and a variety of alternative variant classification strategies and recapitulation of each of the 241 DOCKET NO.: 38120-4362 (11652WO01) PCT association signals was attempted. This benchmark is quite challenging since even a perfect classifier may not be able to reconstruct a pLoF association using missense variants alone. In the context of genetic association studies, the optimal balance between a small number of very deleterious variants and a larger number of more modestly deleterious variants is not obvious. To facilitate this balancing act, variants were ranked from most deleterious to least deleterious using each method and then a method specific grid search was used to find the optimal method-specific split of likely deleterious variants to include ina genetic association burden test. When using singleton missense variants, it was found thatESM-1b+proteomics recapitulated the 88 gene-trait associations with p<0.01 (FIG.27A) This outperformed all alternatives considered, including using all singleton missense variants (63), using the 5 method ensemble (69), using ESM-1v (73), using the original ESM- 1b (83), and using AlphaMissense (87). The relative advantage of ESM-1b+proteomics over other methods increased with more stringent p-value thresholds (See FIGS. 29A and 29B). FIGS. 29A and 29B plot the number of significant burden associations across 241 gold standard genes (using different p-value thresholds) using different allele frequency bins, where each bin includes all variants up to the specified maximum frequency. Scores were ranked from lowest to highest, and all thresholds were tested with a step size of 0.02. The threshold per method with the highest Winsorized mean -log10(p-value) using singleton missense variants was selected. The same threshold was used for all analyses in these figures. Using p-value < 2.1x10-4and singleton missense variants, ESM-1b+proteomics recapitulated 56 gene-trait pairing compared to 51 for AlphaMissense and 50 for standard ESM-1b. FIG.27B shows that the method of the disclosure increased not only the number of significant gene-trait pairs but also increased the average association signal strength relative to the other methods (measured by the mean -log10(p-value) after winsorization to reduce the impact of outliers). Specifically, ESM-1b+p increased mean -log10(p-value) by 7.5% relative to standard ESM-1b and 3.8% relative to AlphaMissense (FIG. 27B). Using median - log10(p-value) shows a similar trend as Winsorized mean -log10(p-value) (See FIG.30). As a sensitivity analysis, the evaluation was expanded to higher frequency missense variants and found similar patterns as in the analysis restricted to singletons (FIGS.27A, 27B). Relative to AlphaMissense, ESM-1b+p improved 32 associations by >1 order of magnitude in p-value DOCKET NO.: 38120-4362 (11652WO01) PCT compared to 17 improved associations for AlphaMissense, the next best method (FIG.27C). Similarly, the model notably boosted 26 associations compared to 9 for standard ESM-1b (p=value = 0.003)(FIG.27D). Finally, it was investigated whether the proteomic-guided LLM increased gene-level effect size estimates in burden testing. It was hypothesized that a larger effect size would indicate more deleterious variants. To test this, the mean absolute effect size across 241 gene-traits for each deleteriousness threshold was computed. The mean absolute effect size was plotted against the Winsorized mean -log10(p-value) and the number of significantassociations (See FIG. 31). It was found that ESM-1b+proteomics exhibited noticeably largereffect sizes and mean -log10(p-values) across most deleteriousness thresholds (FIG.27E). Proteomics-guided LLM maximizes yield of associations for ten example UKB traits The proteomics-guided model was next applied to ten example traits from the UK Biobank (birth weight, body mass index, creatinine, cystatin C, HDL cholesterol, hearing loss, standing height, LDL cholesterol, mean reticulocyte volume, and triglycerides) to contrast association yield with the other approaches for selecting missense variants to include in gene burden tests. To do this evaluation, missense and pLoF variants were combined in groups according to allele frequency (ranging from singletons, 1x10-5, 1x10-4, 1x10-3to 0.01) and deleteriousness (pLoF, pLoF + deleterious missense, pLoF + default deleterious, pLoF + all missense). As a baseline, deleterious missense variants were defined by all 5 methods in the conventional ensemble predicting the variant to be deleterious, whereas default deleterious requires at least one deleterious prediction. For each LLM approach the top 22% scoring missense variants as deleterious (over all possible single nucleotide missense variants) were defined and combined pLoF and deleterious missense burden tests were computed, comparing to the conventional baseline. This percentage was chosen to match the proportion of deleterious missense variants selected by the 5 method ensemble approach. When using variants with allele frequency < 0.1, and a significance threshold of 5x10-9, the 5 method ensemble resulted in 142 significant associations compared to 153, 164, 170 and 177 for ESM-1b, ESM-1v, AlphaMissense and ESM-1b+proteomics respectively (FIG.32A). This translates to a 15.7% increase in yield for ESM-1b+proteomics over standard ESM-1b, and 24.6% increase in yield compared to the 5 method ensemble. The improvement of ESM- DOCKET NO.: 38120-4362 (11652WO01) PCT 1b+proteomics was consistent across allele frequencies and p-value thresholds (FIG. 33). Across these traits, inflation of test statistics was not observed, as shown in Table 5 below. Table 5: ESM-1b+proteomics does not exhibit genomic inflation Trait ESM-1b+p ESM-1b ESM-1v AlphaMissense 1+ / 5 5 / 5 method method BMI 1144 1116 1133 115 1129 1122 Next, novel association signals found by the LLM relative to the other methods were examined. (Table 6 below, and FIGS.34A-C). One of these is highlighted in FIG.32B. FIG.32B shows PCSK6 with triglyceride-levels association. This example shows summary statistics and p-values for burden codings included in the gene-based testing framework (AF < 0.1%). The associations involve pLoF and missense variants with an allele frequency of less than 0.1%. Missense variants were included if they were among the top 22% most deleterious, as determined by AlphaMissense, ESM-1v, ESM-1b, and ESM-1b+proteomics. Table 6: Four associations using variants with AF < 0.1% not found with any other approach

