High-throughput prediction of variant effects from conformational dynamics
By employing molecular dynamics and machine learning to analyze conformational dynamics, the method addresses the scalability and accuracy issues in predicting genetic variant effects, facilitating high-throughput genetic testing without requiring biological expertise.
Patent Information
- Application Number
- JP2024522607
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-10-13
- Filing Date
- 2022-10-13
- Publication Date
- 2024-11-06
AI Technical Summary
Current methods for predicting the functional impact of genetic variations, such as missense mutations, are not scalable and require significant human curation, making them unsuitable for high-throughput genetic testing, and they lack accuracy in determining pathogenicity due to reliance on structural stability changes alone.
A computer-implemented method using molecular dynamics to generate conformational dynamics of molecular structures, followed by dimensionality reduction and machine learning to classify variant effects, enabling high-throughput classification of genetic variants by analyzing conformational dynamics independent of specific biological systems.
This approach allows for rapid and accurate classification of genetic variants, reducing simulation time and eliminating the need for biological expertise, thereby enhancing the throughput and precision of genetic testing.
Smart Images

Figure 2024540883000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to methods for determining the functional impact of genetic variations, e.g., protein missense mutations, using machine learning on the conformational dynamics of biomolecules. [Background technology]
[0002] Accurate variant effect prediction is crucial in genetic research and clinical diagnosis. Missense variants are the most common type of coding genetic variant and a major class of genetic risk across a wide range of common and rare diseases. De novo missense mutations contribute significantly to structural birth defects (Homsy et al. (2015) Science 350:1262-1266, Jin et al. (2017) Nat. Genet. 49:1593-1601) and neurodevelopmental disorders (Iossifov et al. (2014) Nature 514:216-221, De Rubeis et al. (2014) Nature 515:209-215, McRae et al. (2017) Nature 542:433-438). Nevertheless, only a small proportion of missense de novo mutations are pathogenic and result in disease. In clinical genetic testing, many missense variants in well-established risk genes are classified as variants of uncertain significance, indicating that the current clinical information known is not sufficient to determine whether the variant is pathogenic or benign.
[0003] Interpretation of the molecular level effects of genetic variants (e.g., missense gene variants) remains challenging and requires specific investigation of the mutation (either at the amino acid or nucleic acid level) in the context of molecular structure and function. Since it has been reported that the deleterious and / or neutral outcomes of genetic variants are related to the structural properties of the affected molecules (Sivley et al. (2018) Am. J. Hum. Genet. 102:415-426, Hickset al. (2019) Proc. Natl. Acad. Sci. USA 116:8960-8965), resources and tools have been developed to map and visualize variants (e.g., missense variants) on molecular structures (e.g., protein structures), such as ASTRID, mutation3D, COSMIC-3D, VarSite, VarMap, or PhyreRisk. However, inferring the biological effect of a mutation from mere visualization is speculative at best. Current high-throughput methods that use structural features can predict changes in protein structure that are typically associated with changes in the overall stability of the protein, a mechanism by which many loss-of-function (LOF) missense mutations are pathogenic. However, little is known about how changes in conformational dynamics of molecular structures are related to pathogenicity. Thus, determining the impact of variants on conformational dynamics and correlating the observed dynamics with pathogenicity represents a growing field of research. Molecular dynamics (MD) methods, which are computational algorithms for determining conformational dynamics of molecular structures, have been used, for example, in the methods described in U.S. Publication No. 2019 / 0189243A1 and McCoy et al. (2020) Biophysical Journal 120(2):189-204. However, because these methods rely on human curation to identify and select features for machine learning and the resulting predictive models are specific to a particular biological system (e.g., a particular protein associated with a disease or disorder), methods currently known in the art are not scalable and therefore cannot operate in a true high-throughput mode.Furthermore, the computational algorithms used to determine the conformational dynamics of molecular structures are currently too time-consuming to be used for predicting clinical variant effects in scalable genetic testing systems. Summary of the Invention
[0004] The present disclosure provides a computer-implemented method for high-throughput classification of variant effects, the method comprising: (i) generating conformational dynamics of a reference molecular structure (e.g., using molecular dynamics (MD)); (ii) processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; (iii) extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; and (iv) introducing at least one variant into the extracted conformational state of the reference molecular structure by: The method includes: (v) generating a variant molecular structure or a set thereof; (v) generating conformational dynamics of the variant molecular structure or a set thereof (e.g., using MD); (vi) processing the conformational dynamics of the variant molecular structure or a set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or a set thereof; and (vii) training a machine learning prediction model using the low-dimensional representation of the conformational landscape of the reference molecular structure and the variant molecular structure using clinical data as training labels, where the prediction model classifies the variant effect in the reference molecular structure. In some embodiments, generating the conformational dynamics of the reference molecular structure or the variant molecular structure includes (a) obtaining a reference molecular structure or a variant molecular structure; and (b) using the obtained reference molecular structure or the variant protein structure to generate the conformational dynamics. In some embodiments, the reference molecular structure or the variant molecular structure is obtained from the Protein Data Bank (PDB) or AlphaFold. In some embodiments, the conformational dynamics are generated using GROMACS, OpenMM, NAMD, Amber, or LAMPS.
[0005] In some embodiments, the generation of the conformational dynamics is parallelized. In some embodiments, the reference molecular structure is a three-dimensional structure of a wild-type protein. In some embodiments, each of step (i) and / or step (ii) of the method disclosed above independently comprises at least about 10,000, at least about 20,000, at least about 30,000, at least about 40,000, at least about 50,000, at least about 60,000, at least about 70,000, at least about 80,000, at least about 90,000, at least about 100,000, at least about 200,000, at least about 300,000, at least about 400,000, at least about 500,000, at least about 600,000, at least about 700,000, at least about 800,000, at least about 900,000, or at least about 1,000,000 simulated structural conformations including conformational dynamics.
[0006] In some aspects, processing the conformational dynamics to generate a reduced dimensional representation of the conformational landscape of the reference molecular structure comprises extracting structural features from the conformational dynamics. In some aspects, the structural features are selected from the group consisting of bond angles, inter-residue distances, residue variation, surface accessibility, domain relative orientation, binding site interactions, and any combination thereof. In some aspects, the extraction of structural features results in at least a 20-fold, 30-fold, 40-fold, or 50-fold reduction of the structural features.
[0007] In some embodiments, processing the conformational dynamics further comprises dimensionality reduction. In some embodiments, the dimensionality reduction is obtained using principal component analysis (PCA) or an autoencoder. In some embodiments, the dimensionality reduction results in at least a 50-fold, 60-fold, 70-fold, 80-fold, 90-fold, or 100-fold reduction in structural features.
[0008] In some aspects, the low-dimensional representation of the conformational landscape identifies conformations in which the reference molecular structure and / or the variant molecular structure are in metastable conformational states, hi some aspects, the low-dimensional representation comprises motional modes that characterize conformational changes associated with molecular function.
[0009] In some embodiments, extracting metastable conformational states from the low-dimensional conformational landscape comprises clustering the conformational states based on their representation in the low-dimensional space. In some embodiments, clustering the conformational states comprises k-means clustering, hierarchical clustering, density-based clustering, DBSCAN, spectral clustering, Gaussian mixture model, or any combination thereof. In some embodiments, the method further comprises prioritizing the clustered conformational states. In some embodiments, prioritizing the clustered conformational states is based on cluster properties. In some embodiments, the cluster properties are selected from the group consisting of distance from the conformational landscape centroid, cluster occupancy, cluster distribution statistics, structural properties of the cluster, thermodynamics of the cluster, or any combination thereof. In some embodiments, the structural properties are selected from the group consisting of bond angles, inter-residue distances, surface accessibility, domain relative orientation, binding site interactions, structural similarity to known protein conformations, and any combination thereof. In some embodiments, the bond angles include φ-ψ dihedral angles, side chain χ angles, or a combination thereof. In some embodiments, the cluster distribution statistic is selected from the group consisting of Silhouette score, elbow score, Calinski-Harabasz index, Rand index, mutual information, homogeneity, completeness, V-scale, Davies-Bouldin index, and any combination thereof.
[0010] In some embodiments, extracting metastable conformational states from the low-dimensional conformational landscape comprises extracting a representative conformational state. In some embodiments, the representative conformational state is the conformational state closest to the center of gravity of all conformational states of the cluster. In some embodiments, at least one variant is a pathogenic mutation. In some embodiments, at least one variant is a non-pathogenic (benign) mutation. In some embodiments, the variant molecular structure or set thereof comprises a variant that has been or may be observed in a medical genetic test. In some embodiments, the variant molecular structure or set thereof comprises a pathogenic variant. In some embodiments, the variant molecular structure or set thereof comprises a non-pathogenic (benign) variant. In some embodiments, the variant molecular structure or set thereof comprises a variant of unknown significance. In some embodiments, the variant molecular structure or set thereof is generated using a backbone-dependent rotamer library. In some embodiments, the method comprises identifying side chain conformations (rotamers) with existing backbone dihedral angles of the variant molecular structure at the variant positions in the backbone-dependent rotamer library. In some embodiments, the method further comprises testing whether the introduction of the backbone-dependent rotamer library for which the rotamers have been identified causes steric clashes with nearby residues. In some embodiments, the testing is performed until a rotamer that minimizes the steric clashes is identified. In some embodiments, the method further comprises performing minimization and equilibration simulations to regularize the conformational geometry of the variant molecular structure and reduce the energetic impact of introducing the variant into the variant molecular structure.
[0011] In some embodiments, the predictive model is generated using machine learning. In some embodiments, the clinical data used as training labels includes biomarker status (e.g., the presence or absence or expression level of a particular biomarker), biometric data, lifestyle-related data, response to treatment, symptoms of a disease or condition, protein expression data, type of treatment administered, dosage, dosage regimen, route of administration, presence or absence of concomitant therapy, response to therapy, age, weight, sex, ethnicity, ClinVar submissions, INVITAE™ clinical interpretations, data from other experimental or computational models, or any combination thereof. In some embodiments, the classification of variant effects in the reference molecular structure includes the calculation of predicted pathogenicity probability. In some embodiments, the machine learning predictive model is generated using logistic regression. In some embodiments, the machine learning predictive model is generated using random forests. In some embodiments, the machine learning predictive model is generated using artificial neural networks.
[0012] The disclosure also provides a computer-implemented method for scoring the clinical significance of a variant, the method comprising: (a) generating (e.g., using MD) a conformational dynamics of a reference molecular structure; (b) processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; (c) extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; and (d) introducing at least one variant into the extracted conformational state of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the variant molecular structure. (e) generating conformational dynamics of the variant molecular structure or set thereof (e.g., using MD); (f) processing the conformational dynamics of the variant molecular structure or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or set thereof; and (g) training a machine learning predictive model using the low-dimensional representation of the conformational landscape of the reference molecular structure and the variant molecular structure using the clinical data as training labels, where the predictive model outputs a score of clinical significance of the variant.
[0013] The present disclosure also provides a system, comprising: (a) a memory; and (b) at least one processor, coupled to the memory, for (i) generating (e.g., using MD) a conformational dynamics of a reference molecular structure; (ii) processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; (iii) extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; and (iv) generating a variant molecular structure by introducing at least one variant into the extracted metastable conformational state of the reference molecular structure. (v) generate conformational dynamics of the variant molecular structure or set thereof (e.g., using MD); (vi) process the conformational dynamics of the variant molecular structure or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or set thereof; and (vii) train a machine learning predictive model using the low-dimensional representation of the conformational landscape of the reference molecular structure and the variant molecular structure using the clinical data as training labels, where the predictive model classifies the variant effect in the reference molecular structure.
[0014] The present disclosure also provides a tangible computer readable device having instructions stored thereon that, when executed by at least one computing device, cause the at least one computing device to: (i) generate (e.g., using MD) a conformational dynamics of a reference molecular structure; (ii) process the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; (iii) extract metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; and (iv) assign at least one metastable conformational state to the extracted metastable conformational states of the reference protein structure. (v) generating conformational dynamics of the variant molecular structures or set thereof (e.g., using MD); (vi) processing the conformational dynamics of the variant molecular structures or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structures or set thereof; and (vii) training a machine learning predictive model using the low-dimensional representation of the conformational landscape of the reference molecular structures and the variant molecular structures using clinical data as training labels, where the predictive model classifies the variant effect in the reference molecular structures.
[0015] In some aspects of the methods, systems, or tangible computable readable devices disclosed above, the classification of the variant in the variant molecular structure can be used to (i) treat a patient, (ii) select a patient for treatment, (iii) initiate treatment, (iv) discontinue treatment, (v) interrupt treatment, (vi) modify treatment, or (vii) any combination thereof.
[0016] In some aspects, the present disclosure provides a method of (i) treating a patient, (ii) selecting a patient for treatment, (iii) initiating treatment, (iv) ceasing treatment, (v) interrupting treatment, (vi) modifying treatment, or (vii) any combination thereof, comprising applying any of the methods, systems, or tangible computable readable devices disclosed above.
[0017] The present disclosure also provides personalized medicine treatments including therapeutic agents capable of treating a disease or disorder caused by a variant identified as pathogenic by any of the methods disclosed above or by using any of the systems or tangible computable readable devices disclosed above.
[0018] The present disclosure also provides a computer-implemented method for generating a low-dimensional representation of a conformational landscape of a molecular structure, the method comprising: (i) extracting structural features from the conformational landscape of the molecular structure; and (ii) reducing the dimensionality of the conformational landscape. In some aspects, the structural features are selected from the group consisting of bond angles, inter-residue distances, residue variation, surface accessibility, domain relative orientation, binding site interactions, and any combination thereof. In some aspects, extracting the structural features results in at least a 20-fold, 30-fold, 40-fold, or 50-fold reduction in the structural features. In some aspects, reducing the dimensionality includes using principal component analysis (PCA) or an autoencoder. In some aspects, reducing the dimensionality results in at least a 50-fold, 60-fold, 70-fold, 80-fold, 90-fold, or 100-fold reduction in the structural features.
[0019] The present disclosure also provides a computer-implemented method for identifying metastable conformational states of a molecular structure, the method comprising clustering a low-dimensional representation of a conformational landscape of the molecular structure. In some embodiments, clustering the conformational states comprises k-means clustering, hierarchical clustering, density-based clustering, DBSCAN, spectral clustering, Gaussian mixture model, or any combination thereof. In some embodiments, the method further comprises prioritizing the clustered conformational states. In some embodiments, prioritizing the clustered conformational states is based on cluster properties. In some embodiments, the cluster properties are selected from the group consisting of distance from a conformational landscape centroid, cluster occupancy, cluster distribution statistics, structural properties of the cluster, thermodynamics of the cluster, or any combination thereof. In some embodiments, the structural properties are selected from the group consisting of bond angles, inter-residue distances, surface accessibility, domain relative orientation, binding site interactions, structural similarity to known protein conformations, and any combination thereof. In some embodiments, the bond angles include φ-ψ dihedral angles, side chain χ angles, or a combination thereof. In some embodiments, the cluster distribution statistic is selected from the group consisting of Silhouette score, elbow score, Calinski-Harabasz index, Rand index, mutual information, homogeneity, completeness, V-scale, Davies-Bouldin index, and any combination thereof.
[0020] The present disclosure also provides a computer-implemented method for high-throughput classification of variant effects in a molecular structure, the method comprising: (i) generating a low-dimensional representation of a conformational landscape of the molecular structure; and (ii) identifying metastable conformational states from the low-dimensional representation of the conformational landscape of the molecular structure. In some aspects, generating the low-dimensional representation of the conformational landscape of the molecule comprises: (a) extracting structural features from the conformational landscape of the molecular structure; and (b) reducing the dimensionality of the conformational landscape. In some aspects, identifying metastable conformational states from the low-dimensional representation of the conformational landscape of the molecular structure comprises clustering the low-dimensional representation of the conformational landscape of the molecular structure. [Brief description of the drawings]
[0021] [Figure 1] FIG. 1 shows a system diagram of the prediction system of the present disclosure. [Diagram 2] FIG. 1 shows a workflow diagram illustrating successive steps in a prediction system for predicting variant effects from conformational dynamics of biomolecules. [Diagram 3] FIG. 1 shows a schematic diagram of the dimensionality reduction part of the workflow, including the structural feature extraction and dimensionality reduction steps. [Figure 4] 1 shows a schematic diagram of the conformational energy states of a reference molecule and the metastable conformational energy states of a variant molecule. [Diagram 5] We illustrate the relationship between features such as backbone ψ angle distribution and pathogenicity of variant molecules. [Figure 6]FIG. 1 shows a low-dimensional conformational landscape representation resulting from clustering of >76,000 structural conformations from the conformational dynamics of a reference molecule that has been subjected to structural feature extraction and dimensionality reduction. This low-dimensional representation is then subjected to clustering. Labels identify specific conformational clusters, each of which represents a metastable conformation of the reference molecule. Component 1 and Component 2 are the first two representative components from the principal component analysis. [Figure 7] We show that conformational clusters can be used to distinguish pathogenic and benign variants (e.g., missense genetic variants). [Figure 8] The figure shows various amounts of conformational data (dataset definitions) and feature quantification methods used to optimize the machine learning model. For each combination of dataset and feature quantification method (except when they were not calculated: NC), the machine learning model is trained on features extracted from the simulations included in the dataset definition. The performance of the models is evaluated using 5-fold cross-validation on the training set to determine the best performing model according to a balanced weighted F1 score. The performance of the final model (in this case dataset "3ns, 2 iterations, 11 states" and feature quantification method "connectivity tICA") is then determined using the performance of the initial models and the balanced weighted F1 score of a hold-out test set of simulations that were not used in the evaluation. [Figure 9] We compare the performance of our model (the LSS model) across a selection of variants, compared to the true labels, compared to the performance of the same variants in the sequence-based and FoldX-based models. Overall, the predictions of the LSS model for these variants are closer to the true labels than any of the other models. [Figure 10] The performance of a predictive model for MSH2 (MutS homolog 2) tumor suppressor developed according to the method disclosed herein (LSS) is compared to alternative methods, such as FoldX and Polyphen. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] The present disclosure provides a truly high-throughput platform that can accurately predict the phenotype, phenotypic effect, or biological effect of variants (e.g., missense variants) (e.g., whether they are pathogenic and the severity of their pathological phenotype) using conformational dynamics (e.g., generated from molecular dynamics) of molecular structures (e.g., protein structures) that are independent of a specific biological system, eliminating the need for biological and structural biology expertise to train machine learning models.