[0002] ?in av&p 0 1 9 0 9 0 9 Fbo1 -LME- -0-7E01E1E12100pSE . . .21.2 &b 8 8 7 7 F 1 0-E0-E0-E0-oLM 2 1 3E1 pSE .02.42.32.52 &8Fv102- 05- 07- 0-oLMEpS0EE7 2E.31 2E.36 6 .16.2-1)e111sn & 5 6 7 -0es0 0 0 40 OFsoi-E- - -L2Ma6Eh07E E5 2 0Wp. .7.3.l5 6 2 256pA 11(2&es4 7 n)06F- 0 6 1 - 0- 0-3o4Les5 / E9E6E1E9 - psi5(m2.71.71.89.10218&es8 4 5 53Fn)50- 0- 0- 0-:.oLes / + E4E1E2E7 psi1(7.34.4.9.O m8 5 7Ny T nos 8 2 5 4E llnu0-0E-0E-0E-K Ayso n m72635E39C o.n 3.1.71.1 O D F407oL-02010E-E-E-Ep 3 8.5 1 9 25.79.80.4 se lst d or etyed iair L etrsn ac tlnirTecly DL leeouuecgio rTh Mcitoc lyCer igrTe GR1P2K 6 neA B HKSG P C P R RPSP DOCKET NO.: 38120-4362 (11652WO01) PCT The novel association (Effect =-0.103 SD, p-value=2.1x10-9) was for proprotein convertase subtilisin / kexin type 6 (PCSK6) with triglycerides. PCSK6 has common variation associated with triglyceride-levels in the GWAS Catalog and knockout of PCSK6 is associated with salt- sensitive hypertension in mice. The second best approach for deleterious missense variant prediction in PCSK6 with triglycerides was standard ESM-1b (Effect=-0.100 SD, p- value=2.5x10-7). Finally, associations that were novel to the LLM approaches relative to the conventional approaches were explored (Table 7, below). One of these findings is anassociation between hearing loss and rare deleterious coding variants in SIX1, a gene thatplays a critical role in inner ear development during embryogenesis (FIG.32C). This gene has also been connected to hearing loss via de novo mutations. When considering variants with allele frequency < 0.1% in SIX1 in a burden coding, the association using nonsynonymous variants was not significant (Odds ratio = 1.216, p-value=0.003). Using pLoF and missense variants predicted deleterious by 5 algorithms was also not significant (Odds ratio = 2.229, p- value=0.006). All the ESM LLM approaches were significant with the strongest being the ESM- 1b+proteomics approach (Odds ratio = 2.803, p-value=3.6x10-11). Table 7: Novel LLM Hits

[0003] 5esn e 8 8 2 77 9 4 7 5 8 6 3 8 9 2 70 s 6 0- 2 5 68 0- 0- 0 0 57 0 38 6 si74E40 21 9E E-E-E2 -E3 7 7 1 7 7 8 4 0 3 0 5 m7.5 2.010 0 0 0.00 0. .6.2.4.8.0 1(26a3hes1 2 8 0 4 1 4 3 3 6 2 0 2 9 0 8 4-plAnh -E -E -E1-E1-E1-E0-E1-E1-E1-E0-E1-E0-E 02 Fessi igh 9 21.2 3.6 8.9 6.5 6.9 2.4 3.7 6 6 4 1 0 1 1 1 7 3 0 8oLp M 2 7 5 6 1 2 2.3.1.6.1.4.1 3:.1 9 6 OF8 0- 8 6 7 8 8 1 7 9 1 5 2 490- 9 7 00- 00-0- 3 4 NoL1 E 8 2 1 1 1 E 7 E 07 E E 71 45 T p2.0 0 7 01.1 7 9 0 5 2 3 8 00 1 E 20.0.20.0.10. .3.70.0.0 0 0 0 0 K C O 1 1 1 1 1 1 1 1 1 1 1 1 1 DF0 . 0 0 0 0 0 0 0 0 0 0 0 0 A000.00.00.00.00.00.00.00.00.00.00.00.0 l oletetsst rega intCngint gint gintore nyc e gintnyc e g t g t loiLtrsdhgitdhgdhgdhg Ltsalomdhgalomindhg indhg gTD Hleonaihteatn ien ien ieDleeulu n ie eulu n ien ienirCShsy aCtShatShatSh Ho Mchitov ath Mcitov athathaCer Ser S S eH e 2 1 2 2 T neFSA 2 1LFS9P A 1PLB3 L J3 P G B B 2 A2FRB1 GRSO C CG2PT FXICRSD M HTLES LE O PGT SC 8 3 2 8 7 88 7 7 8 7 7 1 8 2 5 16 7 7 764 6 5 0 0 2 0 0 0 0 0 6 0 9 0 0 0070 0 630 4-1- 9 E0- 1 1- 5 0- 0 1- 0 1- 0 1- 0 1- 0 1- 9 0- 7 8 0- 7 0- 9 0- 6 0- 9 0- 8 0- 8 902 1 E E E E E E E E E 1 - 0 3 E E E E E E E 038 0 2 4 6 0 8 5 2 0 0 0 8 5 40 2 5 0 722 4 1.86.1.8.5.1.3.56.58.59.30.0.3.0.5.2.3.5.0 3 8 1 61 3 4 3:.3 4 7 3 26 9 3 5 7 3 7 7 2 8 5 7 48 6 O 2 N 380- 0 2 30 60 061 0 10 0 07 046 32 T 1 E 3 8 5 1 - 0 6 4 E -E 4 3 2 -E -E -E -E 1 0 9 0 20 1 23 9 1 0 91 01 7 4 1 5303 0 E0.0.0 0 0 20 0 0.0.9.43.0 40.0.0.2.22.30.25.01 0 70 0.0.K 0.0.0 0 0 0 0 0.00 0 C O 10 10 10 10 10 10 10 10 10 10 10 10 10 1 1 1 1 1 D0.00.00.00.00.00.00.00.00.00.00.00.00.0 00.0 00.0 00.0 00.0 00.0 etnyetc gstedgsst g g t g t g t log g t g g ttcCeetye ace elomirind dhgir ini ecdhg int i dhg inidhg inidhg inidhgig int n dhg inidhg int i dhg inidhe girn dit ininnctae lom Muc luLna e ly na ena ena ena ena eirna ena ena enaie Lat auc luitoevrDLtShig trShtShtShtShtShaTetShtShtShtSh HDLsyCerMiCtoevr6 1 1 G 3S2 2A 1 2 1 A M B X O R CP NT2FR A A P R R D 1 H C5SR P5T4 G R O RTSMSP K M P N A HOL1LM W C O B P R N RPM D U P OC OSRENEA CSWT 1 60 7 6 5 1 6 1 9 1 34 0 0 61 0 27 0 2 6.6 16.4 21.7 22.0 16.7 10.0 32.3 4 25.48.1 9 0 9 -E08 -08 0 7 0 6 0 8 0 8 0 6 0 1 E -E -E -E -E -E -E -E . -00 21.9 25.0 82.5 31.5 60.3 16.0 11.6 59.5 7 1 ) 1 1 -0O 70 70 50 80 7 7 9 8 7 W2-E-53E- 1E- 3E0-E0-E0-E0-E0-E63.25.21.4 12.4 9.0 5.4 7.9 2.9 9.1 1 1 6 3 3 2 1(2637 8 9 9 9 3 6 6 9 4-0- 92 E 20-0-0- 780-0-02 0 2 E4 E8 E 109 E E -E 0271.01.1.6 4.0 0 2.6 7 9 5 8 20.0 2 2 20. . .0 1 4 4 3:.O 8 6 0 9 5 7 7 9 5 1 6 7 5 6 8 7 N 109 9 -E0-E 2 00-E0-E 6 8 0 4 3 4 9 2 0 6 T4.00 000 011 E0.0.1.10.0. .11.90.0 0 0 K C O 10 10 10 1 1 1 1 1 1 D0.00.00.0 00.0 00.0 00.0 00.0 00.0 00.0 etsnye ed e ientyg t gtg t g t g tace lomirlu ec inn tiaac ie lo ndhtovng inidhng inidhng inidhng inidhgiMuc lyierMucit atehatehatehatehnateh ergrTCer S S S S S2K 6 1 2 2 5 HKSLI1 IPR X A 1 GKNLBLPSC PVSRT LB N R A O C PBFO P DOCKET NO.: 38120-4362 (11652WO01) PCT Predictions from proteomics-guided LLM improve the classification of ClinVar pathogenic variants over unguided baseline LLM As a final assessment of the value of proteomics-refinement over standard ESM-1b, its predictions on ClinVar variants from the ProteinGym benchmark were investigated and the results shown in Table 8 below. To ensure fair evaluation, ClinVar variants used during training for either model fitting or convergence checking were removed. To ensure balance of pathogenic and benign variants per protein, proteins with at least 4 benign and 4 pathogenic variants were selected. This filter resulted in a total of 34,531 variants and 760 genes. ESM-1b+proteomics outperformed standard ESM-1b in both the AUROC and AUPRC for all variants, achieving 0.940 versus 0.919 for AUROC and 0.955 versus 0.936 for area under the precision recall curve (AUPRC), respectively. AlphaMissense performed the best on this metric, achieving an AUROC of 0.947 and an AUPRC of 0.958. A similar pattern also occurred when assessing mean and median AUROC per gene as shown in Table 8 below. Table 8: ESM-1b+proteomics improves performance over ESM-1b on ProteinGym ClinVar variants Model AUROC all variants AUPRC all variants Mean AUROC per Median AUROC gene per gene Proteomic refinement improves performance on a variety of ESM models with differing numbers of parameters showing the robustness of the proteomic training labels The analyses so far shows that ESM-1b+p outperforms the original ESM-1b model at the task of identifying deleterious missense variants for genetic association studies and, also, that ESM-1b+p outperforms all other approaches considered for this task. It was also evaluated whether proteomics data could be used to improve other LLM predictors of variant deleteriousness. For this, a model utilizing ESM-1v model embeddings as input features was trained (See FIGS.9 and 10) and again found that performance improved over standard ESM- DOCKET NO.: 38120-4362 (11652WO01) PCT 1v when classifying missense singleton variants for a genetic association study (83 versus 73 genes recapitulated), although the refined model performance was lower than for ESM- 1b+proteomics (83 versus 88 genes recapitulated) as shown in FIGS. 35A and 35B. Proteomics-refinement was applied to various ESM-2 models with sizes ranging from 8 million parameters to 150 million parameters using 2,000 ClinVar variants (early stopping set). Performance of these models improved with proteomic refinement, but models with fewer parameters improved notably more than models with more parameters as shown in FIG.36. For example, proteomic-refinement improved ESM-2’s AUROC (with 8 million parameters), from 0.653 to 0.841 exceeding the performance of the 150 million parameter version ESM-2. The 150 million parameter model’s AUROC improved from 0.807 to 0.888 exceeding the performance of the 3 billion parameter model which had an AUROC of 0.878. Here, four important advances for genetic association studies of rare coding variation are highlighted. First, the value of human proteomics data was confirmed for comparing various algorithms for damaging missense prediction using summary statistics alone. Second, it was shown that previously described LLMs trained on protein sequences can improve the power of these studies relative to ensembles of conventional models for variant classification. Third, it was shown that when these models were trained on the results of proteomic association studies using UK Biobank data, their performance further improves. Fourth, the results suggest that there is orthogonality between the proteomics-guided LLM, AlphaMissense and standard ESM-1b suggesting omnibus tests combining all of these may lead to the most genetic associations. All these show the value of large-scale human proteomic data for empowering genomic discovery. These advantages were evident in the analysis of 241 pLoF-defined positive control gene-trait pairs both in terms of ability to recapitulate the pLoF defined association signals and in the improvement of burden association yield of ten UKB traits. Furthermore, the LLM could be leveraged in omnibus tests (including SKAT tests) potentially combining other missense variant classification strategies. In principle, protein sequence LLMs could also be refined with data from sources such as ClinVar, HGMD, and Deep Mutational Scans. However, large-scale human proteomic data is appealing relative to these alternatives because, when combined with DNA sequence DOCKET NO.: 38120-4362 (11652WO01) PCT data, it annotates large numbers of variants and is expected to be relatively free of bias. Evidence was also provided that showed that the proteomic