[0023] These insights can aid in the interpretation of variants of uncertain significance from medical genetic testing and guide pharmacological interventions using precision medicine principles.
[0024] In contrast, conventional variant (e.g., missense variant) classification systems in the art cannot be truly high-throughput platforms that can accurately predict variant effects from conformational dynamics, and such systems typically require biological and structural biology expertise to train machine learning models. Aspects of the present disclosure achieve superior performance and throughput over other methods in the art by characterizing and reducing the dimensionality of conformational dynamics in a system-independent manner. Thus, the need for biological and structural biology human expertise to train machine learning models is eliminated. Furthermore, the process of determining metastable conformational states from a reference molecular system, determining the conformational dynamics of variants introduced into these states, and training machine learning models on low-dimensional representations of these states significantly reduces the simulation time required to train accurate models, thereby increasing the throughput of this approach in variant interpretation systems. Finally, this approach provides a way to rapidly make variant effect predictions for newly observed variants by storing pre-trained models for inference.
[0025] The present disclosure provides a computer-implemented method for high-throughput prediction of biological effects of variants (e.g., missense gene variants) from conformational dynamics of molecular structures, the method including, for example, (a) generating conformational dynamics of a reference molecule (e.g., a reference protein); (b) defining a low-dimensional conformational landscape of the reference molecule using an unsupervised dimensionality reduction method; (c) determining a metastable conformational state or set thereof from the low-dimensional conformational landscape of the reference molecule; and (d) constructing a variant molecule (e.g., a variant protein) or set thereof by introducing at least one molecular change to the reference molecule in the metastable conformational state or set thereof; (e) determining conformational dynamics of the variant molecule or set thereof in a metastable conformational state or set thereof; (f) using an unsupervised dimensionality reduction method to generate a low-dimensional representation of the conformational dynamics of the variant molecule or set thereof in a metastable conformational state or set thereof; (g) training a predictive model using the low-dimensional representation of the conformational dynamics of the variant molecule or set thereof using biological or clinical data as training labels, where the model predicts a biological effect of the variant (e.g., a missense mutation) relative to a reference molecule; and, optionally, (h) rapidly extending the predictive model to newly observed variants from medical genetic testing.
[0026] In order to generate conformational dynamics of a molecular structure (eg, a protein structure), the molecular structure must be defined within a molecular system.
[0027] As used herein, a "molecular system" refers to a collection of molecular structures of biomolecules such as proteins, peptides, nucleic acids, or other biologically relevant molecules (e.g., lipids, glycans, or small molecules). A molecular system may also include solvents or other molecules (e.g., ions, lipids) that define the realistic biological composition of the molecular components in the system. A "molecular system" includes a collection of molecular components that unite the system (e.g., proteins, solvated water, ions, etc.), and a "conformational state" of a molecular system is the three-dimensional arrangement of the molecular components of a molecular system at a point in time. Thus, a conformational state represents a "snapshot" of a molecular system. As used herein, the terms "conformation," "conformational structure," "structure," and "molecular structure" may be used interchangeably with "conformational state."
[0028] It should be understood that the reference molecule does not assume a structure in a vacuum, and therefore additional components of the system (e.g., water molecules, ions, etc.) must be present. Thus, in the context of this disclosure, reference to a "reference molecular structure" or a "variant molecular structure" refers to the presence of those additional components. Thus, the terms "reference molecular system" and "reference molecular structure" are considered equivalent and interchangeable. Similarly, the terms "variant molecular system" and "variant molecular structure" are considered equivalent and interchangeable.
[0029] As used herein, "molecular structure," "molecule structure," and grammatical variations thereof, refer to the three-dimensional coordinates and identities of the atoms that make up some or all of a biological molecule (e.g., a protein) or set thereof, and the interactions (e.g., covalent and non-covalent bonds) between those atoms. Molecular structures can be represented, for example, by "coarse-graining" methods (such as the MARTINI algorithm) that reduce the complexity of a molecular structure by replacing a set or sets of atoms with "dummy" atoms that have properties intended to represent combinations of atoms (such as amino acids) and their interactions with other atoms.
[0030] The molecular system can be generated by obtaining molecular structures (e.g., crystallographic protein structures) from a database of experimental structures, such as the Protein Data Bank (PDB) or the Electron Microscopy Data Bank (EMDB), independently determining molecular structures through experimental or computational methods and combining them into one or more data files, or obtaining molecular structures from an existing database of computationally predicted structures, such as the ModBase database of AlphaFold predicted structures or the European Biological Laboratory database. Alternatively, molecular structures can be predicted using protein structure prediction algorithms such as EVFold, AlphaFold, and RoseTTAFold, or generated by homology modeling. Molecular structures can be further informed by experimental methods (e.g., single molecule biophysics, large-scale mutagenesis experiments (e.g., deep mutation scanning), mass spectrometry). Molecular structures can also be determined by any combination of the above in an approach called integrative modeling.
[0031] As used herein, the term "molecule" refers to a biological entity such as a protein, lipid, nucleic acid, carbohydrate, and combinations thereof. As used herein, the term "molecule" encompasses both single and multiple molecular entities. Thus, in some aspects, a molecule as defined herein may be, for example, a multimeric complex comprising multiple protein subunits, or a protein complexed with a nucleic acid. In some aspects, the term molecule refers to a biological molecule such as a protein or nucleic acid complexed with a ligand, for example, an antagonist or agonist such as a receptor, a cofactor, a drug, etc. Particular types of molecules according to the present disclosure are reference molecules and variant molecules. "Molecule" refers to the composition of an entity, and "conformation" or "structure" refers to the arrangement of a molecule in space.
[0032] As used herein, the term "reference molecule" refers to a molecule (e.g., a protein) that is to be investigated by the introduction of a variant (e.g., a missense variant). The reference molecule (e.g., a protein) may be a natural protein (e.g., a wild-type protein). Thus, in some aspects, the reference molecule structure may be the structure of a natural protein (e.g., a wild-type protein) or a set thereof. In other aspects, instead of being wild-type, the reference molecule (e.g., a reference protein) may contain a common or non-common variant (e.g., a missense variant). Thus, in other aspects, the reference molecule structure may be the structure of a variant molecule (e.g., a missense variant protein) or a set thereof. Thus, the reference molecule (e.g., a reference protein) does not have to contain a genetic variant, but should be used as a reference molecule for comparison with a set of other variant molecules having the same background system, i.e., the variant molecule (e.g., a variant protein) may be used as a reference molecule to determine the effect of a variant (e.g., a mutation) in a set of other variant molecules (e.g., other variant proteins having a different set of mutations).
[0033] The reference molecule is defined by its composition (identity and relationships of atoms within the molecule). The method disclosed herein includes obtaining a structure (reference molecule structure) from a database that represents the reference molecule in a particular conformational state. Methods such as MD can be used to generate conformational dynamics of the reference molecule. These conformational dynamics approximate the conformational landscape of the reference molecule. The conformational landscape is the collection of energetically accessible (taking into account parameters, i.e., temperature, pressure of the system) conformational states of the reference molecule.
[0034] A "reference molecular structure," "reference molecule structure," and grammatical variations thereof, can refer to a single conformational state of the structure of a reference molecule (e.g., a reference protein) or an ensemble of conformational states of the structures of a reference molecule.
[0035] The term "variant" as used herein refers to a specific change in a molecule (e.g., a protein or gene) relative to an established reference molecule. For example, a variant can be a change in a nucleic acid (e.g., a gene), protein, lipid, carbohydrate, etc., relative to a reference molecule (i.e., a corresponding molecule without the change). In some aspects, a variant is a genetic variant. As used herein, the term "variant molecule" refers to a molecule (e.g., a protein) that includes at least one variant (e.g., a missense variant). As used herein, "variant molecular structure," "variant molecular structure," and grammatical variations thereof refer to the three-dimensional structure of a variant molecule. In some, a variant molecular structure can refer to a single conformational state of the structure of a variant molecule (e.g., a variant protein), or an ensemble of conformational states of the structure of a variant molecule.
[0036] The term "gene" is used broadly to refer to any segment of DNA or other nucleic acid associated with a biological function (e.g., mRNA, miRNA, etc.). Thus, a gene includes a coding sequence and regulatory sequences necessary for its expression. A gene also optionally includes non-expressed nucleic acid segments that, for example, form recognition sequences for other proteins. Genes may be derived from a variety of sources, including cloning from a source of interest or synthesis from known or predicted sequence information, and may include sequences designed to have desired parameters.
[0037] The term "genetic variant" refers to any specific change in a nucleotide or nucleotide sequence relative to an established reference nucleotide or nucleotide sequence. Variations in genomic information are useful to identify because they may indicate the cause of phenotypic variation between subjects. One theory is that invariant regions of the genome of normal subjects are likely important for encoding essential components necessary for the development and survival of those subjects. Variants may explain normal phenotypic differences between subjects and / or variants may explain disease-associated variations. In some aspects, variants are mutants. Variants are polypeptides or polynucleotide sequences that have been modified by one or more substitutions, insertions, crossovers, deletions, and / or other genetic manipulations. For purposes of this disclosure, variants and variants are not limited to any particular method of generating them. In some aspects, mutated or variant sequences have increased, decreased, or substantially similar activity or characteristics compared to the parent sequence. In some aspects, variant polypeptides include one or more amino acid residues that are mutated compared to the amino acid sequence of a wild-type polypeptide (e.g., a parent polypeptide). In some embodiments, one or more amino acid residues of a polypeptide of a variant polypeptide comprising a plurality of polypeptides remain constant, non-variant, or non-mutated relative to the parent polypeptide, hi some embodiments, the parent polypeptide is used as a basis for generating variants with improved stability, activity, or any other desired property.
[0038] In some embodiments, the variant is a missense variant gene. As used herein, the term "missense variant gene" refers to a gene in which nucleotide substitution results in the replacement of one amino acid with another amino acid in the protein produced by the gene. The amino acid change can change the function of the protein. In some embodiments, the amino acid change can result in a pathogenic variant. In some embodiments, the amino acid change can result in a benign variant.
[0039] The terms "protein," "polypeptide," and "peptide" are used interchangeably and refer to any polymer of two or more amino acids covalently linked by amide bonds. In some cases, the polymer has at least about 30 amino acid residues, and generally at least about 50 amino acid residues. More typically, the polymer contains at least about 100 amino acid residues. The disclosure is not intended to be limited to any particular length of amino acid sequence. The term generally encompasses compositions that are considered to be fragments of full-length proteins or peptides. D- and L-amino acids, as well as mixtures of D- and L-amino acids, are also included in this definition. The polypeptides described herein are not limited to genetically encoded amino acids. Indeed, in addition to genetically encoded amino acids, the polypeptides described herein may be composed, in whole or in part, of naturally occurring amino acids and / or synthetic non-encoded amino acids. In some aspects, the polypeptides have amino acid sequences that include amino acid additions, or deletions (e.g., gaps), and / or substitutions, compared to the amino acid sequence of a full-length parent polypeptide, while retaining the functional activity of the parent polypeptide. In some aspects, the polypeptides may be modified through post-translational modifications (eg, methylation, phosphorylation, glycosylation).
[0040] In some aspects, the variant may be a missense variant protein. As used herein, the term "missense variant protein" refers to a protein encoded by a missense variant gene, which has an amino acid substitution with respect to a reference protein, e.g., a wild-type protein or a variant protein prevalent in a population. A missense variant protein has a different amino acid sequence with respect to a wild-type or other reference protein, and the change in the sequence of a protein may affect both its physical and / or functional properties. A single amino acid mutation in a protein may cause the protein to become inactive, which is a dramatic change in functional activity and may cause a fatal disease in a subject. In some aspects, the substitution in a missense variant protein may have a biological effect, e.g., causing a disease or disorder in a subject.
[0041] As used herein, the term "biological effect" includes the molecular effect of a given biological entity (e.g., a variant), where the variant is a protein (e.g., a missense variant protein), as well as the effect of such an entity at the level of a cell, tissue, or organism. Biological effects encompass, for example, effects at the molecular level, as well as physiological effects. Thus, in some aspects, a change in the molecular structure (e.g., three-dimensional structure) of a protein, or a change in binding affinity for its ligand, is considered a biological effect. In some aspects, the gain or loss of function of a protein can be considered a biological effect. In some aspects, the onset of, for example, a disease or disorder due to the gain or loss of function of a protein, or protein aggregation due to protein misfolding, can be considered a biological effect. In some aspects, an increase or decrease in the risk of developing a disease or disorder can be considered a biological effect. In some aspects, the manifestation of a pathological phenotype can be considered a biological effect. In some aspects, the terms "biological effect" and "clinical significance" can be used interchangeably.
[0042] Generating a "conformational dynamics" of a reference molecule (e.g., a reference protein) involves using computational algorithms to determine conformational changes in the molecular structure of a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein). In some aspects, these algorithms are referred to as molecular dynamics (MD) simulations, in which a reference molecular system is simulated over time by completing computational cycles that involve (1) calculating physical interaction energies between atoms, and (2) substituting atoms according to these interaction energies. These simulation cycles, also known as steps, generate a trajectory of the reference molecule (e.g., a reference protein) or variant molecule (e.g., a variant protein) over time, characterizing the conformational dynamics of the reference molecule or variant molecule. As used herein, the term "trajectory" refers to a set of data obtained from a conformational dynamics simulation, including the distinct conformational states that the reference molecule or variant molecule adopts during a simulation over a particular period of time.
[0043] In some embodiments, at least one (1) independent trajectory can characterize the conformational dynamics of the reference molecule or the variant molecule. In some embodiments, the MD simulation is performed using GROMACS, OpenMM, NAMD, Amber, or LAMMPS. In some embodiments, the sampling of the MD simulation can be improved by using advanced sampling (AS) methods, such as umbrella sampling, parallel tempering, replica exchange, metadynamics, and simulated annealing. In some embodiments, deep learning algorithms can be used to predict the conformational dynamics of the reference molecule (e.g., reference protein) or variant molecule (e.g., variant protein).
[0044] In some embodiments, multiple conformational states of the same reference molecule (e.g., a reference protein) or variant molecule (e.g., a variant protein) or their respective molecular structures can be used to infer the conformational dynamics of the reference molecule (e.g., a reference protein) or variant molecule (e.g., a variant protein). The conformational dynamics of the reference molecule (e.g., a reference protein) or variant molecule (e.g., a variant protein) can include, for example, about 10,000 to about 1,000,000 or more conformational states of the reference molecule (e.g., a reference protein) or variant molecule (e.g., a variant protein). The conformational dynamics of the reference molecular structure (e.g., a reference protein structure) or variant molecular structure (e.g., a variant protein structure) can include a generative model that can be used to generate conformational states according to a distribution representing the conformational dynamics.
[0045] In the context of this disclosure, the term "reduced-dimensional conformational landscape" refers to multiple conformational states of the conformational dynamics of a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein) resulting from the application of dimensionality reduction. Generating a reduced-dimensional conformational landscape of a molecule's conformational dynamics (e.g., a reference molecule or a variant molecule, e.g., a reference protein or a variant protein) may include, for example, (1) applying a molecular featurerization method to the conformational states that define the conformational dynamics of the molecule, and (2) using an unsupervised dimensionality reduction method on these characterized conformational states to further reduce the number of dimensions.
[0046] The low-dimensional conformational landscape can reveal relevant conformational dynamics involved in the function and / or behavior of components (e.g., atoms) of a reference or variant molecule (e.g., a reference or variant protein). The conformational landscape of a reference or variant molecule (e.g., a reference or variant protein) can be represented, for example, as a discrete distribution of individual conformational states from a trajectory of conformational dynamics in a low-dimensional space. The conformational landscape can also be represented as a continuous distribution in a low-dimensional space. The density of the distribution reflects the tendency of the reference or variant molecule (e.g., a reference or variant protein) to occupy certain regions in the conformational landscape, and in some aspects, this density correlates with the overall energy of the reference or variant molecule (e.g., a reference or variant protein), such that high-density regions of the conformational landscape reflect low-energy conformations. The conformational dynamics of a variant molecule (e.g., a variant protein) that contains a pathogenic molecular alteration (variant or mutation) may differ from the conformational dynamics of a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein) that has a benign molecular alteration (variant or mutation).
[0047] The methods disclosed herein include a feature quantification or feature extraction step. As used herein, "feature quantification", "feature extraction", and grammatical variations thereof, refer to methods used to calculate structural features of molecules, e.g., reference molecules of the present disclosure (e.g., reference proteins). Structural features of molecules may include, for example, angles, orientations, distances, interactions, dispersion, fluctuations, energy, solvent accessibility or lack thereof, or any combination thereof, as disclosed herein. Structural features may be defined by individual atoms or groups of atoms according to structural elements (e.g., spines, side chains, residues, motifs, secondary structure elements, or domains). In some aspects, molecular feature quantification results in at least a 10-fold, 20-fold, 30-fold, 40-fold, or 50-fold reduction in the number of features that define the molecular structure (e.g., reference molecules or variant molecules of the present disclosure). The feature quantification process does not require a human operator (i.e., the process is automated), and specific feature quantification is selected based on downstream model performance.