labels are not largely driven by epitope effects given their classification capabilities on ClinVar pathogenic missense variants and correlation with existing deleterious scores, consistent with previous studies. In the future, an abundance of deeply annotated datasets that combine DNA sequence information, proteomics and other omics might provide even broader training sets that can be used to refine LLMs, either individually or in combination with sources such as ClinVar or HGMD. Training labels and validation data are provided. Furthermore, precomputed classification (deleterious, default, likely benign) for all missense variants in canonical transcripts from Ensembl 100 and annotation files for REGENIE are provided. Example 2: More detailed methods Protein quantitative trail loci (pQTL) and function annotation Data from UK Biobank Pharma Proteomics Project for pQTL analysis and subsequent model training, by way of processor 202, was used. Starting with the normalized protein expression (NPX) data from the consortium, principal component analysis was performed, and the NPX data was adjusted by the top 100 principal components. This was done to adjust for confounders and concomitantly increase power for cis-pQTL detection. pQTL analysis was then run using REGENIE for variants that were captured by UKBiobank exome sequencing efforts in the 46,665 individuals with blood plasma proteomicsdata as well as TOPMED variants. pQTL regression analysis was run adjusting for age, sex, age x sex, age, and four genetic ancestry principal components as covariates. VEP (variant effect predictor) was used to annotate variants as synonymous, missense, or loss-of-function (stop_gained, essential splice site variants and frameshift indels), using ENSEMBL Release 100 transcript annotations. Canonical transcripts were defined similar to other approaches. Specifically, transcripts with "MANE Select" annotations were selected as the canonical transcript set, the current standard for representing biologically relevant transcripts. In the absence of MANE annotation, APPRIS was selected and "Ensembl canonical" annotation tags to define canonical transcripts. APPRIS annotation was prioritized over “Ensembl canonical” because it uses cross-species conservation as a key feature to identify the likely functional DOCKET NO.: 38120-4362 (11652WO01) PCT transcript. Putative loss-of-function (pLoF) variants were defined as those annotated as frameshift, stop_gained, splice donor, and splice acceptor. Positive control gene-trait pairs A list of gene-trait pairs with an established pLoF burden signal was defined as the positive control set of genes. The inclusion criteria for these gene-trait pairs were that there was a burden signal (p < 2.18x10-11) driven by pLoF variants with allele frequency < 0.01. Since a gene can be associated with many traits, it was desired to reduce the chance a single gene would drive the conclusions. To this end, each gene was selected at most one time, focusing on the most significant trait per gene. This resulted in 276 gene-trait pairs. To further reduce bias in the conclusions from proteomic training, pairs that overlapped the Panel I Olink proteins used for training were excluded, and autosomal genes were focused on, resulting in 242 gene-trait pairs. One gene LARP1 (ENSG00000155506) was neither 5’ nor 3’ complete using the canonical transcript defined above so it was removed leading to 241 gene-traits. Proteomic training labels derived from pLoFs and synonymous variants To refine the LLMs, a target label was needed to perform supervised training for every missense variant included in the learning panel. Several options were considered for creating proteomics training labels, including the effect size with inverse variance weighting, but found that combining proteomic summary statistics (effect size, allele frequency, p-value and standard error) into a single missense label improved the classification of ClinVar pathogenic variants. Specifically, a model was created (trained on synonymous and pLoF variants), by way of processor 202, that generates a continuous score between zero and one indicating the degree to which a missense variant is acting like a pLoF or synonymous variant. Since some proteins may have imprecise antibody binding or a large percentage of its readout below the limit of detection, training proteins were limited to those with a cis-pQTL (p < 3.4x10-11), where cis was defined as being within 1 mb of the start or the end of the gene. It was hypothesized that if there is a cis-pQTL for a protein, then the protein is likely to have a reasonably specific antibody binding. This resulted in 1,175 out of 1,454 genes (from Panel I proteins) having at least one associated cis-pQTL and a missing rate across samples less than 10% were used for training. From this group of proteins, synonymous and pLoF variants were DOCKET NO.: 38120-4362 (11652WO01) PCT selected and used as binary labels (0 and 1, respectively) for training in step one of the two step machine learning approach shown in FIG.4. An ensemble of two models were created (which were averaged to get a continuous score per missense variant), by way of processor 202. A neural network trained, by way of processor 202, with pytorch (version 1.13) and pytorch-lightning (version 1.8.1), and a XGBoost (version 1.6.2) model trained using scikit-learn (version 1.02) was used. The structure of the neural network was a standard three layer fully connected network with a hidden size of 16 and GELU nonlinear activations. Synonymous and pLoF variants were used as labeled training data and missense variants were used for prediction. The features used for training consisted of log10 p-value, minor allele frequency, effect size, and standard error. The models were trained, by way of processor 202, using all chromosomes except for chromosome 20, which was used to tune hyperparameters optimizing AUROC. To account for the class imbalance (since there are many more synonymous than pLoF variants), the samples were weighted, by way of processor 202, so they would have an equal total weight across classes. The AUROC of the XGBoost classifier was 0.88 for distinguishing between pLoF and synonymous variants on chromosome 20. The feature importance was (using Shapley values) 0.673 for effect size, 0.149 for p-value, 0.095 for minor allele frequency, and 0.083 for standard error. The neural network classifier with the same features had an AUROC of 0.86 on chromosome 20. Averaging the predictions of these models slightly improved AUROC on overlapping ClinVar variants (See Table 2).The predictions generated, by way of processor 202, from this ensemble model were then used as continuous training labels for refining the ESM models. These labels are called y_assay. All AUROC p-value calculations were done using DeLong’s method. The p-values from this approach were close to that form 2,000 bootstraps. Refining ESM-1b model with proteomic derived labels The transformers library from Hugging Face was used to import ESM models. A convolutional neural network head was added to ESM-1b. The input to the modified ESM- model was three amino acid sequences: (1) the reference amino acid sequence, (2) the amino acid sequence with at least one altered amino acid, and (3) the amino acid sequence masked DOCKET NO.: 38120-4362 (11652WO01) PCT out. On the forward pass the ESM-1b model was applied to each of these inputs. The last hidden layer was extracted and concatenated into separate channels. A 1x1 convolution kernel was applied to merge the channels into a single matrix. After this, seven kernels of size