[0048] An example of the rationale behind the selection of a particular feature (e.g., backbone ψ angle) for dimensionality reduction is shown in FIG. 5. This shows that there is a different distribution of backbone ψ angles at a particular position (F650) in pathogenic and benign variants of a molecule (e.g., variants containing mutations at positions P349L, L341P, etc.). Thus, there is a relationship between the feature (backbone ψ angle distribution) and the pathogenicity of the variant molecule. Thus, the backbone ψ angle is a structural feature that can be extracted for dimensionality reduction. The method disclosed herein also includes a dimensionality reduction step. This dimensionality reduction step also does not require a human operator. Thus, the dimensionality reduction step is unsupervised. Dimensionality reduction can be used to further reduce the number of global structural features extracted from the conformational dynamics of a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein) to generate a low-dimensional representation of the molecule. As used herein, the term "dimensionality reduction" includes feature selection (e.g., selecting features based on feature correlation, variance, dispersion, or any combination thereof), linear decomposition (e.g., principal component analysis, PCA, and / or singular value decomposition, SVD), non-linear decomposition (e.g., manifold learning (e.g., UMAP), isomap, kernel PCA, t-distributed stochastic neighbor embedding (t-SNE), and multidimensional scaling (MDS)), neural network embedding (e.g., using artificial neural networks (ANN) to generate a low-dimensional representation of the input data, such as through the use of (variational) autoencoders or graph embedding), or any combination thereof. Unsupervised dimensionality reduction may result in a further reduction of features by at least 50x, 60x, 70x, 80x, 90x, or 100x.
[0049] FIG. 3 shows a schematic diagram of the dimensionality reduction portion of the disclosed workflow. It includes (i) a structural feature extraction step, in which certain original features or characteristics are selected, and (ii) a dimensionality reduction step, in which the selected original features or characteristics are transformed into new features or characteristics using methods such as PCA or autoencoders. As a result, a significant reduction in the dimensionality of the conformational dynamics can be achieved. For example, a system that includes multiple trajectories and has 10,000-100,000 degrees of freedom can be subjected to a structural feature extraction step as disclosed herein (e.g., extracting features such as backbone angles, residue pairing, distances, residue variation, or solvent accessibility, to name a few) to achieve, for example, an approximately 50-fold reduction in the number of features. After structural feature extraction, the trajectories can be subjected to a dimensionality reduction step as disclosed herein (e.g., using PCA, autoencoders, etc.) to obtain a set of trajectories with a further reduced number of features (e.g., approximately 100-fold reduction).
[0050] The methods disclosed herein further include determining a metastable conformational state or a set thereof from the low-dimensional conformational landscape of the reference molecule (e.g., a reference protein). The process of determining such metastable conformational states generally includes: (1) defining a set of conformational clusters in the low-dimensional conformational landscape of the reference molecule (e.g., a protein) using a clustering technique (e.g., (a) defining a set of conformational states in the low-dimensional space by projecting conformational states from the conformational dynamics of the reference protein structure into the low-dimensional space, and (b) using the projected conformational states as input to a clustering algorithm to define the conformational clusters); (2) selecting a subset of conformational clusters from the set of all clusters; and (3) defining a representative conformational state from each selected conformational cluster as a metastable conformational state for further analysis. A "conformational cluster" is a collection of conformational states that are close to each other in a low-dimensional conformational landscape (eg, a region in a low-dimensional conformational landscape).
[0051] The disclosed methods may also include selecting or prioritizing a subset of conformational clusters based on cluster properties, such as distance from the conformational landscape centroid, cluster occupancy, cluster distribution statistics, structural properties of the clusters, thermodynamics of the clusters, or any combination thereof. The structural properties may be selected from structural characterizations used for dimensionality reduction of any number of structural components of the molecular system.
[0052] In some embodiments, the term "metastable conformational state" refers to a conformational state sampled in a conformational cluster that is closest or one of the closest to the center of gravity of all conformational states of the cluster in the low-dimensional representation. In the methods disclosed herein, a metastable conformational state may be identified for each selected conformational cluster. In some embodiments, a metastable conformational state may also be defined as (i) a conformational state in a region of the conformational cluster generated by a generative model, such as a decoder network of an autoencoder, or a state interpolated from nearby conformational states in the conformational cluster, (ii) a local minimum in the conformational landscape, or (iii) a conformational state in which the molecular component of interest differs from other metastable conformational states by at least about 2 Å backbone atom root mean square deviation (RMSD).
[0053] FIG. 4 shows a schematic diagram of the conformational energy states of the reference molecule and the metastable conformational energy states of the variant molecule. This representation illustrates that, as discussed above, it is possible to simplify or compress the conformational space through trajectory featurization (reducing the degrees of freedom of the conformational dynamics by selecting certain features) and dimensionality reduction (e.g., converting selected features to other features via PCA) to generate simplified conformational dynamics while still distinguishing between the reference and variant molecules. The simplified representation can then be used as input to generate a machine learning model, as described below. FIG. 6 shows an exemplary low-dimensional conformational landscape representation obtained from clustered structural conformations of MutS homolog 2 protein (MSH2), which is associated with diseases such as Lynch syndrome I or mismatch repair cancer syndrome 2. Each of the clusters shown in FIG. 6 represents a metastable conformational state.
[0054] Conformational dynamics in low-dimensional representations can distinguish between pathogenic and benign variants (e.g., missense genetic variants), for example, as shown in Figure 7, and therefore can be effectively used as features in predictive machine learning models. Constructing a variant molecule (e.g., variant protein) or set thereof can include introducing at least one variant (e.g., mutation) into a reference molecule (e.g., reference protein) in a metastable conformational state or set thereof. This process may include, for example, (1) defining a variant or set thereof (e.g., a mutation or set of mutations) for constructing a variant molecule, e.g., a set of variants that are or may be observed in a medical genetic test; (2) introducing the variant (e.g., a mutation) into a relevant molecular component of a molecular system (as disclosed above, a molecular system may include, for example, a reference molecule, such as a protein, and additional components, such as a solvent or a ligand); and (3) performing a steric and / or energy-based minimization of the variant (e.g., a mutation in the reference molecule that would be considered a variant molecule after introduction of the variant) to mitigate abrupt conformational changes in the molecular system.
[0055] The set of variants may include (i) a set of variants with known clinical significance that are used as training labels for the predictive model, and (ii) a set of variants whose clinical significance is unknown, but for which further variant effect information is desired. The set of variants should include variants known to change the composition of one or more molecular components of the system (e.g., missense variants known to change an amino acid residue at a specific position in a protein). The set of variants should include at least one variant known to cause a pathogenic effect on the molecular system and one variant known to cause a non-pathogenic (benign) effect on the molecular system (which itself may be a reference molecular system).
[0056] A variant (e.g., a mutation) may, for example, alter the identity, properties, interactions, or position of atoms in a molecular structure. A variant may, for example, be a mutation in an amino acid side chain or a nucleobase. A variant may be directly encoded by a variant gene (e.g., a substitution, insertion, deletion, missense mutation, or nonsense mutation) or may be indirect, for example, resulting in a change in the composition of a molecular system (e.g., an increase or decrease in the concentration of a molecular component, or a difference in the chemical composition of a molecular component) or a non-genetic modification of a molecular component (e.g., a post-translational modification such as phosphorylation or glycosylation).
[0057] In some aspects, introducing a variant (e.g., a missense variant) into a reference molecular structure to generate a variant molecular structure may include some or all of the following steps: (1) removing specific amino acid side chain atoms and interactions at the variant position from a reference molecule (e.g., a reference protein); (2) determining the original positions of the backbone φ-ψ angles; (3) iteratively identifying possible rotamer conformations of the missense amino acid using a backbone-dependent rotamer library, calculating the steric clashes caused by this rotamer placement at the variant position, and repeating until the steric clashes are below a threshold; and (4) completing relaxation to further reduce the steric and energetic costs associated with introducing the variant.
[0058] Determining the conformational dynamics of the variant molecular structure in a metastable conformational state can include using computational methods to sample the conformational changes that occur after the introduction of the variant. These conformational changes can be determined, for example, using molecular dynamics (MD) simulations.
[0059] The predictive models of the present disclosure can be generated using machine learning (ML) methods such as, for example, linear or nonlinear regression, decision trees (e.g., random forests), or artificial neural networks (ANNs). Clinical data can be used as training labels, for example, biomarker status (e.g., the presence or absence or expression level of a particular biomarker), biometric data, lifestyle-related data, response to treatment, symptoms of a disease or condition, protein expression data, type of treatment administered, dosage, dosage regimen, route of administration, presence or absence of concomitant therapy, response to therapy, age, weight, sex, ethnicity, data from NCBI ClinVar database submissions, INVITAE™ clinical interpretations, other experimental or computational models, or any combination thereof. For example, the ANN may be trained on a training set that includes variant protein structures obtained after determining the conformational dynamics of metastable conformational states of variant molecules. The training set can include the structural changes caused by the introduction of the variant and a label (training label) applied to each variant protein structure that corresponds, for example, to the observed clinical characteristic (e.g., the likelihood or severity of a disease in the presence of a particular variant).
[0060] Predicting variant effects may involve training a predictive model using conformational dynamics of metastable conformational states of variant molecules and on training labels such as those described above applied to such variant molecules, and using the trained model to predict the clinical significance of other variants of unknown significance.
[0061] Predictive models can be extended to newly observed variants from genetic tests in a process that includes, for example, (1) constructing new variant molecules (e.g., variant proteins) from stored metastable conformational states; (2) generating conformational dynamics of the new variant molecules; (3) determining low-dimensional conformational dynamics of the new variant molecules using a previously computed unsupervised dimensionality reduction model; and (4) predicting the clinical significance of the newly observed variants using the previously trained predictive model.
[0062] I. High-throughput prediction of variant effects from conformational dynamics This disclosure describes novel computational methods that allow characterization of the functional consequences, i.e., biological effects, of variants (e.g., missense gene and protein variants). These methods integrate conformational dynamics generation and machine learning (ML) to identify and quantify structural features of molecular systems (e.g., genes or proteins) and their variants (e.g., missense variants) to obtain measurements or predictions of the effects of variants (e.g., mutations) and their potential disruption of biological function. Identification of functional disruptions (i.e., biological effects) with respect to a reference molecular system (e.g., a wild-type protein or gene, or common allelic forms), and their severity, can then be used, for example, to treat diseases and disorders associated with such functional disruptions.
[0063] In some aspects, the disclosure provides a computer-implemented method for high-throughput prediction of a biological effect of a variant (e.g., a missense variant), the method comprising: (a) generating (e.g., from molecular dynamics) a conformational dynamics of a structure of a reference molecule (e.g., a wild-type protein or gene, or a common allelic form); (b) defining a low-dimensional conformational landscape of a reference molecular system that includes a reference molecule (e.g., a reference protein) using an unsupervised dimensionality reduction method; (c) determining a metastable conformational state or a set thereof from the low-dimensional conformational landscape of a reference molecular system that includes a reference molecule (e.g., a reference protein); and (d) introducing in silico at least one variant (e.g., mutation) into a metastable conformational state of a molecular component in a reference molecular system, e.g., a reference molecule (e.g., a reference protein), or a set thereof, thereby generating, e.g., a variant molecular system including a variant molecule (e.g., a variant protein) or a set thereof; (e) determining the conformational dynamics of the variant molecular system or set thereof in a metastable conformational state or set thereof; (f) generating a low-dimensional representation of the conformational dynamics of the variant molecular system or set thereof in a metastable conformational state or set thereof using an unsupervised dimensionality reduction method; (g) training a predictive model (e.g., an ANN) using the low-dimensional representation of the conformational dynamics of the variant system or set thereof and the biological or clinical data as training labels, where the predictive model predicts the biological effect of the variant on a molecular component (e.g., a protein) of a reference molecular system (e.g., a reference protein); and optionally (h) rapidly extending the predictive model to newly observed variants from medical genetic testing (e.g., introducing new variants for evaluation in step (d) of the workflow).
[0064] Since the different steps of the workflow are precomputed after the first iteration, the evaluation of additional variants will occur very quickly.
[0065] Generally, the methods described herein include one or more of the steps described above. In some aspects, the order of the steps described above may vary, e.g., generating (e.g., using MD) conformational dynamics of a variant molecule (e.g., a variant protein) or set thereof may be performed before or simultaneously with simulating the structural conformation of a reference molecule (e.g., a reference protein). In other aspects, certain groups of steps may be repeated in a loop. For example, steps including generating a variant molecular system (e.g., a molecular system including a variant protein structure) or set thereof in silico (e.g., a molecular system including a reference molecule (e.g., a reference protein such as a wild-type protein or a common allelic form) by introducing at least one variant (e.g., a missense mutation) into a metastable conformational state of a reference molecular system, and determining the conformational dynamics of the variant system or set thereof (i.e., simulating the structural conformation), e.g., by applying molecular dynamics (MD), may be applied sequentially or simultaneously to different batches or families of variants.
[0066] In some aspects, the methods described herein include (a) generating conformational dynamics of a reference protein; (b) defining a low-dimensional conformational landscape of the reference protein structure using an unsupervised dimensionality reduction method; (c) determining a metastable conformational state or set thereof from the low-dimensional conformational landscape of the reference protein structure; (d) constructing a variant protein or set thereof by introducing in silico at least one variant (e.g., mutation) into the metastable conformational state or set thereof of the reference protein structure, thereby generating a variant protein structure or set thereof; and (e) determining a metastable conformational state or set thereof from the low-dimensional conformational landscape of the reference protein structure. (f) using an unsupervised dimensionality reduction method to generate a low-dimensional representation of the conformational dynamics of the variant protein or set in a metastable conformational state or set, (g) using the low-dimensional representation of the conformational dynamics of the variant protein or set using biological or clinical data as training labels to train a predictive model, where the model predicts the biological effect of the variant relative to the reference protein, and optionally (h) rapidly extending the predictive model to newly observed variants from medical genetic testing. In some aspects, new variants for evaluation may be introduced into the pipeline in step (d).
[0067] A schematic diagram of an exemplary workflow of the present disclosure is shown in Figures 1 and 2. The diagram in Figure 1 shows the main modules (conformational dynamics generation, unsupervised dimensionality reduction, identification of metastable conformational states, variant construction, model extension, and predictive modeling; white background) and the data objects that are input and output of the various modules (reference molecule, conformational dynamics of reference molecule, low-dimensional conformational landscape, metastable conformational states, variant list, variant molecule, conformational dynamics of variant molecule, low-dimensional conformational dynamics of variant molecule, newly observed variants from genetic testing, clinical training data of variants, and predicted variant effects; grey background).
[0068] The prediction system for predicting variant effects from molecular conformational dynamics disclosed herein includes the following steps: (a) generating conformational dynamics of a reference molecule; (b) defining a low-dimensional conformational landscape of the reference molecule using an unsupervised dimensionality reduction method; (c) determining a metastable conformational state or a set of metastable conformational states from the low-dimensional conformational landscape of the reference molecule; (d) constructing a variant molecule or a set of variants by introducing at least one molecular change to a molecular component of the reference molecule in the metastable conformational state or set of metastable conformational states; and (e) determining a variant effect in the metastable conformational state or set of metastable conformational states. (f) using an unsupervised dimensionality reduction method to generate a low-dimensional representation of the conformational dynamics of the variant molecule or set thereof in a metastable conformational state or set thereof, (g) using the low-dimensional representation of the conformational dynamics of the variant molecule or set thereof to train a predictive model using biological or clinical data as training labels, where the model predicts the biological effect of the variant on the molecular components of the reference molecule, and optionally (h) rapidly extending the predictive model to newly observed variants from medical genetic testing. See FIG. 2.
[0069] Generation of IA conformational dynamics The methods disclosed herein use conformational dynamics of molecular structures. Conformational dynamics can be generated, for example, using computational methods (e.g., molecular dynamics (MD) simulations) to study the motion of atoms and molecules in systems that include reference or variant molecules of the present disclosure. Conformational dynamics describes the complex time-dependent behavior of atoms and molecules in systems that may include large molecules with highly dynamic structures, such as proteins and nucleic acids. Such dynamics include, for example, molecular stability, molecular folding, conformational changes of molecules, and molecular binding. Thus, in the case of proteins, such dynamics include, for example, protein stability, protein folding, conformational changes of proteins, and molecular binding, including receptor / ligand binding. These studies can provide useful information for designing drugs for the treatment of various diseases, including, for example, cancer and genetic diseases. They also provide information for predicting the severity of disruption of a variant (e.g., mutation) or other such molecule's (e.g., macromolecule, such as a protein (e.g., a reference protein or a variant protein, such as a missense variant protein of the disclosure)) biological function.
[0070] The study of conformational dynamics of molecules such as proteins is complex and requires computational and in silico studies of macromolecular structures and macromolecular dynamics to obtain accurate information. In some aspects, the methods disclosed herein can use all-atom conformational dynamics simulations (e.g., MD simulations) performed using any software, any parameter set or force field used to calculate potential energy, and any protein / ligand interaction system. The parameter set includes a set of rules that control the behavior of individual atoms in the system. Molecular structures generated by conformational dynamics simulations (e.g., MD simulations) can be characterized by several metrics to compare the effect of the variant (e.g., mutation) on the conformational dynamics of the reference molecule or variant molecule of the present disclosure, such as the overall structural features, the dynamics of subdomains in the structure, energy interactions, and the statistical features of the overall ensemble of structures.