seven x hidden size was applied to the resulting matrix with adaptive pooling. GELU nonlinear activations were used, and a drop out operator after this layer was included. This output was passed into a final fully connected layer. A skip connection from the final hidden layer (selecting out the embedding at the amino acid position for the variant) was also added to the fully connected layer. To reduce dimensionality, a 1x1 convolution was applied to merge across channels before passing into the fully connected layer. Next, the log probability difference between the alternative amino acid and reference amino acid at the amino acid sequence position was added and this was put in the final fully connected layer. To adjust for protein length, the log10 of the amino acid sequence length was also added. The length adjustment covariate did not significantly alter the results. To update model weights, binary cross entropy between yassay and the predicted value from the forward pass called passay was used. log(pREF) from the masked sequence prediction was also added, where pREFis the likelihood of the reference amino acid. These were weighted by a constant ^^ which was set to 0.01. ^^^^^^^^(^^^^^^^^^^^^ ,^^^^^^^^^^^^ , ^^^^^^^^ , ^^^^^^^^) was used, and for training the pytorch-lightning was used.2,000 random benign / likely benignand pathogenic / likely pathogenic variants from ClinVar were used for early stopping to prevent overfitting. Spearman correlation of the passay and the binary labels were checked every 20% of an epoch. Once performance stopped improving for one entire epoch, training was stopped and the best model was taken. To speed up performance, a maximum context size of 768 was used. A noticeable decrease in performance on the 2,000 ClinVar variants was not observed using this context size. Augmenting ESM-1v embeddings using supervised learning with proteomics derived labels DOCKET NO.: 38120-4362 (11652WO01) PCT The missense scores from the proteomic data were limited to the observed variants and genes, but could be expanded to the whole exome by utilizing these missense scores and refining LLMs. The refined LLM can then be applied, by way of processor 202, to all possible missense variants. First, features from the ESM-1v were extracted, by way of processor 202, according to the scheme shown in FIG.9. The approach aimed to capture information about gene context and amino acid change for a variant of interest. A 1281x4 matrix was designed, by way of processor 202, per missense variant that captured vectors related to context and sequence change. Specifically, embedding vectors were extracted, by way of processor 202, from the final transformer layer of ESM-1v to develop the 1281x4 matrix from a) the average (length 1280) vector of all sequence position vectors spanning a gene; b) the (length 1280) vector at a sequence position being considered (which could vary due to a genetic variant); c) another version of the gene context (length 1280) vector, this time including the sequence variant, another (length 1280) sequence vector, but this time for the variant sequence (as opposed to the reference). A row was also added, by way of processor 202, to this feature matrix (which is how the column vectors go from length 1280 to length 1281) with ESM-1v scores for the variant change in columns 1 and 3 and zeros in columns 2 and 4. The output layer depicted in the schematic in FIG. 9 is for illustrative purposes (i.e., to clarify how the original ESM-1v score was calculated), and was only used in the feature matrix in the final appended row just described (i.e., the ESM-1v score per variant). As depicted, features were concatenated, by way of processor 202, as columns after sequentially inputting the reference and altered amino acid sequences. Finally, the missense scores predicted by the step 1 ensemble model (as continuous training labels on the probability scale) were used for training / tuning, by way of processor 202, this ESM-1v derived feature matrix. Training was done, by way of processor 202, with a convolutional neural network (CNN) implemented in pytorch (version 1.13) and pytorch-lightning (version 1.8.1), and using the architecture depicted in FIG. 10. A CNN was used to take advantage of the input structure that was extracted from ESM-1v, as each row corresponds to the same position from the transformer or output layer from ESM-1v. By leveraging a CNN, each row of this matrix could have a kernel applied to it, thus dramatically reducing the parameter space as compared to a standard feed forward network. Layer normalization and GELU nonlinear activations were used and a drop DOCKET NO.: 38120-4362 (11652WO01) PCT out operator was included after the CNN layer. Finally, L2 normalization was used for weight regularization and binary cross entropy with continuous labeling used as loss function. When training this model, by way of processor 202, 22 of 23 chromosomes were used while chromosome 20 was held out of training and used to determine early stopping criteria (i.e. it was used as the “tuning” chromosome). Training progression was evaluated, by way of processor 202, using the Spearman correlation statistic, and if this did not increase for 10 epochs, training was stopped and the best model for subsequent prediction was selected. The developers of ESM-1v trained five models and the final ESM-1v prediction was an average of the 5 models. A similar approach was taken in this Example, where a separate CNN per ESM-1v model was trained, by way of processor 202, and the average of the 5 trained models taken to make the final prediction for a variant. For genes longer than 1022 amino acids (the maximum context size of ESM-1v), the context was centered at the variant position. Predictions were then created, by way of processor 202, for all possible missense variants for canonical transcripts in Ensembl 100. Burden analysis using positive control gene-trait pairs REGENIE whole genome regression framework was used for gene burden association testing. Genetic variants were determined using the 430,998 European exomes from the UK Biobank. Variants annotated as “start lost” and “stop lost” were included as missense variants in the missense gene burden codings. Predicted ESM-1v scores and predicted ESM- 1v+proteomics scores per missense variant were taken and these scores (scaled from 0 to 1) were rank-transformed over observed variants with allele frequency (AF) < 1% among pLoF- associated genes. When doing gene burden analysis, a threshold was defined such that all variants above would be considered a variant deleterious (and therefore consequential enough to be included in gene burden codings). In order to avoid erroneous conclusions, the threshold was varied from zero to 0.94 by 0.02 and then gene burden variant inclusion was tested at each threshold. To evaluate model performance as a function of gene burden association yield, multiple criteria were used. The first criterion used was the fraction of gold standard gene- trait pairs discovered with nominally significant association signal (p-value < 0.01). Multiple thresholds were also assessed to evaluate the robustness of the conclusion. The DOCKET NO.: 38120-4362 (11652WO01) PCT denominator was calculated as the actual number of gene trait pairs, so that gene-trait pairs missing association p-values (i.e., no variants meeting deleterious threshold) would be penalized. The second criterion considered was the Winsorized mean -log10(p-value), where -log10(p-values) were truncated at 20, to adjust for outliers. The third criterion considered was median -log10(p-value). The fourth