[0071] In addition to being computationally intensive, conformational dynamics (e.g., from MD simulations) are high-dimensional because they include the motion of all atoms in a molecule in three dimensions over time. Because molecules often have 10,000 or more atoms, analysis of these high-dimensional systems typically requires expert analysis to determine features of interest. The need to use human experts to determine features of interest creates a bottleneck that prevents the application of these methods in a high-throughput mode. Thus, there is a need to develop methods that can extract relevant structural features and reduce the dimensionality of conformational dynamics data (e.g., MD data) in an automated manner (i.e., without human intervention) in order to use conformational dynamics data (e.g., MD data) for high-throughput variant classification, for example, using machine learning.
[0072] In some aspects, simulating a structural conformation of a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein) using conformational dynamics includes (i) obtaining a three-dimensional structure of the reference molecule (e.g., a reference protein) or the variant molecule (e.g., a variant protein); and (ii) generating conformational dynamics (e.g., from MD) using the obtained three-dimensional structure of the reference molecule (e.g., a reference protein) or the variant molecule (e.g., a variant protein). In some aspects, the three-dimensional molecular structure is experimentally determined (e.g., by crystallography, electron microscopy, or nuclear magnetic resonance). In some aspects, the three-dimensional molecular structure is obtained from an experimental three-dimensional protein structure repository, such as the PDB.
[0073] Knowing the structure of a molecule (e.g., a protein) provides an important resource for understanding how it functions, but experiments to determine the structure of a molecule can take months or more, and some prove intractable. For this reason, researchers have developed computational methods to predict molecular structures. For example, researchers have developed computational methods to predict protein structures from their primary amino acid sequences. When the three-dimensional structures of similar proteins have already been experimentally determined, for example, using X-ray crystallography, algorithms based on "template modeling" can provide accurate predictions of protein structures. Thus, in some embodiments, the three-dimensional structure is a template model structure.
[0074] In some embodiments, the three-dimensional molecular structure corresponds to a molecule (e.g., a protein) for which an experimentally determined molecular structure is not available. Thus, in some embodiments, the three-dimensional molecular structure is obtained via computational modeling. In some embodiments, the free modeled three-dimensional molecular structure is obtained from a database with computationally derived structures, such as modbase, Pfam, EMDB, or EVFold. In some embodiments, the three-dimensional molecular structure is obtained by prediction using existing computational structure prediction software, such as AlphaFold or trRosetta. See, for example, Senior et al. (2020) Nature 577:706-710 and Yang et al. (2020) PNAS 117:1496-1503, both of which are incorporated herein by reference in their entirety. In some embodiments, the three-dimensional molecular structure is determined by a combination of experimental and computational methods.
[0075] In some aspects, the three-dimensional molecular structures used in the methods disclosed herein are selected according to their resolution. In some aspects, the three-dimensional molecular structures used in the methods disclosed herein are high-resolution molecular structures.
[0076] As used herein, the term "resolution" refers to a measure of the quality of data collected for an experimentally derived molecular structure (e.g., the structure of a reference molecule or variant molecule of the present disclosure). For example, in the case of a molecular structure obtained using X-ray crystallography, if all the proteins in the diffracted crystal are aligned in the same way to form a very perfect crystal, all the proteins will scatter the X-rays in the same way and the diffraction pattern will show the fine details of the crystallized protein. On the other hand, if the proteins in the crystal are all slightly different, due to local flexibility or motion, the diffraction pattern will not contain as much fine information. Resolution is thus a measure of the level of detail present in the diffraction pattern and the level of detail that will be seen when the electron density map is calculated. "High resolution" as applied to crystal structures refers to structures with a resolution value on the order of 1 Å, that are highly ordered and all the atoms in the electron density map can be easily seen. "Low resolution" as applied to crystal structures refers to structures with a resolution of 3 Å or higher that show only the basic outline of the protein chain and the atomic structure must be inferred. Most of the crystallographically defined structures of proteins lie between these two extremes. As a general rule of thumb, molecular structures with smaller resolution values, i.e. "high resolution structures," have higher confidence in the positions of atoms.
[0077] In some embodiments, the term "high resolution" refers to a resolution of 3.0 Å or less, for example, in an experimentally derived structure, for example, a crystal structure of a reference molecule (e.g., a reference protein). Molecular structures of any resolution, such as 6.0 Å, 5.0 Å, 4.0 Å, 3.0 Å, 2.0 Å, or less, can be used in the disclosed methods. Thus, in some embodiments, the resolution is about 1 Å, about 1.5 Å, about 2 Å, about 2.5 Å, about 3 Å, about 3.5 Å, about 4 Å, about 4.5 Å, about 5 Å, about 5.5 Å, or about 6 Å. In some embodiments, the resolution is better (higher) than about 6 Å, better than about 5 Å, better than about 4 Å, better than about 3 Å, better than about 2 Å, or better than about 1 Å. In some embodiments, the resolution is about 1 Å to about 2 Å, about 2 Å to about 3 Å, about 3 Å to about 4 Å, about 4 Å to about 5 Å, about 5 Å to about 6 Å, about 1 Å to about 3 Å, about 2 Å to about 4 Å, about 3 Å to about 5 Å, about 4 Å to about 6 Å, about 1 Å to about 4 Å, about 2 Å to about 5 Å, about 3 Å to about 6 Å, about 1 Å to about 5 Å, about 2 Å to about 6 Å, or about 1 Å to about 6 Å.
[0078] In some embodiments, the three-dimensional structure of the molecule used in the methods disclosed herein is selected according to the energetics of conformational dynamics, for example, from MD simulation. Resolution is a consideration for experimentally derived structures, but does not apply to computationally predicted structures. In addition to using structure and prediction quality metrics (e.g., resolution) as a criterion for selecting targets, conformational dynamics, for example, from MD simulation, can be used to evaluate whether a reference molecular structure (e.g., a three-dimensional structure of a reference protein) is suitable for constructing a structural model of a variant molecule. Conformation dynamics (e.g., from MD simulation of a protein structure) can be evaluated for changes in secondary, tertiary, and in the case of complexes, quaternary structure, and / or changes in solvent-accessible surface area to determine whether the conformational dynamics of a reference molecule (e.g., a reference protein) can be accurately calculated.
[0079] The conformational dynamics of a reference or variant molecule (e.g., a reference or variant protein) are determined by applying computational methods (e.g., MD simulations) to the three-dimensional structure of the reference or variant molecule (e.g., a reference or variant protein). MD is a computer simulation method for analyzing the physical motions of atoms and molecules. The atoms and molecules are allowed to interact for a period of time, providing a picture of the dynamic "evolution" of the system. In the most common version, the trajectories of the atoms and molecules are determined by numerically solving Newton's equations of motion for a system of interacting particles, and the forces between the particles and their potential energies are often calculated using interatomic potentials or molecular mechanics force fields. Because molecular systems are typically composed of a vast number of particles, it is not possible to analytically determine the properties of such complex systems, and MD simulations circumvent this problem by using numerical methods.
[0080] Other methods of generating conformational dynamics other than MD can be used to determine the conformational landscape of a protein (www.sciencedirect.com / science / article / pii / S0969212619301145). The methods do not rely on simulation of conformational dynamics (e.g., MD simulation) per se, but have a collection of distinct conformational states that represent the conformational landscape. Such methods incorporating deep learning can be used to determine the conformational landscape of a molecule instead of traditional MD. Normal mode analysis (pubs.acs.org / doi / 10.1021 / cr900095e) can also be used to calculate the conformational states of a protein, similar to traditional MD.
[0081] The simulation size, time step, and total time must be chosen so that the calculations can be completed within a reasonable time. However, the conformational simulation needs to be long enough to be appropriate for the time scale of the natural process being studied. To make statistically valid conclusions from a conformational simulation, the simulated time span needs to match the dynamics of the natural process. For example, for protein dynamics, data from simulations spanning nanoseconds (10<+-><+9> seconds) to microseconds (10<+-><+6> seconds) are used. To obtain these conformational simulations, several CPU days to CPU years are required. Parallel algorithms allow the load to be distributed among CPUs, an example is the space partitioning algorithm or the force decomposition algorithm. During classical MD simulations, the most CPU-intensive task is the evaluation of the potential as a function of the internal coordinates of the particles. Another factor that affects the total CPU time required for a simulation is the size of the integration time step. This is the length of time between evaluations of the potential. The time step must be chosen to be small enough (i.e., smaller than the period associated with the fastest vibrational frequency in the system) to avoid discretization errors. A typical time step in classical MD is on the order of 1 femtosecond (10<+-><+1><+5> seconds).
[0082] Given the CPU time requirements for generating the conformational dynamics of a single protein, MD-based methods are not suitable for high-throughput analysis. However, as disclosed in detail below, the presently disclosed methods can reduce the dimensionality of the data used in MD simulations and dramatically reduce the CPU time, thus enabling the application of conformational dynamics from, for example, MD simulations in a high-throughput mode.
[0083] In some embodiments, for example, conformational dynamics from MD simulations are generated using the GROMACS package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the NAMD package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the OpenMM package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the AMBER package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the LAMMPS package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the CHARMM package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the Abalone package / program. In some embodiments, for example, conformational dynamics from MD simulations are generated using the TINKER package / program.
[0084] In some embodiments, for example, conformational dynamics from MD simulations are generated using the GROMOS force field. In some embodiments, for example, conformational dynamics from MD simulations are generated using the OPLS (Optimized Potentials for Liquid Simulations) force field. In some embodiments, for example, conformational dynamics from MD simulations are generated using the AMBER (Assisted Model Building with Energy Refinement) force field. In some embodiments, the AMBER force field is ff14SB. In some embodiments, for example, conformational dynamics from MD simulations are performed using the CHARMM (Chemistry at Harvard Macromolecular Mechanics) force field. In some embodiments, the CHARMM force field is CHARMM19 (bonded atoms, sometimes called extended atoms), CHARMM22 (all atoms), CHARMM22 / CMAP (a dihedral angle potential corrected variant of CHARMM22), CHARMM27, CHARMM36, CHARMM36m, or CHARMM36IDPSFF. In some embodiments, for example, conformational dynamics from MD simulations are performed using the MMFF force field (Merck molecular force field).
[0085] In some embodiments, the reference or variant molecule (e.g., a reference protein or a variant protein) may be coarse-grained, such as to reduce the number of atoms (and therefore the computation and time) required to simulate the reference or variant molecule (e.g., a reference protein or a variant protein) using the MARTINI force field. By coarse-graining the reference or variant molecule (e.g., a reference protein or a variant protein), the simulation time of the reference or variant molecule (e.g., a reference protein or a variant protein) can be reduced.
[0086] In some embodiments, solvent molecules are required to simulate a reference molecule or a variant molecule (e.g., a reference protein or a variant protein) in a solvent environment. In some embodiments, these solvent molecules are explicit water molecules (such as TIP3P). In some embodiments, these solvent molecules include ionic molecules (such as Na<++> or Cl<+->). In some embodiments, these solvent molecules are either lipids, all atoms, or coarse-grained atoms. In some embodiments, these solvent molecules can be represented by an implicit solvent model to reduce the number of atoms in the simulation.
[0087] In one particular embodiment, the CHARMM36 force field with explicit TIP3P solvent is used. Minimization simulations of the target system are performed for over 5,000 simulation steps with backbone atoms constrained, followed by equilibration from 125 ps to a temperature of 303 K, followed by a production simulation sampled every 2 femtoseconds using a particle mesh Ewald electrostatic range of 1.2 nm radius and a 1.2 nm radius cutoff for Van der Waals interactions.
[0088] While parallelizing the simulation, reducing its complexity using coarse-graining or implicit solvent, and improving sampling using advanced sampling methods reduces the required simulation time, simulating each variant molecule with the same amount of sampling as the reference protein is still too computationally, time- and cost-intensive for use in a high-throughput clinical variant interpretation platform. However, as described herein, the total amount of simulation time for the variants can be reduced by using conformational simulation of the reference molecule to generate a low-dimensional representation of the conformational landscape and clustering the conformations in the landscape to identify metastable conformational states for simulating the variant molecules. In one particular embodiment, the reference protein was simulated with a simulation time of about 5 microseconds, requiring a total of more than 300 hours of computation time. In comparison, for each variant protein simulation, only 100 nanoseconds of simulation time (7 hours of computation time) was required to generate a high-performance model, i.e., by selecting the pre-computed metastable conformational states identified as described herein, the total computation time was reduced to about 2% of the original time. In other words, the present method can perform conformational dynamics simulations for over 40 variant proteins while traditional conformational dynamics approaches calculate conformational simulations for a single variant protein. It is important to emphasize that these simulation time reductions are a result of using a low-dimensional representation of the conformational landscape and clustering of conformations in the landscape to identify metastable conformational states, and are applicable to any simulation approach independent of the simulation package / program and / or advanced sampling methods used.
[0089] In some embodiments, each of the generation of conformational dynamics steps (e.g., via MD simulation) in the methods disclosed herein independently generates at least about 1,000, at least about 5,000, at least about 10,000, at least about 20,000, at least about 30,000, at least about 40,000, at least about 50,000, at least about 60,000 conformational dynamics data for each molecule (e.g., a reference molecule such as a reference protein or any variant thereof). , at least about 70,000, at least about 80,000, at least about 90,000, at least about 100,000, at least about 200,000, at least about 300,000, at least about 400,000, at least about 500,000, at least about 600,000, at least about 700,000, at least about 800,000, at least about 900,000, or at least about 1,000,000 simulated structural conformations. In some aspects, each of the generation of conformational dynamics steps (e.g., via MD simulation) in the methods disclosed herein includes generating, independently, about 1,000, about 5,000, about 10,000, about 20,000, about 30,000, about 40,000, about 50,000, about 60,000, about 70,000, about 80,000, about 90,000, about 100,000, about 200,000, about 300,000, about 400,000, about 500,000, about 600,000, about 700,000, about 800,000, about 900,000, or about 1,000,000 simulated structural conformations for each molecule (e.g., a reference molecule such as a reference protein or any variant thereof).In some embodiments, each of the generation of conformational dynamics steps (e.g., via MD simulation) in the methods disclosed herein independently generates from about 1,000 to about 50,000, from about 50,000 to about 100,000, from about 100,000 to about 150,000, from about 150,000 to about 200,000, from about 200,000 to about 250,000, for each molecule (e.g., a reference molecule such as a reference protein or any variant thereof). , about 250,000 to about 300,000, about 300,000 to about 350,000, about 350,000 to about 400,000, about 450,000 to about 500,000, about 500,000 to about 600,000, about 600,000 to about 700,000, about 700,000 to about 800,000, about 800,000 to about 900,000, or about 900,000 to about 1,000,000 simulated structural conformations. In some embodiments, each of the generation of conformational dynamics steps (e.g., via MD simulation) in the methods disclosed herein includes generating at least 1,000,000 simulated structural conformations for each molecule (e.g., a reference molecule such as a reference protein or any variant thereof) independently. In some embodiments, the term "simulation" refers to all simulated structural conformations for a given protein. "Conformational dynamics," "simulated structural conformation," "conformational simulation," and grammatical variations thereof are used interchangeably and refer to the three-dimensional conformation of a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein) obtained by application of a method to generate conformational dynamics (e.g., molecular dynamics (MD)).
[0090] In some embodiments, each of the conformational dynamics generation steps (e.g., via MD simulation) in the methods disclosed herein, e.g., steps (a) and (e) of the workflow above, independently comprises about 1,000, about 5,000, about 10,000, about 20,000, about 30,000, about 40,000, about 50,000, about 60,000, about 70,000, about In some embodiments, the method comprises generating about 80,000, about 90,000, about 100,000, about 150,000, about 200,000, about 250,000, about 300,000, about 350,000, about 400,000, about 450,000, about 500,000, about 600,000, about 700,000, about 800,000, about 900,000, or about 1,000,000 simulated structural conformations. In some embodiments, each of the conformational dynamics generation steps (e.g., via MD simulation) in the methods disclosed herein, e.g., steps (a) and (e) of the workflow described above, can independently generate a conformational dynamics (D) sequence of about 1,000 to about 5,000, about 5,000 to about 10,000, about 10,000 to about 20,000, about 20,000 to about 50,000, about 50,000 to about 100,000, about 100,000 to about 200,000, about 200,000 to about 300,000, about 300,000 to about 400,000, about 400 The method includes generating about 25,000 to about 500,000, about 500,000 to about 600,000, about 600,000 to about 700,000, about 700,000 to about 800,000, about 800,000 to about 900,000, about 900,000 to about 1,000,000, about 10,000 to about 50,000, about 10,000 to about 100,000, about 25,000 to about 100,000, about 10,000 to about 150,000, about 10,000 to about 200,000, or about 50,000 to about 150,000 simulated structural conformations. Generating conformational dynamics is time-consuming and costly, so the advanced sampling developed by the MD community can increase the effective simulation time for the same amount of computational time / computational cost.Thus, in some embodiments, generating conformational dynamics as disclosed herein involves using enhanced sampling MD methods. The majority of enhanced sampling methods fall into two categories: collective variable-based and collective variable-free methods. The bottleneck in generating conformational dynamics lies in the fact that there are high energy barriers separating different structures / conformations, so that transitions between them appear as rare events. A straightforward and effective idea to accelerate thermodynamic calculations is to modify the potential energy surface by adding a bias potential to the Hamiltonian of the system, i.e., to reduce the energy barrier and increase the sampling transition region. Such methods include, but are not limited to, the widely used umbrella sampling, potential smoothing methods, J-walking, local elevation, conformational flooding, hyperdynamics, conformational space annealing, adaptive bias force methods, metadynamics (MetaD), local elevation umbrella sampling, and variational enhanced sampling (VES). In some embodiments, to effectively guide the simulation, these methods use a predetermined reaction coordinate or collective variable (CV). However, it is well known that in many systems, a suitable reaction coordinate is not easily identified. Under such circumstances, methods that do not use collective variables may be useful. In this latter type of method, the canonical probability distribution is changed to a distribution that will induce a wider sampling of potential energies. Several such methods have been proposed, such as parallel tempering or replica exchange molecular dynamics (REMD), simulated tempering, multicanonical simulation, temperature accelerated dynamics, Wang-Landau algorithm, accelerated molecular dynamics simulation, statistical temperature sampling, temperature accelerated molecular dynamics, envelope distribution sampling, integrated tempering sampling (ITS), and accelerated envelope distribution sampling. In some embodiments, the enhanced sampling method is MetaD, VES, REMD, or ITS.In some embodiments, the enhanced sampling method is a hybrid method that combines features of different enhanced sampling methods, see Yi et al (2019) J. Chem. Phys. 151:070902, which is incorporated by reference in its entirety.