criterion used was the mean absolute effect size across all gene-trait pairs. For the head-to-head comparisons, a one-sided binomial sign test was used to see if ESM-1b proteomics had higher numbers of traits with increased association strength. Yield analysis on ten UKB traits To assess the performance of ESM-1b+proteomics compared to the other approaches, combined gene-based testing was applied to ten UKB traits (birth weight, body mass index, creatinine, cystatin C, HDL cholesterol, hearing loss, height, LDL cholesterol, mean reticulocyte volume, and triglycerides). Using canonical Ensembl 100 genes (see Table 4 above), scores were generated from ESM-1b, ESM-1b+proteomics, AlphaMissense and ESM- 1v for all possible missense variants and genes. These scores were rank-transformed to a scale from zero to one, with a higher score indicating a more deleterious variant. The top 22% of scores were defined as deleterious based on all possible missense variants. This percentage matched the proportion of variants with 5 deleterious predictions from the conventional ensemble approach. This approach allowed a different fraction of deleterious variants per gene. Gene-based association testing was carried out using REGENIE. More precisely, multiple groupings of missense and pLoF variants were tested based on the annotation strategy and alternate allele frequency cutoffs (singletons, 0.01%, 0.1%, 0.5% and 1%), and all the tests were combined through a unified testing framework to result in a single p-value for each gene / trait pair. A genome-wide significance threshold of 5x10-9corresponding to a Bonferroni correction for the number of traits, genes and annotation strategies compared were defined. Example 3: Auxiliary Features ESM-1v+p was further refined, by way of processor 202, using auxiliary features, including AlphaMissense by Google DeepMind and / or conservations scores to generate ESM- 1v+Alpha+p, ESM-1v+Conserve+p and ESM-1v+Alpha+Conserve+p models. Overall, it was DOCKET NO.: 38120-4362 (11652WO01) PCT found that the further refined LLM also increased the power of deleterious missense burden- based tests, both compared to the original ESM-1v approach and when compared to an ensemble of conventional variant classifiers, and also when expanded beyond singleton missense variants to include all missense variants with allele frequency (AF) < 1%. The results are shown below in Table 8. Table 8: LLMs with auxiliary features improve power over all missense variants and 5 method ensemble in singleton missense variants and variants with allele frequency (AF) < 1% Singleton missense variants Allele frequency (AF) < 1% ClinVar247 missense variants genes 10 1000 3Optimal threshold for missense masks created with ESM-1v.4Optimal threshold for missense masks created with ESM-1v using -log10(MAF) as a training label for same set of variants as ESM-1v+p.5Optimal threshold for missense masks created with ESM-1v using proteomics labels.6Optimal threshold for missense masks created using AlphaMissense.7Optimal threshold for missense masks created with ESM-1v using proteomics labels (adding GERP / PhyLoP as additional features.8Optimal threshold for missense masks created with ESM-1v & AlphaMissense using proteomics labels.9Optimal threshold for missense masks created with ESM-1v & AlphaMissense & conservation features using proteomics labels.10Percentage of signals with p<0.01 across 247 traits. Table 9 further illustrates the power of deleterious missense burden-based tests applied to 186 of 247 non-proteomic gold standard traits without missense 5 holes. Table 9: Application to 186 of 247 non-proteomic gold standard traits without 5 holes Singleton missense variants Allele frequency (AF) < 1% missense variants DOCKET NO.: 38120-4362 (11652WO01) PCT All missense 414.4 25.8% 1577.0 67.2%5 / 5 missense 468.0 33.3% 1463.3 64.0%1+ / 5 missense 442.5 27.4% 1575.0 66.7% Tables 6 and 7 above was also conducted. In one assessment, the increase in yield using the AlphaMissense, ESM-1v, ESM-1v+Alpha+Conserve+p and ESM-1v+p models as a replacement for the 5-method ensemble approach was assessed, and the results shown in FIG.21A and Table 10 below. Table 10: Replacement of 5 method ensemble approach Method Threshold Increase AlphaMissense 4.60E-09 53 (3.68%) In another assessment, the increase in yield using the AlphaMissense, ESM-1v, ESM- 1v+Alpha+Conserve+p and ESM-1v+p models in combination with the 5 method ensemble approach was assessed, and the results shown in FIG.21B and Table 11 below. Table 11: Combination with 5 method ensemble Method Threshold Increase DOCKET NO.: 38120-4362 (11652WO01) PCT AlphaMissense 5.00E-10 68 (5.41%)ESM-1v 5.00E-10 48 (3.82%) ESM-1v+Al ha+Conserve+ 500E-10 79 (629%) FIG. 22 illustrates a CNN model for ESM-1v feature refinement for a single variant using auxiliary feature data. The CNN model illustrated in FIG.22 can be used to in the ESM- 1v+Alpha+p, ESM-1v+Conserve+p and ESM-1v+Alpha+Conserve+p models described in Example 3. In some embodiments, the second predictive model 718 (shown in FIG. 7) comprises a convolutional neural network. By leveraging a convolutional neural network, each row of the input matrix (e.g., matrix 920 shown in FIG.9) could have a kernel applied to it, as shown in FIG.22, thereby dramatically reducing the parameter space as compared to a standard feed forward network. As a non-limiting example, in FIG. 22, 301x4 kernels 2204 are applied to the 1281x4 input matrix 2202 to create a 1281x30 matrix 2206. Next, one 1x30 kernel 2208 is applied to summarize the resulting tensor to a 1281x1 vector. Next, 30 fully connected linear units 2212 with Gaussian Error Linear Unit (GELU) activation are applied to the 1281x1 vector 2210. In the embodiment illustrated in FIG.22, the fully connected linear units 2212 include auxiliary feature data. In some embodiments, the auxiliary feature data includes one or more auxiliary features which can include, but are not limited to, any feature from one or more other predictive models configured to predict the effect of genetic variants, conservation scores, and the like. The one or more other predictive models may include, but are not limited to, AlphaMissense by Google DeepMind. A final fully connected layer is applied with a sigmoid activation function to give an output 2214 that ranges from 0 to 1, where 1 most closely resembles a pLoF variant. In some embodiments, the auxiliary feature data is used to re-train the second predictive model 718 (shown in FIG. 7). For example, the auxiliary feature data, such as features derived from another predictive model such as AlphaMissense, can be compared to the output from the trained model, and the trained model can be re-trained, as described DOCKET NO.: 38120-4362 (11652WO01) PCT above with respect to FIGS.2 and 3, to improve performance of the second predictive model 718. Machine learning methods are disclosed herein. The computer-implemented methods discussed herein may include additional, less, or alternate actions, including those discussed elsewhere herein. The methods may be implemented via one or more local or remote processors, transceivers, servers and / or via computer-executable instructions stored on non-transitory computer-readable media or medium. Additionally, the computer systems discussed herein may include additional, less, or alternate functionality, including that discussed elsewhere herein. The computer systems discussed herein may include or be implemented via computer-executable instructions stored on non-transitory computer-readable media or medium. A processor or a processing element may be trained