[0091] IB Unsupervised Dimensionality Reduction In some aspects, the methods disclosed herein include a processing step in which the dimensionality of the conformational dynamics (e.g., from an MD simulation) is reduced. For example, the step of processing the conformational dynamics of a reference molecule (e.g., a reference protein) or a corresponding variant molecule (e.g., a missense variant protein) to generate a reduced dimensional representation of the conformational landscape of the reference molecule or the variant molecule includes independently processing the conformational dynamics of the molecule or variant molecule (or set of variant molecules) to generate a reduced dimensional representation of the conformational landscape of each of the reference molecule (e.g., a reference protein) and each of its variants (e.g., variant proteins).
[0092] As used herein, the term "conformational landscape" refers to the set of conformations (e.g., structural conformations) that can be adopted by a reference molecule (e.g., a reference protein) or a corresponding variant molecule (e.g., a missense variant protein) of the present disclosure. The large number of atoms in a molecular system creates a high-dimensional data set.
[0093] As used herein, the term "dimensionality" as applied to conformational dynamics (e.g., MD) of the present disclosure refers to the number of dimensions or degrees of freedom present in a set of conformational dynamics (i.e., simulated structural conformations) generated according to the methods disclosed herein.
[0094] High-dimensional data is often sparse, which is problematic for any method that requires statistical significance. Machine learning problems that involve learning "natural states" from a finite number of data samples in a high-dimensional feature space where each feature has a range of possible values typically require vast amounts of training data to ensure that there are several samples for each combination of values.
[0095] The conformational dynamics (e.g., from MD) of a reference molecule (e.g., a reference protein) or a corresponding variant molecule (e.g., a missense variant protein) of the present disclosure may include, for example, 10,000 to 100,000 degrees of freedom (n non-solvent atoms x three dimensions x f frames x t trajectories). The high dimensionality of the conformational dynamics (e.g., from MD) prevents the meaningful application of machine learning methods and, as a consequence, prevents the system from being scaled up to a true high-throughput mode. Thus, a low-dimensional representation of the conformational space is necessary. These low-dimensional representations of the conformational space are sufficiently simplified to be used to train machine learning systems.
[0096] The process of generating a reduced dimensional representation of conformational space according to the present disclosure includes two defined stages: structural feature extraction and dimensionality reduction. Structural feature extraction returns a subset of features (e.g. bone angles), while dimensionality reduction creates new features from features of the original features (e.g. the extracted feature subset).
[0097] In some aspects, processing the conformational dynamics (e.g., from MD) to generate a reduced dimensional representation of the conformational landscape of a reference molecule (e.g., a reference protein) or a corresponding variant molecule (e.g., a missense variant protein) includes a feature selection process in which structural features are selected and calculated from the conformational dynamics (e.g., from MD) as a first dimensionality reduction step. This feature selection approach seeks to find a subset of input variables (also called features or attributes).
[0098] As used herein, the term "structural feature" refers to a feature or attribute that defines or influences a particular conformational state of a reference molecule (e.g., a reference protein) or a corresponding variant molecule (e.g., a missense variant protein) in the conformational dynamics (e.g., from MD) of the reference molecule (e.g., a reference protein) or the corresponding variant molecule (e.g., a missense variant protein). Structural features are not limited to purely structural features of a molecule (e.g., a protein) and may also include features or attributes related to molecular function (e.g., protein function) or thermodynamics. In some aspects, a structural feature relates to molecular geometry (e.g., protein geometry). In some aspects, structural features related to protein geometry include, for example, atom positions, bond angles (e.g., backbone bond angles), backbone bond lengths, inter-residue distances and / or orientations, stereochemical conformation, secondary structure, tertiary structure, etc. In some aspects, a structural feature relates to backbone bond angles. In some aspects, structural features related to backbone bond angles are phi (φ) angles, psi (ψ) angles, omega (ω) angles, or any combination thereof. In some embodiments, the structural features include side chain chi (χ) angles. In some embodiments, the structural features relate to physicochemical properties of the protein. In some embodiments, structural features related to physicochemical properties of the protein include, for example, surface accessibility, surface energy, solvent accessibility, etc. In some embodiments, the structural features relate to domain architecture, for example, relative orientation of domains, or subdomains, for example, protruding loops or strands, which may have different local dynamics relative to the motion of the overall structure. Thus, the dynamics of the subdomains in the structure can be obtained.
[0099] In some embodiments, the structural features refer to the binding properties of protein / ligand interactions, such as protein-protein interactions, protein-ion interactions, protein-small molecule interactions, or protein-nucleic acid interactions. In some embodiments, the structural features refer to, for example, the shape, affinity, or specificity of a cofactor binding site or a substrate binding site. In some embodiments, the ligand can be a substrate when the target protein and the corresponding missense variant protein are enzymes. In some embodiments, the ligand can also be an agonist, antagonist, effector molecule, or antibody. An energy interaction can be obtained that refers to the sum of all atomic interactions between the protein and the ligand.
[0100] In some embodiments, the structural characteristic refers to the root mean square variation (RMSF) and / or the root mean square error (RMSD).
[0101] In some embodiments, the structural features used to reduce the total number of features in the conformational dynamics (e.g., from MD) of a reference molecule (e.g., a reference protein) or a corresponding variant molecule (e.g., a missense variant protein) include phi (φ) angles, psi (ψ) angles, omega (ω) angles, residue pair distances (inter-residue distances), residue variability, solvent accessibility, distance from the center of mass to a binding site or to a ligand bound to such a binding site, global structural features related to protein shape (e.g., lengths along different axes), domain relative orientation, binding site interactions, energetic features (e.g., van der Waals forces, electrostatic forces, non-bonded energies occurring (e.g., between a ligand and a protein)), and any combination thereof.
[0102] In some aspects, extraction of structural features from a conformational dynamics (e.g., MD) for a reference molecule (e.g., a reference protein) or corresponding variant molecule (e.g., a missense variant protein) of the present disclosure can result in at least about a 3-fold, about a 4-fold, about a 5-fold, about a 6-fold, about a 7-fold, about a 8-fold, about a 9-fold, about a 10-fold, about a 15-fold, about a 20-fold, about a 25-fold, about a 30-fold, about a 35-fold, about a 40-fold, about a 45-fold, about a 50-fold, about a 60-fold, about a 70-fold, about a 80-fold, about a 90-fold, or about a 100-fold reduction in structural features relative to the total number of features present in the conformational dynamics (e.g., MD) prior to extraction of the structural features. In some embodiments, extraction of structural features from a conformational dynamics (e.g., MD) of a reference molecule (e.g., a reference protein) or corresponding variant molecule (e.g., a missense variant protein) of the present disclosure can result in at least about a 2-fold to about 5-fold, about 5-fold to about 10-fold, about 10-fold to about 20-fold, about 20-fold to about 30-fold, about 30-fold to about 40-fold, about 40-fold to about 50-fold, about 50-fold to about 60-fold, about 60-fold to about 70-fold, about 70-fold to about 80-fold, about 80-fold to about 90-fold, or about 90-fold to about 100-fold reduction of structural features relative to the total number of features present in the conformational dynamics (e.g., MD) prior to extraction of the structural features.
[0103] The processing of conformational dynamics can further include a dimension reduction step. Thus, in some embodiments, the extraction of structural features from conformational dynamics (e.g., from MD) can be followed by a dimension reduction step, for example, using Principal Component Analysis (PCA) or an autoencoder, and the total number of features present in the simplified conformational states resulting from the structural feature extraction step can be further reduced (or converted into new features). Methods such as PCA or autoencoders transform data from a high-dimensional space to a space with fewer dimensions. The data transformation can be linear, such as PCA, but there are also many non-linear dimension reduction techniques.
[0104] Principal component analysis (PCA) is the process of computing the principal components of a collection of data points and using them to base variations in the data, sometimes by using only the first few principal components and ignoring the rest. A variation of PCA called time-structure independent component analysis or tICA is particularly useful for simulations because it identifies components that vary slowly along the time dimension, eliminating variations that are unlikely to be useful for understanding the conformational motion of proteins.
[0105] Autoencoders are a type of artificial neural network used to learn efficient data coding in an unsupervised fashion. Examples include regularized (sparse, denoising, and shrinking) autoencoders, which are effective at learning representations for subsequent classification tasks, and variational autoencoders, which have applications as generative models.
[0106] In the case of PCA and autoencoders, the number of output features is chosen to account for a significant portion of the variation in the data.
[0107] In some embodiments, dimensionality reduction of conformational dynamics (e.g., from MD) may be performed using, for example, PCA, time-structure independent component analysis (tICA), non-negative matrix factorization (NMF), kernel PCA, graph-based kernel PCA, linear discriminant analysis (LDA), generalized discriminant analysis (GDa), autoencoders, t-distributed stochastic neighbor embedding (t-SNE), or uniform manifold approximation and projection (UMAP).
[0108] In some embodiments, structural feature extraction and dimensionality reduction can be combined into a single step, e.g., as a pre-processing step following K-NN on feature vectors in a reduced dimensional space using PCA, LDA, NMF, or canonical correlation analysis (CCA) techniques. In some embodiments, the dimensionality reduction step results in at least about 3-fold, about 4-fold, about 5-fold, about 6-fold, about 7-fold, about 8-fold, about 9-fold, about 10-fold, about 15-fold, about 20-fold, about 25-fold, about 30-fold, about 35-fold, about 40-fold, about 45-fold, about 50-fold, about 60-fold, about 70-fold, about 80-fold, about 90-fold, or about 100-fold reduction of structural features relative to the total number of features present in the conformational dynamics (e.g., from MD) prior to extraction of the structural features. In some aspects, the dimensionality reduction step results in at least a 3-fold, 4-fold, 5-fold, 6-fold, 7-fold, 8-fold, 9-fold, 10-fold, 15-fold, 20-fold, 25-fold, 30-fold, 35-fold, 40-fold, 45-fold, 50-fold, 60-fold, 70-fold, 80-fold, 90-fold, or 100-fold reduction in structural features relative to the total number of features present in the conformational dynamics (e.g., from MD) after extraction of the structural features. In some embodiments, the dimensionality reduction step results in at least about 3-fold to about 5-fold, about 5-fold to about 10-fold, about 10-fold to about 20-fold, about 20-fold to about 30-fold, about 30-fold to about 40-fold, about 40-fold to about 50-fold, about 50-fold to about 60-fold, about 60-fold to about 70-fold, about 70-fold to about 80-fold, about 80-fold to about 90-fold, or about 90-fold to about 100-fold reduction in structural features relative to the total number of features present in the conformational dynamics (e.g., from MD) after extraction of the structural features.
[0109] FIG. 8 shows the specific feature quantification methods specified according to the features used (e.g., dihedral angles, tlPCA; connectivity, tlCA; dihedral angles, PCA; contacts, PCA; and connectivity, PCA) followed by the dimensionality reduction method applied.
[0110] Identification of metastable conformational states by IC In some aspects, the methods disclosed herein include processing steps related to extracting conformational states (e.g., metastable conformational states) from a low-dimensional conformational landscape of a reference molecule (e.g., a reference protein) to seed a conformational dynamics generation (e.g., MD) of a variant molecule (e.g., a variant protein). In some aspects, the methods disclosed herein may also include processing steps related to extracting conformational states (e.g., metastable conformational states) from a low-dimensional conformational landscape of a variant molecule (e.g., a variant protein) to seed a conformational dynamics generation (e.g., MD) of a further variant molecule (e.g., a variant protein).
[0111] In some embodiments, extracting specific conformational states (e.g., metastable conformational states from the low-dimensional conformational landscape) comprises clustering the conformational states according to structural similarity, e.g., the various metastable conformational states in the low-dimensional conformational landscape are the result of clustering the conformational states. In some embodiments, the average distance between backbone atoms of the identified conformational states is measured according to RMSD, and the RMSD is at least about 2 Å. In other embodiments, different RMSD thresholds can be used, e.g., 1 Å, 1.5 Å, 2.5 Å, 3 Å, 3.5 Å, 4 Å, 4.5 Å, 5 Å, or any intermediate value.
[0112] By extracting specific conformational states (e.g., metastable conformations) from the conformational landscape (e.g., MD conformational landscape) of a reference molecule (e.g., a reference protein, which can be a wild-type protein) to seed a simulation (e.g., an MD simulation) of a variant, the overall simulation time required to compare variants can be dramatically reduced. For example, in some embodiments, the extraction of a set of conformational states and the subsequent simulation of the variant molecular structures in these conformational states achieves at least about a 10-fold reduction in simulation time. In some embodiments, the extraction of a set of conformational states and the subsequent simulation of the variant molecular structures in these conformational states achieves at least about a 3-fold, at least about a 4-fold, at least about a 5-fold, at least about a 6-fold, at least about a 7-fold, at least about a 8-fold, at least about a 9-fold, at least about a 10-fold, at least about a 15-fold, or at least about a 20-fold reduction in simulation time. In some embodiments, extraction of a set of conformational states and subsequent simulation of variant molecular structures in these conformational states achieves a reduction in simulation time of about 3x, about 4x, about 5x, about 6x, about 7x, about 8x, about 9x, about 10x, about 15x, or about 20x. In some embodiments, extraction of a set of conformational states in these states and subsequent simulation of variant molecular structures achieves a reduction in simulation time of about 3x to about 5x, about 4x to about 6x, about 5x to about 8x, about 6x to about 9x, about 5x to about 10x, about 10x to about 15x, or about 15x to about 20x. All performance improvements are for workflows that do not use representative metastable conformations to simulate the effects of variants.
[0113] Clustering states from the low-dimensional conformational landscape of a reference molecule (e.g., a reference protein) to seed a simulation of a variant molecule (e.g., a variant protein) can be done using, for example, k-means clustering, hierarchical clustering, density-based clustering, DBSCAN, spectral clustering, Gaussian mixture models, or any combination thereof.
[0114] Clustering allows for the extraction of multiple conformational states from a low-dimensional conformational landscape, typically in the range of 10-100 conformational states. In some embodiments, at least 5, at least 10, at least 15, at least 20, at least 25, at least 30, at least 35, at least 40, at least 45, at least 50, at least 55, at least 60, at least 65, at least 70, at least 75, at least 80, at least 85, at least 90, at least 95, or at least 100 conformational states are extracted. In some embodiments, about 5, about 10, about 15, about 20, about 25, about 30, about 35, about 40, about 45, about 50, about 55, about 60, about 65, about 70, about 75, about 80, about 85, about 90, about 95, or about 100 conformational states are extracted. In some embodiments, about 10 to about 20, about 20 to about 30, about 30 to about 40, about 40 to about 50, about 50 to about 60, about 60 to about 70, about 70 to about 80, about 80 to about 90, or about 90 to about 100 conformational states are selected.
[0115] In some embodiments, the methods disclosed herein further comprise prioritizing the clustered conformational states prior to generating the conformational dynamics (e.g., MD simulations). In some embodiments, the clustered conformational states are prioritized according to cluster properties, such as distance from the conformational landscape centroid, cluster occupancy, cluster distribution statistics, structural properties of the clusters, thermodynamics of the clusters, or any combination thereof. In some embodiments, the structural properties of the clustered conformational states are selected from the group consisting of bond angles (e.g., φ-ψ dihedral angles), inter-residue distance, surface accessibility, domain relative orientation, binding site interactions, structural similarity to known protein conformations, and any combination thereof. In some embodiments, the cluster distribution statistics are selected from the group consisting of Silhouette score, elbow score, Calinski-Harabasz index, Rand index, mutual information, homogeneity, completeness, V-scale, Davies-Bouldin index, and any combination thereof. In some cases, extracting states from the low-dimensional conformational landscape includes extracting a representative conformational state (e.g., a metastable conformational state), which includes selecting a protein conformation that has the closest distance in the low-dimensional representation of the conformational landscape to the conformational centroid of the selected conformational cluster.
[0116] ID variant molecules and construction of variant molecule structures The methods disclosed herein include generating a variant molecule (e.g., a variant protein) or set thereof and its corresponding variant molecular structure (e.g., a variant protein structure) by introducing at least one variant (e.g., a missense mutation) into the simulated protein conformation of the variant molecule in the extracted state. For example, a mutation identified in a sample obtained from a subject can be introduced into the MD-simulated conformation disclosed herein to generate a variant protein structure. The effect of such a mutation can then be evaluated according to the methods disclosed herein. In other aspects, the variant (e.g., a mutation) can be a speculative variant introduced into the MD-simulated conformation disclosed herein to gain knowledge about the functional and physiological effects, and optionally their severity, when, for example, a particular domain, region, functional site (e.g., a ligand-binding site) is mutated into a reference molecule (e.g., a reference protein) or a variant molecule (e.g., a variant protein that already contains other variants). In some aspects, a collection of variant molecules (e.g., variant proteins) generated according to the methods of the present disclosure can be stored as a library.