using supervised or unsupervised machine learning, and the machine learning program may employ a neural network, which may be a convolutional neural network, a deep learning neural network, a reinforced or reinforcement learning module or program, or a combined learning module or program that learns in two or more fields or areas of interest. Machine learning may involve identifying and recognizing patterns in existing data in order to facilitate making predictions for subsequent data. Models may be created based upon example inputs in order to make valid and reliable predictions for novel inputs. Additionally, or alternatively, the machine learning programs may be trained by inputting sample data sets or certain data into the programs, such as synonymous variant data, pLoF variant data, missense variant data, proteomic data, and / or other data. The machine learning programs may utilize deep learning algorithms that may be primarily focused on pattern recognition, and may be trained after processing multiple examples. The machine learning programs may include Bayesian Program Learning (BPL), voice recognition and synthesis, image or object recognition, optical character recognition, and / or natural language processing – either individually or in combination. The machine learning programs may also include natural language processing, semantic analysis, automatic reasoning, and / or machine learning. DOCKET NO.: 38120-4362 (11652WO01) PCT Supervised and unsupervised machine learning techniques may be used. In supervised machine learning, a processing element may be provided with example inputs and their associated outputs, and may seek to discover a general rule that maps inputs to outputs, so that when subsequent novel inputs are provided the processing element may, based upon the discovered rule, accurately predict the correct output. In unsupervised machine learning, the processing element may be required to find its own structure in unlabeled example inputs. In some embodiments, machine learning techniques may be used to extract data about a missense variant based on data derived from protein language models, proteomic data, and / or other data. In some embodiments, the voice bots or chatbots discussed herein may be configured to utilize ML and / or AI techniques. For instance, the voice bot or chatbot may be an AI chatbot. The voice bot or chatbot may employ supervised or unsupervised machine learning techniques, which may be followed by, and / or used in conjunction with, reinforced or reinforcement learning techniques. As will be appreciated based upon the foregoing specification, the above-described embodiments of the disclosure may be implemented using computer programming or engineering techniques including computer software, firmware, hardware or any combination or subset thereof. Any such resulting program, having computer-readable code means, may be embodied or provided within one or more computer-readable media, thereby making a computer program product, i.e., an article of manufacture, according to the discussed embodiments of the disclosure. The computer-readable media may be, for example, but is not limited to, a fixed (hard) drive, diskette, optical disk, magnetic tape, semiconductor memory such as read-only memory (ROM), SD card, memory device and / or any transmitting / receiving medium, such as the Internet or other communication network or link. The article of manufacture containing the computer code may be made and / or used by executing the code directly from one medium, by copying the code from one medium to another medium, or by transmitting the code over a network. These computer programs (also known as programs, software, software applications, “apps”, or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object-oriented programming language, DOCKET NO.: 38120-4362 (11652WO01) PCT and / or in assembly / machine language. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The “machine-readable medium” and “computer-readable medium,” however, do not include transitory signals. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor. That is, also described here is a computer program comprising instructions, which, when the program is executed by a computer cause the computer to carry out the method of any of the claims. As used herein, a processor may include any programmable system including systems using micro-controllers, reduced instruction set circuits (RISC), application specific integrated circuits (ASICs), logic circuits, and any other circuit or processor capable of executing the functions described herein. The above examples are example only, and are thus not intended to limit in any way the definition and / or meaning of the term “processor.” As used herein, the terms “software” and “firmware” are interchangeable, and include any computer program stored in memory for execution by a processor, including RAM memory, ROM memory, EPROM memory, EEPROM memory, and non-volatile RAM (NVRAM) memory. The above memory types are example only, and are thus not limiting as to the types of memory usable for storage of a computer program. In some embodiments, a computer program is provided, and the program is embodied on a computer readable medium. In an example embodiment, the system is executed on a single computer system, without requiring a connection to a server computer. In a further embodiment, the system is being run in a Windows® environment (Windows is a registered trademark of Microsoft Corporation, Redmond, Washington). In yet another embodiment, the system is run on a mainframe environment and a UNIX® server environment (UNIX is a registered trademark of X / Open Company Limited located in Reading, Berkshire, United Kingdom). The application is flexible and designed to run in various different environments without compromising any major functionality. DOCKET NO.: 38120-4362 (11652WO01) PCT In some embodiments, the system includes multiple components distributed among a plurality of computing devices. One or more components may be in the form of computer- executable instructions embodied in a computer-readable medium. The systems and processes are not limited to the specific embodiments described herein. In addition, components of each system and each process can be practiced independent and separate from other components and processes described herein. Each component and process can also be used in combination with other assembly packages and processes. The present embodiments may enhance the functionality and functioning of computers and / or computer systems. As used herein, an element or step recited in the singular and preceded by the word “a” or “an” should be understood as not excluding plural elements or steps, unless such exclusion is explicitly recited. Furthermore, references to “example embodiment” or “one embodiment” of the present disclosure are not intended to be interpreted as excluding the existence of additional embodiments that also incorporate the recited features. The patent claims at the end of this document are not intended to be construed under 35 U.S.C. § 112(f) unless traditional means-plus-function language is expressly recited, such as “means for” or “step for” language being expressly recited in the claim(s). This written description uses examples to disclose the disclosure, including the best mode, and also to enable any person skilled in the art to practice the disclosure, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the disclosure is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insubstantial differences from the literal language of the claims.