[0117] In some aspects of the methods disclosed herein, at least one variant (e.g., a missense mutation) is introduced, for example, into a reference molecule (e.g., a reference protein such as a wild-type protein), or the variant is a pathogenic variant (e.g., a pathogenic mutation). The term "pathogenic variant" as used herein refers to a malignant variant associated with the appearance of a disease. In some aspects, the pathogenic variant is a pathogenic mutation. The term "pathogenic mutation" as used herein refers to a malignant mutation associated with the appearance of a disease. These terms also refer to variants or mutations that are risk susceptible to a disease, and the variant or mutation can cause a corresponding disease phenotype, i.e., is a genetic change that causes a disease or disorder and / or increases the susceptibility or predisposition of an individual to that particular disease or disorder. However, it is not limited to the case where only the variant or mutation causes a disease or disorder. In some embodiments, for example, at least one variant or mutation (e.g., a missense mutation) introduced into a reference molecule (e.g., a reference protein such as a wild-type protein) or a variant thereof is a non-pathogenic (benign) variant or mutation, i.e., a variant or mutation that does not cause a disease or disorder in a subject or does not predispose the subject to a particular disease or disorder. In some embodiments, for example, at least one variant (e.g., a mutation such as a missense mutation) introduced into a reference molecule (e.g., a reference protein such as a wild-type protein) or a variant thereof is a variant or mutation of unknown significance. That is, the available empirical data on the variant does not allow such variant or mutation to be classified as either pathogenic or non-pathogenic. One of the objectives of the methods disclosed herein is to classify such variants or mutations of unknown significance as either pathogenic or non-pathogenic, and to determine the likelihood that a variant or mutation of unknown relevance, for example, classified as pathogenic, will cause a disease or disorder, and / or, for example, the severity of such a disease or disorder.
[0118] In some aspects, the variant molecule (e.g., protein) or set thereof includes protein missense variants observed in medical genetic testing. The term "genetic testing" as used herein refers to the analysis of DNA, RNA, or protein in a sample from an individual, whereby, without limitation, the sequence of one or more genes in the sample, the presence or absence of one or more genetic markers, variance, mutation, polymorphism, or microsatellite sequences associated with the genes, the presence of one or more viral, viral-like, or repetitive sequences, haplotypes spanning one or more genes, copy number of one or more genes, the amount or characteristics of RNA or protein expressed from one or more genes, the location of the gene in the genome, chromosome number, or chromosomal integrity is determined. In some aspects, the present disclosure relates to genetic tests useful in medicine for diagnosing genetic diseases, determining an individual's propensity for multifactorial diseases, and predicting an individual's response to therapeutic drugs. Genetic tests have been developed for many genetic diseases, including, but not limited to, Huntington's disease, cystic fibrosis, and phenylketonuria. Genetic tests have also been developed for genes that cause diseases, including but not limited to atherosclerosis, heart failure, stroke, anemia, cancer, coagulation disorders, dementia, endocrine disorders, and lung diseases. Genetic tests have also been reported to predict the pharmacokinetic and pharmacodynamic properties of many drugs, including but not limited to drugs for treating high cholesterol, drugs for treating cancer, drugs to reduce high blood pressure, and drugs to treat dementia. Despite the potentially large impact of genetic testing on healthcare, the availability of genetic testing is currently very limited. To date, only a small portion of the genes in the genome have been characterized, and only a small subset of variations that predict disease or response to therapy have been reported.
[0119] In some embodiments, the variant molecular structures (e.g., variant protein structures) or sets thereof used in the methods disclosed herein or included in the libraries assembled according to the methods disclosed herein are generated in silico using backbone-dependent rotamer libraries. Other methods that can be used to generate these variant molecular structures in silico include, for example, FoldX or Rosetta. The term "rotamer" refers to a combination of torsion angles that describe the geometry of amino acid side chains.
[0120] Rotamer libraries are known to those skilled in the art and can be obtained from a variety of sources, including the Internet. Rotamers are low-energy side chain conformations. The use of a rotamer library allows the modeling of a structure to attempt the most likely side chain conformations, saving time and generating structures that are more likely to be accurate.
[0121] In some aspects, the in silico generation of variant molecular structures (e.g., variant protein structures) used in the methods disclosed herein involves identifying side chain conformations (rotamers) with existing backbone dihedral angles of the variant molecular structure (e.g., variant protein) at variant positions in a backbone-dependent rotamer library.
[0122] In some aspects, the method of in silico generation of variant molecular structures further comprises testing whether the introduction of a rotamer identified from a backbone-dependent rotamer library causes steric clashes with nearby residues. In some aspects, the testing is performed until a rotamer that minimizes the steric clashes is identified. In some aspects, the method of in silico generation of variant molecular structures (e.g., variant protein structures) of the present disclosure further comprises performing minimization and equilibration simulations to regularize the geometry of the variant molecular structures (e.g., variant protein structures) and reduce the energetic impact of introducing variants (e.g., missense mutations) into the variant molecular structures.
[0123] IE Predictive Modeling In some aspects, the methods disclosed herein include a processing step that includes training a machine learning (ML) predictive model using a low-dimensional representation of the conformational landscape of a reference molecule (e.g., a reference protein) and a variant molecule (e.g., a variant protein) using clinical data as training labels, where the predictive model classifies or scores the effect of the variant (e.g., a protein missense mutation) on the reference molecule (e.g., a reference protein). Thus, in some aspects, the training set of the selected ML method includes both MD simulation related data as well as other datasets (e.g., clinical data).
[0124] The phrase "training set" refers to a set of observations that are the subject of fitting one or more models and on which the models are built. For example, in the case of a predictive model in which protein activity is predicted, the training set typically includes complete or partial protein sequence information, along with activity values for each protein in the library. In some cases, multiple activity types (e.g., kinetic constant data and thermostability data) are provided together in the training set. In addition to data related to protein sequence, protein structure, protein dynamics, or protein activity, the training set may include, for example, clinical data.
[0125] In some embodiments, clinical data used as training labels in the training set include any parameters related to a disease or condition, symptoms, patients suffering or at risk of suffering from sequelae associated with the target protein or variants thereof, treatment of a disease or condition, outcome of treatment, etc. For example, clinical data can be used as training labels, including, for example, biomarker status (e.g., the presence or absence of a particular biomarker or its expression level, e.g., RNA or protein expression level), biometric data, lifestyle-related data, response to treatment, symptoms of a disease or condition, protein expression data, type of treatment administered, dosage, dosage regimen, route of administration, presence or absence of concomitant therapy, response to therapy, age, weight, sex, ethnicity, or any combination thereof. In some embodiments, data that can be used as training labels include, for example, data from NCBI ClinVar database submissions, INVITAE™ clinical interpretations, other experimental or computational models.
[0126] In some embodiments, the training labels are derived from a combination of individually specified clinical data parameters, e.g., several protein expression values, biometric parameters, etc., and can be combined into an index (e.g., BMI) or score (e.g., the ratio between the expression levels of two proteins).
[0127] In some embodiments, the classification of the effect of the variant (e.g., missense mutation) on the reference molecule (e.g., reference protein) comprises calculating a predicted pathogenicity probability based on the output of the ML prediction model. In some embodiments, the probability is calculated as a confidence score from the ML prediction model. In some embodiments, the ML output of the methods disclosed herein can predict a specific phenotypic disruption and rank the relative severity of the disruption compared to the reference molecule (e.g., reference protein).
[0128] In general, the computer-implemented methods disclosed herein generate a predicted pathogenicity score, which is a confidence score of how likely the model is to predict that a variant is pathogenic. In some aspects, a performance threshold is applied to the pathogenicity score to determine whether to assign evidence that a particular variant is pathogenic or benign.
[0129] In the context of the present disclosure, the term predictive model should be understood to include one or more predictive models or combinations of predictive models that can belong to the same or different classes. For example, predictive models are generated using different ML techniques or using different sets of clinical data as training labels. The term predictive model is used to describe the output of a particular model. For example, in addition to ML predictive models that predict the likelihood and / or severity of phenotype disruption, other ML predictive models may predict the binding of the generated variant to a particular drug, a candidate therapy, an outcome associated with treatment with a particular therapy, or any combination thereof.
[0130] In some aspects, pre-processing of the training set prior to generating the ML predictive model includes feature selection and, optionally, feature normalization. In some aspects, the ML model is trained using hyperparameter tuning. In ML, hyperparameter optimization or tuning is the problem of selecting an optimal set of hyperparameters of a learning algorithm. Hyperparameters are parameters whose values are used to control the learning process. In contrast, values of other parameters (typically node weights) are learned. The same kind of ML predictive model may require different constraints, weights, or learning rates to generalize different data patterns. These measures are called hyperparameters and need to be tuned so that the model can optimally solve the ML problem. Hyperparameter optimization finds a tuple of hyperparameters that produces an optimal model that minimizes a predefined loss function for given independent data. The objective function receives a tuple of hyperparameters and returns the associated loss. Cross-validation is often used to estimate this generalization performance.
[0131] In some embodiments, the ML predictive models disclosed herein are obtained by application of machine learning techniques selected from the group consisting of logistic regression, random forest, artificial neural network (ANN), support vector machine (SVM), XGBoost (XGB; an implementation of gradient boosting decision trees designed for speed and performance), Glmnet (a package that fits generalized linear models via penalized maximum likelihood), cforest (an implementation of random forest and a bagging ensemble algorithm that utilizes conditional inference trees as base learners), classification and regression trees for machine learning (CART), Treebag (bagging, i.e., bootstrap aggregation, an algorithm that builds multiple models from isolated subsets of training data to improve model accuracy for regression and classification problems, as well as an algorithm that builds a final aggregate model), K-nearest neighbors (kNN), or a combination thereof. In certain embodiments, the ML predictive models disclosed herein are obtained by application of at least one ANN (e.g., 1, 2, 3, 4, or 5 ANNs).
[0132] IEi Logistic Regression ML Prediction Model In some aspects, the machine learning (ML) predictive models of the present disclosure are generated using logistic regression. Logistic regression is often considered one of the best predictors for small data sets. However, tree-based models (e.g., Random Forest, ExtraTrees) and ANNs can reveal latent interactions between features. However, when there are few interactions, logistic regression and more complex models have similar performance. In some aspects of the present disclosure, the ML predictive models are generated using logistic regression. In one aspect, the following hyperparameters were tuned for the logistic regression model: stopping criterion tolerance, penalty, solving algorithm (e.g., LIBLINEAR, stochastic average gradient).
[0133] IEii Random Forest ML Prediction Model In some aspects, the machine learning (ML) predictive models of the present disclosure are generated using random forests (RF). RF, or random decision forests, is an ensemble learning method for classification, regression, and other tasks that operates by building a large number of decision trees at training time. For classification tasks, the output of RF is the class selected by most trees. RF adjusts for habits of decision trees that cause them to overfit the training set. RF generally outperforms decision trees. The RF algorithm uses bootstrap samples to generate individual decision trees. Trees are grown by selecting random subsamples of the independent variables at each node and selecting the variables that produce the best results.
[0134] In one particular aspect, the predictive model can optimize validation performance (using class-weighted F1 score as tuning metric) between both logistic regression and random forest. For random forest, the following hyperparameters were tuned: number of estimators, splitting criterion, maximum tree depth, minimum samples to split tree, minimum samples in leaf node, and maximum number of features.
[0135] IEiii Artificial Neural Network ML Prediction Model In some aspects, the machine learning (ML) predictive models of the present disclosure are generated using artificial neural networks (ANNs). A "neural network" is a model that includes an interconnected group of processing elements or "neurons" that process information using a connectivity approach to computation. Neural networks are used to model complex relationships between input and output values or to find data patterns. Most neural networks process data in a nonlinear, distributed, and parallel manner. In most cases, neural networks are adaptive systems that change their structure during the learning phase. Functions are performed collectively and in parallel by processing elements rather than using an explicit description of the subtasks to which various units are assigned.
[0136] Neural networks are described in Duda et al., 2001, Pattern Classification, Second Edition, John Wiley & Sons, Inc., New York, and Hastie et al., 2001, which are incorporated by reference in their entireties. In some embodiments, a neural network of the present disclosure, such as a backpropagation neural network (see, e.g., Abdi, 1994, "A neural network primer", J. Biol System. 2, 247-283, which is incorporated by reference in its entirety), may include an input layer, a hidden layer, and an output layer. In some embodiments, the neural network may be implemented using the EasyNN-Plus version 4.0g software package (Neural Planner Software Inc.), scikit-learn (scikit-learn.org), PyTorch (github.com / pytorch / pytorch), TensorFlow (github / tensorflow / tensorflow), Keras, Torch, Flux, Caffe, CNTK (Microsoft Cognitive Toolkit), or any other machine learning package or program known in the art.
[0137] In some embodiments, the ML predictive model comprises a model derived from an ANN. In some embodiments, the ANN is a feedforward neural network. In some embodiments, the ANN is a simple perceptron network.
[0138] In some aspects, the ANN is a multi-layer perceptron (MLP). In some aspects, the activation function is a sigmoid function. In some aspects, the activation function is a rectified linear unit (ReLU) or a variant thereof (e.g., noisy ReLU, leaky ReLU, parametric ReLU, or exponential LU). In some aspects, the MLP includes three or more layers of non-linear activation nodes (input layer and output layer with one or more hidden layers). The multiple layers and non-linear activation distinguish the MLP from a linear perceptron. It can distinguish data that is not linearly separable. The MLP is fully connected, so each node in one layer connects to every node in the next layer with a specific weight w<-i><-j>. Learning is done in the perceptron by changing the connection weights after each portion of data is processed based on the amount of error in the output compared to the expected result. This is an example of supervised learning, which is done by backpropagation.
[0139] In some embodiments, the MLP has three layers. In other embodiments, the MLP has four or more layers. In some embodiments, the MLP has a single hidden layer. In other embodiments, the MLP has multiple hidden layers.
[0140] In some aspects, the input layer is: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155 5, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121 1, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, or 150 neurons. In some embodiments, the input layer comprises 70-100 neurons. In some embodiments, the input layer comprises 70-80 neurons. In some embodiments, the input layer comprises 80-90 neurons. In some embodiments, the input layer comprises 90-100 neurons. In some embodiments, the input layer comprises 70-75 neurons. In some embodiments, the input layer comprises 75-80 neurons. In some embodiments, the input layer comprises 80-85 neurons. In some embodiments, the input layer comprises 85-90 neurons. In some embodiments, the input layer comprises 90-95 neurons. In some embodiments, the input layer comprises 95-100 neurons.In some embodiments, the input layer is at least about 1 to at least about 5, at least about 5 to at least about 10, at least about 10 to at least about 15, at least about 15 to at least about 20, at least about 20 to at least about 25, at least about 25 to at least about 30, at least about 30 to at least about 35, at least about 35 to at least about 40, at least about 40 to at least about 45, at least about 45 to at least about 50, at least about 50 to at least about 55, at least about 55 to at least about 60, at least about 60 to at least about 65, at least about 65 to at least about 70, at least about 70 to at least about 75, at least about 75 to at least about 80 at least about 80 neurons, at least about 80 to at least about 85 neurons, at least about 85 to at least about 90 neurons, at least about 90 to at least about 95 neurons, at least about 95 to at least about 100 neurons, at least about 100 to at least about 105 neurons, at least about 105 to at least about 110 neurons, at least about 110 to at least about 115 neurons, at least about 115 to at least about 120 neurons, at least about 120 to at least about 125 neurons, at least about 125 to at least about 130 neurons, at least about 130 to at least about 135 neurons, at least about 135 to at least about 140 neurons, at least about 140 to at least about 145 neurons, or at least about 145 to at least about 150 neurons. In some embodiments, the input layer includes at least about 1 to at least about 10, at least about 10 to at least about 20, at least about 20 to at least about 30, at least about 30 to at least about 40, at least about 40 to at least about 50, at least about 50 to at least about 60, at least about 60 to at least about 70, at least about 70 to at least about 80, at least about 80 to at least about 90, at least about 90 to at least about 100, at least about 100 to at least about 110, at least about 110 to at least about 120, at least about 120 to at least about 130, at least about 130 to at least about 140, or at least about 140 to at least about 150 neurons.In some embodiments, the input layer includes at least about 1 to at least about 20 neurons, at least about 20 to at least about 40 neurons, at least about 40 to at least about 60 neurons, at least about 60 to at least about 80 neurons, at least about 80 to at least about 100 neurons, at least about 100 to at least about 120 neurons, at least about 120 to at least about 140 neurons, at least about 10 to at least about 30 neurons, at least about 30 to at least about 50 neurons, at least about 50 to at least about 70 neurons, at least about 70 to at least about 90 neurons, at least about 90 to at least about 110 neurons, at least about 110 to at least about 130 neurons, or at least about 130 to at least about 150 neurons. In some embodiments, the input layer comprises more than about 1, more than about 5, more than about 10, more than about 15, more than about 20, more than about 25, more than about 30, more than about 35, more than about 40, more than about 45, more than about 50, more than about 55, more than about 60, more than about 65, more than about 70, more than about 75, more than about 80, more than about 85, more than about 90, more than about 95, more than about 100, more than about 105, more than about 110, more than about 115, more than about 120, more than about 125, more than about 130, more than about 135, more than about 140, more than about 145, or more than about 150 neurons. In some embodiments, the input layer comprises less than about 1, less than about 5, less than about 10, less than about 15, less than about 20, less than about 25, less than about 30, less than about 35, less than about 40, less than about 45, less than about 50, less than about 55, less than about 60, less than about 65, less than about 70, less than about 75, less than about 80, less than about 85, less than about 90, less than about 95, less than about 100, less than about 105, less than about 110, less than about 115, less than about 120, less than about 125, less than about 130, less than about 135, less than about 140, less than about 145, or less than about 150 neurons. In some embodiments, weights are applied to the inputs of each of the neurons in the input layer. In some embodiments, the ANN comprises a single hidden layer. In some embodiments, the ANN comprises 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 hidden layers. In some embodiments, a single hidden layer comprises 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 neurons.In some embodiments, the single hidden layer comprises at least 1, at least 2, at least 3, at least 4, at least 5, at least 6, at least 7, at least 8, at least 9, or at least 10 neurons. In some aspects, the single hidden layer comprises fewer than 10, fewer than 9, fewer than 8, fewer than 7, fewer than 6, fewer than 5, fewer than 4, or fewer than 3 neurons. In some aspects, the single hidden layer comprises two neurons. In some aspects, the single hidden layer comprises three neurons. In some aspects, the single hidden layer comprises four neurons. In some aspects, the single hidden layer comprises five neurons. In some aspects, a bias is applied to the neurons in the hidden layer. In some aspects, the output layer classification is binary. In some aspects, the output layer classification is normalized to a probability distribution over the predicted output classes (e.g., pathological vs. non-pathological variants), and the components will be summed to 1 so that they can be interpreted as probabilities. In some aspects, the classification of the output layer values into the n phenotypic classes is supported by applying a logistic regression function. In some embodiments, classification of the output layer values into the n phenotypic classes is supported by applying a logistic regression prediction model (e.g., a Softmax function). Softmax assigns a fractional probability that adds up to 1.0 to each class. In some embodiments, the use of a logistic regression prediction model such as a Softmax function helps training converge more quickly. In some embodiments, a logistic regression prediction model including a Softmax function is implemented through a neural network layer immediately prior to the output layer. In some embodiments, such a neural network layer immediately prior to the output layer has the same number of nodes as the output layer. In some embodiments, various cutoffs are applied to the results of the logistic regression prediction model (e.g., a Softmax function) depending on the particular dataset used (see, for example, cutoffs applied to select a particular type of pathological phenotype or a particular severity of a particular pathological phenotype).Thus, by applying different sets of cutoffs, the phenotypic impact of missense variants can be classified or stratified not only in a binary manner (e.g., pathogenic vs. non-pathogenic), but also, for example, according to severity. In some embodiments of the present disclosure, the ANN is a multi-layer perceptron with independent state NNs with at least one hidden layer, which are then combined into a combined NN with a number of nodes equal to the number of nodes for each state NN, with at least one hidden layer and a final Softmax output layer. The multi-layer perceptron can use ReLU or tanh activation, SGD or Adam optimization, NLL, hinge or cross-entropy loss function, and dropout layers to prevent overfitting.
[0141] In some embodiments, the quality of the resulting ML prediction model is evaluated. Thus, in one embodiment of the present disclosure, each prediction model is evaluated for its ability to adequately characterize the phenotypic impact of each variant molecule (e.g., variant protein or variant gene) in the training population using methods known to those skilled in the art. For example, standard statistical methods can be used to evaluate the prediction model using cross-validation, leave-one-out cross-validation (LOOCV), N-fold cross-validation, or jackknife analysis. In another embodiment, each prediction model is evaluated for its ability to adequately characterize variant molecules (e.g., variant proteins or variant genes) that were not used to generate the prediction model. In some embodiments, a predictive model can be trained using one dataset and evaluated on another, different dataset. In one embodiment, the method used to evaluate the predictive model for its ability to adequately characterize each variant molecule (e.g., variant protein or variant gene) in the training set is a method to evaluate the sensitivity (TPF, true positive rate) and specificity (FPF, false positive rate) of the predictive model. In one aspect, the method used to test the predictive model is the Receiver Operating Characteristic ("ROC"), which provides several parameters for assessing both the sensitivity and specificity of the results of a generated predictive model (e.g., a predictive model derived from the application of an ANN).
[0142] In some aspects, the metrics used to evaluate the predictive model for its ability to adequately characterize each variant molecule (e.g., variant protein or variant gene) in the training population include classification accuracy (ACC), area under the receiver operating characteristic curve (AUC ROC), sensitivity (true positive fraction, TPF), specificity (true negative fraction, TNF), positive predictive value (PPV), negative predictive value (NPV), or any combination thereof. In a particular aspect, the metrics used to evaluate the predictive model for its ability to adequately characterize each variant molecule (e.g., variant protein or variant gene) in the training population are classification accuracy (ACC), area under the receiver operating characteristic curve (AUC ROC), sensitivity (true positive fraction, TPF), specificity (true negative fraction, TNF), positive predictive value (PPV), and negative predictive value (NPV).
[0143] II. Clinical, Diagnostic, and Drug Design Applications In some aspects, the computational methods for determining the phenotypic impact of a variant (e.g., mutation) disclosed herein may further include: (i) obtaining and processing a sample from the subject to determine the presence of any variants (e.g., missense variants) of a reference molecule (e.g., a reference protein), e.g., by obtaining a tissue sample (e.g., obtaining a tissue sample, processing the tissue sample to obtain RNA or DNA, and sequencing the RNA or DNA, e.g., using NGS methods); and / or (ii) developing a treatment plan (which may include, for example, administering a therapy or treatment to the subject).
[0144] As used herein, the terms "therapy" and "therapeutic" include prophylaxis and prophylactic, and encompass the prevention and amelioration of symptoms associated with a disease or disorder, inhibiting or slowing the progression of a disease or disorder, and treating a disease or disorder.
[0145] As used herein, the terms "treat", "treating" and "treatment" refer to any type of intervention or process performed on a subject or administration of an active agent to a subject with the goal of reversing, mitigating, ameliorating, inhibiting or slowing down or preventing the progression, onset, severity or recurrence of symptoms, complications, conditions or biochemical signs associated with a disease and improving overall survival. Treatment can be the treatment of a subject with a disease or a subject not having a disease (e.g., for prophylaxis). As used herein, the terms "treat", "treating" and "treatment" refer to the administration of an effective dose or effective dosages.
[0146] As used herein, the term "disease" refers to any phenotype or phenotypic trait of interest, including, for example, a disease or disease state, a predisposition or susceptibility to a disease, or an abnormal drug response. Illustrative and non-limiting examples of disease states include cancer, high cholesterol levels, congestive heart failure, high blood pressure, diabetes, glucose intolerance, depression, anxiety, infectious diseases, toxic conditions, side effects of drug therapy, ineffectiveness of drug therapy, alcoholism, poisoning, trauma, etc. In general, a disease is a specific characteristic process in the body with a specific cause and characteristic symptoms. As used herein, the term "disorder" refers to an irregularity, disturbance, or interruption of the normal functioning of the body.
[0147] "Subject" generally refers to a relevant organism and sequence information, and optionally phenotypic information, available for analysis. As used herein, the term "subject" includes any human or non-human animal. The terms subject and patient are used interchangeably herein. The term "non-human animals" includes, but is not limited to, dogs, cats, horses, cows, pigs, wild boars, sheep, goats, buffalo, bison, llamas, deer, elk, and other large animals, as well as their young animals, including calves and lambs, and primates, such as mice, rats, rabbits, guinea pigs, monkeys, and other laboratory animals. Among animals, mammals are preferred, most preferably precious and valuable animals, such as household pets, race horses, and animals used to directly (e.g., meat) or indirectly (e.g., milk) produce food for human consumption, but also include laboratory animals. In certain aspects, the subject is a human. Thus, the present disclosure is applicable to clinical, veterinary, and research applications.
[0148] For example, variants (e.g., missense mutations) identified in a variant molecule (e.g., a variant gene) from a subject can be used as input to generate variant molecular structures (variant protein structures) in silico, which can be used, for example, to determine the phenotypic impact of the mutation, to generate MD simulation structures for incorporation into a library or stored for future analysis, to expand ML training sets to improve ML predictive models, or any combination thereof.
[0149] In some embodiments, a sample for genetic testing used to identify variants of a reference molecule (e.g., a reference protein) is requested, for example, by a health care provider (e.g., a doctor) or a health care benefit provider. In some embodiments, a sample for genetic testing can be obtained and / or processed by the same or a different health care provider (e.g., a nurse, a hospital) or a clinical laboratory, and after processing, the results can be transferred to the original health care provider or to a further health care provider, health care benefit provider, or facility that performs computational analysis of the variant-related information obtained from the subject (e.g., performing part or all of the workflow disclosed herein). Based on the results from the methods disclosed herein, a decision can be made regarding whether to administer a treatment.
[0150] In some embodiments, the term "administering" can also include initiating a therapy, stopping or discontinuing a therapy, temporarily interrupting a therapy, or modifying a therapy (e.g., increasing the dosage or frequency of dosage, or adding one of multiple therapeutic agents in a combination therapy).
[0151] As used herein, the term "healthcare provider" refers to an individual or institution that directly interacts with and administers to living subjects, e.g., human patients. Non-limiting examples of health care providers include doctors, nurses, technicians, therapists, pharmacists, counselors, alternative medicine practitioners, medical facilities, physician offices, hospitals, emergency rooms, clinics, urgent care centers, alternative medicine clinics / facilities, and any other entity that provides general and / or specialized treatment, evaluation, maintenance, therapy, medication, and / or advice related to all or a portion of a patient's health condition, including, but not limited to, general medical, specialized medical, surgical, and / or any other type of treatment, evaluation, maintenance, therapy, medication, and / or advice.
[0152] As used herein, the term "clinical laboratory" refers to a facility for the testing or processing of material derived from a living subject, e.g., a human. Non-limiting examples of processing include biological, biochemical, serological, chemical, immunohematological, hematological, biophysical, cytological, pathological, genetic, or other testing of material derived from the human body, for example, for the purpose of providing information for the diagnosis, prevention, or treatment of any disease or disorder, or for the assessment of the health of a living subject, e.g., a human. These tests may also include procedures for taking or otherwise obtaining samples, preparing, determining, measuring, or otherwise describing the presence or absence of various substances in the body of a living subject (e.g., a human) or in samples obtained from the body of a living subject (e.g., a human).
[0153] As used herein, the term "health care benefits provider" encompasses an individual party, organization, or group that offers, provides, pays in whole or in part for, or is otherwise associated with, a patient's access to one or more health care benefits, benefit plans, health insurance, and / or health care account programs.
[0154] The healthcare provider may obtain a sample, process a sample, submit a sample, receive a sample, transport a sample, analyze or measure a sample, quantify a sample, provide results obtained after analyzing / measuring / quantifying a sample, receive results obtained after analyzing / measuring / quantifying a sample, compare / score results obtained after analyzing / measuring / quantifying one or more samples, provide a comparison / score from one or more samples, obtain a comparison / score from one or more samples, administer a therapy, begin administration of a therapy, stop administration of a therapy, continue administration of a therapy, temporarily interrupt administration of a therapy, increase the amount of a therapeutic agent administered, decrease the amount of a therapeutic agent administered, continue administration of the amount of a therapeutic agent, increase the frequency of administration of a therapeutic agent, decrease the frequency of administration of a therapeutic agent, maintain the same frequency of administration of a therapeutic agent, replace a therapy or therapeutic agent with at least another therapy or therapeutic agent, combine a therapy or therapeutic agent with at least another therapy or additional therapeutic agent.
[0155] In some aspects, the health care benefit provider may approve or deny, for example, collection of a sample, processing of a sample, submission of a sample, receipt of a sample, transfer of a sample, analysis or measurement of a sample, quantification of a sample, provision of results obtained after analysis / measurement / quantification of a sample, transfer of results obtained after analysis / measurement / quantification of a sample, comparison / scoring of results obtained after analysis / measurement / quantification of one or more samples, transfer of comparison / score from one or more samples, administration of a therapy or therapeutic agent, initiation of administration of a therapy or therapeutic agent, discontinuation of administration of a therapy or therapeutic agent, continuation of administration of a therapy or therapeutic agent, temporary interruption of administration of a therapy or therapeutic agent, increase in the amount of a therapeutic agent administered, decrease in the amount of a therapeutic agent administered, continue administration of the amount of a therapeutic agent, increase the frequency of administration of a therapeutic agent, decrease the frequency of administration of a therapeutic agent, maintain the same dosing frequency of a therapeutic agent, replacement of a therapy or therapeutic agent with at least another therapy or therapeutic agent, or combination of a therapy or therapeutic agent with at least another therapy or additional therapeutic agent.
[0156] Additionally, a health care provider may, for example, approve or deny prescription of a therapy, approve or deny coverage for a therapy, approve or deny reimbursement for treatment costs, and determine or deny eligibility for a therapy.
[0157] In some aspects, a clinical laboratory may, for example, collect or obtain a sample, process a sample, submit a sample, receive a sample, transport a sample, analyze or measure a sample, quantify a sample, provide results obtained after analyzing / measuring / quantifying a sample, receive results obtained after analyzing / measuring / quantifying a sample, compare / score results obtained after analyzing / measuring / quantifying one or more samples, provide a comparison / score from one or more samples, obtain a comparison / score from one or more samples, or other related activities.
[0158] Using the methods, systems, tangible computable readable devices, etc. disclosed herein, variants (e.g., protein missense mutations) in variant molecules (e.g., variant proteins) can be classified, for example, as pathogenic or non-pathogenic, and such classification can then be used to (i) treat a patient, (ii) select a patient for treatment, (iii) initiate treatment, (iv) discontinue treatment, (v) interrupt treatment, (vi) modify treatment, or (vii) any combination thereof.
[0159] The present disclosure provides methods of (i) treating a patient, (ii) selecting a patient for treatment, (iii) initiating treatment, (iv) ceasing treatment, (v) interrupting treatment, (vi) modifying treatment, or (vii) any combination thereof, including applying the methods, systems, tangible computable readable devices, etc. disclosed herein.
[0160] The present disclosure also provides methods of determining whether to treat a subject with a particular therapy, including applying the methods, systems, tangible computable readable devices, etc. disclosed herein. For example, determining that a variant molecule (e.g., a variant protein) is likely to be pathogenic and result in the manifestation of a particular symptom can be used to select a particular therapy that will treat or ameliorate such symptom.
[0161] The present disclosure also provides for personalized medicine treatments, including: (i) identifying a particular variant molecule (e.g., a variant protein) in a subject; (ii) determining the phenotypic impact of the particular variant molecule (e.g., a variant protein) using a method disclosed herein, e.g., using an ML predictive model to diagnose and / or predict the phenotypic impact of the variant molecule (e.g., a variant protein such as a missense variant protein), where the workflow used generates such an ML predictive model or associates, determines a library (e.g., a library of precomputed MD models, a library of low dimensional conformational landscapes, a library of low dimensional states, or a library of ML predictive models); and (iii) selecting a treatment to treat, prevent, or ameliorate a disease or condition associated with the variant molecule.
[0162] In one aspect, the method disclosed herein includes making a diagnosis, which may be a differential diagnosis, based at least in part on a classification or prediction of a phenotypic impact of a variant molecule (e.g., a variant protein) in a subject, e.g., by a predictive model of the present disclosure, where the variant molecule (e.g., a variant protein) has been classified by application of the ML predictive model disclosed herein. This diagnosis may be recorded in the patient's medical record. The medical record may be in paper form and / or may be maintained on a computer readable medium. The medical record may be maintained by a laboratory, a clinic, a hospital, a health care maintenance organization, an insurance company, and / or a personal medical record website.
[0163] Factors known in the art to diagnose and / or suggest, select, prescribe, recommend, or otherwise determine a course of treatment for a patient or class of patients having or at risk of having a disease or disorder can be used, for example, in combination with measurements of expression of a reference molecule or variant molecule (e.g., RNA or protein expression levels of a reference or variant protein), as well as in combination with pathogenicity classifications, predictive models, or other output from the methods disclosed herein.
[0164] The present disclosure also provides a portfolio of pre-calculated variant molecular structures (e.g., variant protein structures) or pre-calculated pathogenicity scores for variant molecules (e.g., variant proteins). Such portfolios can be generated using the methods disclosed herein.
[0165] In some aspects, the diagnosis based on application of the ML predictive models disclosed herein may be recorded on or in a medical alert article, such as a card, a wearable article, and / or a radio frequency identification (RFID) tag. As used herein, the term "wearable article" refers to any article that can be worn on the subject's body, including, but not limited to, a tag, a bracelet, a necklace, or an armband.
[0166] In some aspects, the methods disclosed herein, e.g., methods for generating ML predictive models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants), workflows used to generate such ML predictive models, and / or associated libraries (e.g., libraries of pre-computed conformational dynamics (e.g., MD) models, libraries of low-dimensional conformational landscapes, libraries of low-dimensional states, or libraries of ML predictive models), can be used to design or select therapeutic agents to treat, prevent, or ameliorate a disease or condition associated with a reference molecule (e.g., a reference protein) or a variant thereof. For example, pre-computed MD simulation structures of a reference molecule or a variant molecule can be used to identify new or existing drugs that may optimally interact with the reference molecule or the variant molecule by a docking program.
[0167] In some aspects, a portfolio of pre-computed variant molecules (e.g., pre-computed protein structures), pre-computed pathogenicity scores of variant molecules, generated models or libraries of models as disclosed herein can be used to select patients for clinical trials.
[0168] III. Reference and variant molecules and indications The methods described herein can be used with missense variant proteins to provide information for predicting, diagnosing, and treating various diseases (e.g., cancer, neurodegenerative diseases, and genetic diseases). Examples of cancer include breast cancer, ovarian cancer, lung cancer, brain cancer, leukemia, prostate cancer, and pancreatic cancer. Examples of neurodegenerative diseases include Alzheimer's disease, Parkinson's disease, and Huntington's disease. Genetic diseases include diseases and disorders caused by mutations in a single gene (monogenic diseases) or multiple genes (polygenic diseases), a combination of mutations and environmental factors, or damage to chromosomes (changes in the overall number or structure of chromosomes). In monogenic diseases, the mutations can be present in one or both chromosomes. Examples of monogenic diseases include Marfan syndrome, Tay-Sachs disease, sickle cell anemia, cystic fibrosis, and polycystic kidney disease. Polygenic diseases are caused by a combination of small genetic mutations in multiple genes. Examples of polygenic diseases include cancer, heart disease, diabetes, and autoimmune diseases.