Claims

DOCKET NO.: 38120-4362 (11652WO01) PCT What Is Claimed Is:

1. A system for predicting one or more effects of genetic variants, the system comprising: at least one memory storing computer-executable instructions; and at least one processor in communication with the at least one memory, wherein the at least one processor is configured to execute the computer-executable instructions to: receive an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and in response to the numerical value being above a predetermined threshold, associate the variant with a trait.

2. The system of claim 1, wherein the at least one processor is further configured to execute the computer-executable instructions to: build a classifier for at least one of synonymous variants or putative loss of function (pLoF) variants using summary statistics from a proteomic-trait-loci (pQTL) analysis; and apply the classifier to missense variants to generate continuous labels for fine- tuning the machine learning model.

3. The system of claim 1, wherein the at least one processor is further configured to execute the computer-executable instructions to train the machine learning model using low-rank adaptation (LoRA).DOCKET NO.: 38120-4362 (11652WO01) PCT 4. The system of claim 1, wherein the one or more outputs of the machine-learning model comprise a prediction of a deleterious effect of the variant.

5. The system of claim 1, wherein the variant is a missense variant.

6. The system of claim 1, wherein the at least one processor is further configured to, during execution of the neural network: extract a last hidden layer and concatenating into separate channels; apply a 1x1 convolutional kernel to merge the channels into a single matrix; apply one or more kernels of a predetermined hidden size to a resulting matrix with adaptive pooling; apply nonlinear activations and a drop out operator to generate an output; and pass the output to the fully connected layer.

7. The system of claim 6, wherein the at least one processor is further configured to, when executing the neural network, add one or more features selected from log10 p-value, beta value, allele frequency, proteomic-trait-loci (pQTL) effect size, and standard error.

8. A computer-based method for predicting one or more effects of a genetic variant, the method implemented using a system including a computing device including a processor communicatively coupled to memory device, the method comprising: receiving an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural networkDOCKET NO.: 38120-4362 (11652WO01) PCT receiving one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of the variant; and in response to the numerical value being above a predetermined threshold, associating the variant with a trait.

9. The computer-based method of claim 8, further comprising: building a classifier for at least one of synonymous variants or putative loss of function (pLoF) variants using summary statistics from a proteomic-trait-loci (pQTL) analysis; and applying the classifier to missense variants to generate continuous labels for fine- tuning the machine learning model.

10. The computer-based method of claim 8, further comprising training the machine learning model using low-rank adaptation (LoRA).

11. The computer-based method of claim 8, wherein the one or more outputs of the machine-learning model comprise a prediction of a deleterious effect of the variant.

12. The computer-based method of claim 8, wherein the variant is a missense variant.

13. The computer-based method of claim 8, further comprising, when executing the neural network: extracting a last hidden layer and concatenating into separate channels; applying a 1x1 convolutional kernel to merge the channels into a single matrix; applying one or more kernels of a predetermined hidden size to a resulting matrix with adaptive pooling; applying nonlinear activations and a drop out operator to generate an output; andDOCKET NO.: 38120-4362 (11652WO01) PCT passing the output to the fully connected layer.

14. The computer-based method of claim 13, further comprising, when executing the neural network, adding one or more features selected from log10 p-value, beta value, allele frequency, proteomic-trait-loci (pQTL) effect size, and standard error from proteomic association models.

15. At least one non-transitory computer-readable storage media having computer- executable instructions embodied thereon, wherein when executed by at least one processor, the computer-executable instructions cause the at least one processor to: receive an amino acid sequence, the amino acid sequence with at least one altered amino acid, and the amino acid sequence masked out; apply the amino acid sequence, the amino acid sequence with the at least one altered amino acid, and the amino acid sequence masked out to a machine learning model previously trained using data labelled with at least one proteomic label, wherein the machine learning model includes a neural network; receive one or more outputs from the machine learning model, the one or more outputs comprising a numerical value corresponding to one or more effects of a variant; and in response to the numerical value being above a predetermined threshold, associate the variant with a trait.

16. The at least one non-transitory computer-readable storage media of claim 15, wherein the computer-executable instructions further cause the at least one processor to: build a classifier for at least one of synonymous variants or putative loss of function (pLoF) variants using summary statistics from a proteomic-trait-loci (pQTL) analysis; and applying the classifier to missense variants to generate continuous labels for fine- tuning the machine learning model.DOCKET NO.: 38120-4362 (11652WO01) PCT 17. The at least one non-transitory computer-readable storage media of claim 15, wherein the computer-executable instructions further cause the at least one processor to train the machine learning model using low-rank adaptation (LoRA).

18. The at least one non-transitory computer-readable storage media of claim 15, wherein the one or more outputs of the machine-learning model comprise a prediction of a deleterious effect of the variant.

19. The at least one non-transitory computer-readable storage media of claim 15, wherein the variant is a missense variant.

20. The at least one non-transitory computer-readable storage media of claim 15, wherein the computer-executable instructions further cause the at least one processor to, during execution of the neural network: extract a last hidden layer and concatenating into separate channels; apply a 1x1 convolutional kernel to merge the channels into a single matrix; apply one or more kernels of a predetermined hidden size to a resulting matrix with adaptive pooling; apply nonlinear activations and a drop out operator to generate an output; and pass the output to the fully connected layer.