[0169] In some embodiments, the reference molecule of the present disclosure may be a polynucleotide encoding a gene disclosed below, or a domain thereof. In other embodiments, the reference molecule of the present disclosure may be a protein encoded by a gene disclosed below, or a domain thereof. In some embodiments, the reference molecule may be a complex comprising at least one protein encoded by a gene disclosed below, or a domain thereof.
[0170] In some aspects, the variant molecules of the disclosure can be polynucleotides encoding genes disclosed below, or domains thereof, that contain one or more variants (e.g., mutations) with respect to the canonical sequence of the gene. In other aspects, the variant molecules of the disclosure can be proteins encoded by genes disclosed below, that contain one or more variants (e.g., mutations) with respect to the canonical sequence of the protein or domains thereof. In some aspects, the variant molecules can be complexes that contain at least one protein encoded by a gene disclosed below, or domains thereof.
[0171] The full names and sequences of the genes disclosed below, and the full names and sequences of their corresponding proteins, can be found in their corresponding Uniprot (www.uniprot.org) and RefSeq (www.ncbi.nlm.nih.gov / refseq) entries in the latest database releases available as of September 9, 2021, the entire contents of which are incorporated by reference herein.
[0172] In some embodiments, the reference molecule (e.g., a reference protein) is selected from the group consisting of SLC12A3, MLH1, GAA, MSH2, LDLR, COL6A2, TP53, GLDC, KCNQ1, COL6A1, CAPN3, MUTYH, VHL, TGM1, IGHMBP2, PYGM, ABCD1, PMS2, CLCN1, CHEK2, MFN2, WFS1, SLC22A5, and CPT2. In some embodiments, the reference molecule (e.g., a reference protein) is MSH2, a domain thereof, or a complex comprising MSH2, e.g., MSH2 / MSH6.
[0173] Specific proteins that can be used as reference molecules to predict the phenotypic impact caused by the variants are disclosed below. In some embodiments, the reference molecule (e.g., reference protein) or variants thereof (i.e., variant molecules such as missense variant proteins) are related or associated with a Mendelian disorder, and the reference molecule or variants thereof ... 1, TSC2, NF2, PTEN, RB1, RYR1, GLA, RYR2, TGFBR1, TGFBR2, ACTC1, CACNA1S, COL3A1, DSC2, DSG2, DSP, FBN1, MEN1, MYL2, MYL3, PKP2, PMS2, PRKAG2, SMAD3, TMEM43, TPM1, WT1, BMPR1A, SMAD4, ATP7B, and OTC. The term "associated with" means that there is a relationship or correlation between the items discussed. For example, a particular protein or gene may be associated with a disease. This would mean that the protein or gene is associated with or correlated to the presence of the disease. This could also mean that the protein or gene has been shown to be responsible for the disease, in whole or in part.
[0174] In some aspects, a gene may appear to be a driver gene based on the results of training a machine learning model (trained predictive model). The term "driver gene" as used herein refers to a gene that contains a driver gene mutation. In some aspects, a driver gene is a gene in which one or more acquired mutations (e.g., driver gene mutations) may be causally related to, for example, cancer progression. In some aspects, a driver gene may regulate one or more cellular processes including cell fate determination, cell survival, and genome maintenance. A driver gene may be associated with (e.g., may regulate) one or more signal transduction pathways, such as the TGF-β pathway, the MAPK pathway, the STAT pathway, the PI3K pathway, the RAS pathway, the cell cycle pathway, the apoptosis pathway, the Notch pathway, the Hedgehog (HH) pathway, the APC pathway, the chromatin modification pathway, the transcriptional regulation pathway, the DNA damage control pathway, or combinations thereof. Exemplary driver genes include oncogenes and tumor suppressors. In some aspects, a driver gene provides a selective growth advantage to the cells in which it occurs. In some embodiments, the driver gene provides proliferation capability to the cell in which it occurs, e.g., allowing proliferation (e.g., clonal expansion) of the cell. In some embodiments, the driver gene is an oncogene. In some embodiments, the driver gene is a tumor suppressor gene (TSG).
[0175] In some embodiments, the reference molecule (e.g., a reference protein) or variant molecule (e.g., a missense variant protein) is a "cancer driver," and the reference molecule or variant molecule is one of the following: TP53, PIK3CA, ARID1A, RB1, PTEN, KRAS, BRAF, CDKN2A, NRAS, FBXW7, STAG2, NFE2L2, NF1, IDH1, ATM, PIK3R1, CASP8, HRAS, MLL2, SF3B1, ERBB2, CREBBP, AKT1, HLA-A, CTCF, ERBB3, CTNNB1, RUNX1, MYD 88, SMARCA4, EP300, SETD2, SMARCB1, EGFR, TBL1XR1, U2AF1, EZH2, RAC1, MLL3, IL7R, CD79B, POU2AF1, MAP2K1, PTPN11, CCND1, MAP2K4, TCF7L2, KIT, CDK 4, FOXA1, TSC1, FAT1, WT1, BCOR, XPO1, PRDM1, KEAP1, NSD1, PPP2R1A, CDKN1B, ASXL1, MET, RPL5, MYCN, TNFRSF14, FLT3, ALK, KDM5C, KDM6A, APC, PBRM1, S TK11, RAD21, EZR, SPOP, TET2, PHF6, IRF4, DDX5, CCDC6, HIST1H3B, CARD11, IDH2, MLL, FGFR2, CDK12, ERCC2, B2M, MED12, CEBPA, NOTCH1, BRCA1, MAP3K1 , VHL, DNMT3A, FGFR3, NPM1, FAM46C, CBFB, GATA3, MYB, CDH1, BAP1, ELF3, ZNF198, MALT1, WIF1, KDR, SFRS3, MXRA5, SS18, TAL1, RXRA, TCEA1, HEAB, THRAP 3, RUNDC2A, SLC44A3, TNF, TAL2, FLJ27352, LAF4, STK19, DDX10, MSI2, NUTM2A, POU5F1, TRIP11, STAT5B, NCOA2, AZGP1, NCOA1, STAT3, NCOA4, OR52N1, CD KN2a(p14), CEP1, TFPT, SUFU, HOXA13, DB2, HOXA11, P2RY8, ECT2L, TRD@, IGH@, SMAD4, RBM10, LASP1, ROS1, KMT2D, WASF3, RBM15, PRKAR1A, KCNJ5, ATRX,EPHA2、BIRC3、HNRNPA2B1、OR4A16、NUTM2B、KLF4、MAP2K2、C15orf21、ERG、CD79A、SRGAP3、MLLT3、MITF、MN1、MLLT2、MLLT7、MLLT6、FAS、C15orf55、POU2 F2、EIF2S2、MLLT4、EPS15、HERPUD1、TBC1D12、MLLT1、ALO17、CNOT3、FIP1L1、CBL、OLIG2、HOXC13、NT5C2、ABL1、ZNF521、PLAG1、TPM4、LMO1、LMO2、BLM、NT N4, SLC4A5, IRTA1, JAK3, PMS2, ATP1A1, TERT, CDH11, PTCH, DDX3X, HEY1, MORC4, TLX3, PALB2, BCR, BRCA2, MDM4, MDM2, BRD4, TFG, CSF3R, RPL10, PER1, IT PKB, PDSS2, CREB1, AF3p21, TRIM27, WRN, KIF5B, CHD8, RAB40A, GATA1, ATIC, CD1D, SETBP1, CRTC3, TNFRSF17, COL1A1, DUX4, ACVR1B, C16orf75, NIN, ZNF 278, MAF, NF2, AKAP9, CCND2, MAX, MECT1, ARHGEF12, SEPT6, CBLB, FACL6, ALKBH6, CHN1, CBFA2T1, IL6ST, TCEB1, MEN1, FBXO11, HIST1H4I, RALGDS, BUB1B, FHIT, CRLF2, RASA1, TLX1, IGK@, SELP, TXNDC8, CACNA1D, GUSB, NUP214, NKX2-1, INPPL1, CBFA2T3, BCLAF1, TSC2, SDH5, CDC73, ZNF384, CDC27, OTUD7A , SIL, RANBP17, NDRG1, SMC3, FH, PAX7, CD273, HLA-B, PHOX2B, CD274, GNAS, GNAQ, PSIP1, ASPSCR1, GPHN, XIRP2, PAX8, MYOCD, FRMD7, RAP1GDS1, PAX3, AJ UBA, SLC34A2, HLF, UBR5, REL, RPS2, GNA11, LHFP, TBX3, SMO, RET, PAPD5, RPS15, SS18L1, MYH11, EIF4A2, LCK, XPA, HSPCA, PPARG, CHIC2, HOXC11, H3F3B,<h2 style=";text-align:left;direction:ltr">JAK2、TFRC、ZNF620、SOX17、MTCP1、JUN、LCTL、TAF15、NONO、SRSF2、CHCHD7、 MAML2、PPM1D、DAXX、H3F3A、JAK1、RIT1、CCND3、TRRAP、MED23、IGL@、SPEN、DI APH1、CMKOR1、ZNF471、STL、POLE、MAP4K3、ING1、FOXO1A、LIFR、CHEK2、LCP1 、AKT2、TPR、NFKB2、FOXL2、COL5A1、FEV、HMGA1、BCL3、HMGA2、CARS、PCSK7、EL L、GMPS、LYL1、BMPR1A、TGFBR2、SLC45A3、GRAF、HLXB9、HIST1H1E、DIS3、WWT R1、PDGFRA、PDE4DIP、ARID5B、ALDH2、STX2、SACS、ARNT、GOPC、SOS1、ITK、DIC ER1、KEL、CIC、RAB5EP、FVT1、PML、ADNP、FANCA、ABL2、C12orf9、BRIP1、MALA T1、FANCD2、PAFAH1B2、MUTYH、POT1、JAZF1、GNPTAB、FGFR1OP、RAD51L1、DNER 、ZNF331、CD70、IKZF1、NCOR1、MLF1、MYH9、SYK、HCMOGT-1、FANCE、FANCF、FA NCG、TPM3、NUP210L、INTS12、SDHC、RUNXBP2、BTG1、TTLL9、EML4、SDHB、CDK6、 PMX1、PDGFRB、FOXO3A、NTRK1、CLTCL1、SH2B3、EBF1、GPC3、FGFR1、ETV6、NR4 A3、SBDS、PIM1、ALPK2、PDGFB、CUL4B、YWHAE、ETV1、BCL10、PBX1、IL21R、CREB 3L1、ATF1、FANCC、C2orf44、HSPCB、CANT1、PTPRC、WAS、NFIB、CREB3L2、AF1Q 、NOTCH2、ABI1、SH3GL1、NBS1、OMD、SUZ12、TRA@、AF5q31、RSBN1L、BCL11B、MS H6、ERCC5、BCL11A、ERCC3、MSH2、NUMA1、KTN1、TFE3、IL2、MYCL1、LPP、HOXA9 、RPL22、MSN、EVI1、BCL7A、AXIN1、NBPF1、ZNF9、MLH1、SFRS2、TRIM33、SIRT4、AXIN2, CIITA, ARHGAP35, SET, ELF4, HIP1, MSF, SOX2, FNBP1, CD74, TCL1A, RAF1, MADH4, COPEB, FLI1, CBLC, GATA2, EXT1, EXT2, MICALCL, DDI T3, D10S170, CDKN2C, MYC, GOLGA5, TRIM23, NTRK3, KLK2, SLC1A3, PRF1, ACSL3, NUP98, ELK4, CYLD, TMPRSS2, DDX6, CCNB1IP1, TTL, ZNF750, T IF1, SOCS1, PNUTL1, FOXQ1, ATP2B3, PMS1, FSTL3, PCBP1, KDM5A, ZNF145, PICALM, EWSR1, AF15Q14, BCL6, GNA13, BCL5, BCL9, ANK3, RHEB, BHD ,QKI,PPP6C,CALR,PRCC,FCGR2B,BCL2,RPN1,SSX4,MDS2,TPX2,RARA,ZFHX3,TRB@,MDS1,MAFB,SLC26A3,SGK1,SDHD,CDX2,SSX1,ZRANB3,K IAA1549, SSX2, HOOK3, MTOR, SNX25, TCF1, MGA, LRIG3, PRDM16, ELKS, RHOA, ACO1, ELN, VTI1A, BRD3, MLLT10, RNF43, CDKN1A, ARID2, LCX, TFE B, WHSC1L1, ETV5, ETV4, HOXD11, GAS7, ARHH, IPO7, GOT1, SMAD2, WHSC1, TNFAIP3, TCL6, HOXD13, SDC4, PAX5, MPL, MPO, SFPQ, TCF3, NACA, REC It is encoded by a gene selected from the group consisting of QL4, SMC1A, ERCC4, TCF12, KLHL8, DNM2, CLTC, SMARCE1, DEK, XPC, USP6, FUBP1, PCM1, TRAF7, ZRSR2, FUS, FOXP1, FLG, TOP1, MUC1, TCP11L2, COX6C, MYST4, MUC17, CAMTA1, C3orf70, CUX1, CAP2, TRAF3, MKL1, CCNE1, TSHR, AMER1, CCDC120, CHD4, and TAP1.
[0176] In some embodiments, the reference molecule (e.g., a reference protein) or variant molecule (e.g., a missense variant protein) is a "pharmacogenomic target protein," and the reference molecule or variant molecule is selected from the group consisting of A2M, ABAT, ABCA1, ABCA12, ABCA3, ABCA8, ABCB1, ABCB11, ABCB4, ABCB5, ABCB6, ABCB9, ABCC1, ABCC10, ABCC11, ABCC2, ABCC3, ABCC4, ABCC5, ABCC6, ABCC8, ABCC9, ABCD1, ABCD2, ABCG1, ABCG2, , ABCG8, ABL1, ABO, ACBD4, ACE, ACE2, ACHE, ACP5, ACSS2, ACTG1, ACY3, ACYP2, ADA, ADAM12, ADAM33, ADAMTS1, ADAMTS14, ADCK4, ADCY2, ADCY9, ADD1, ADH 1A, ADH1B, ADH1C, ADH7, ADIPOQ, ADK, ADM, ADORA1, ADORA2A, ADORA2A-AS1, ADRA1A, ADRA2A, ADRA2B, ADRA2C, ADRB1, ADRB2, ADRB3, ADRBK2, AFAP1L1, AG AP1, AGBL4, AGO1, AGT, AGTR1, AGXT, AHR, AIDA, AK4, AKR1C3, AKR1C4, AKR7A2, AKT1, AKT2, ALDH1A1, ALDH1A2, ALDH2, ALDH3A1, ALDH5A1, ALG10, ALOX12, ALOX15, ALOX5, ALOX5AP, AMHR2, AMPD1, ANGPT2, ANGPTL4, ANKFN1, ANKK1, ANKRD55, ANKS1B, ANXA11, AOX1, APBB1, APEH, APLF, APOA1, APOA4, APOA5, APO B, APOBEC2, APOC1, APOC3, APOE, APOH, AQP2, AQP9, ARAP1, ARAP2, AREG, ARG1, ARHGEF10, ARHGEF4, ARID5B, ARMS2, ARNT, ARNTL, ARRB2, ARVCF, AS3MT, AS IC2, ASPH, ASS1, ATF3, ATG16L1, ATG5, ATIC, ATM, ATP2B1, ATP5E, ATP7A, ATP7B, AXIN2, B4GALT2, BACH1, BAD, BAG6, BAZ2B, BCAP31, BCHE, BCL2, BCL2L11,BCR, BDKRB1, BDKRB2, BDNF, BDNF-AS, BGLAP, BLK, BLMH, BMP5, BMP7, BRAF, BRD2, BTG4, BTRC, C10orf107, C10orf11, C11orf30, C11orf65, C12orf40, C17orf51, C18orf21, C18orf56, C1orf167, C2, C20orf194, C3, C5, C5orf22, C8orf34, C9orf72, CA10, CA12, CACNA1A, CACNA1C, CACNA1E, CACNA1H, CACNA1 S, CACNB2, CACNG2, CALU, CAMK1D, CAMK2N1, CAMK4, CAP2, CAPG, CAPN10, CAPZA1, CARD16, CARTPT, CASP1, CASP3, CASP7, CASP9, CASR, CAT, CBR1, CBR3, CBS, CCDC22, CCHCR1, CCL2, CCL21, CCND1, CCNH, CCNY, CCR5, CD14, CD28, CD38, CD3EAP, CD40, CD58, CD69, CD74, CD84, CDA, CDC5L, CDCA3, CDH13, CDH4, CD K1、CDK4、CDK9、CDKAL1、CDKN2B-AS1、CELF4、CELSR2、CEP68、CEP72、CERKL、CERS6、CES1、CES1P1、CES2、CETP、CFAP44、CFB、CFH、CFI、CFLAR、CFTR、CHAT 、CHIA、CHIC2、CHL1、CHRM2、CHRM3、CHRM4、CHRNA1、CHRNA3、CHRNA4、CHRNA5、CHRNA7、CHRNB1、CHRNB2、CHRNB3、CHRNB4、CHST13、CHST3、CHUK、CLASP1、CL CN6、CLMN、CLNK、CLOCK、CMPK1、CNKSR3、CNOT1、CNPY4、CNR1、CNTF、CNTN4、CNTN5、CNTNAP2、COL18A1、COL1A1、COL1A2、COL22A1、COL26A1、COLEC10、COMT 、COQ2、CPA2、CPS1、CR1、CR1L、CREB1、CRH、CRHR1、CRHR2、CRP、CRTC2、CRY1、CSK、CSMD1、CSMD2、CSMD3、CSNK1E、CSPG4、CSRNP3、CSRP3、CST5、CTH、CTLA4、<h2 style=";text-align:left;direction:ltr">CTNNA2、CTNNA3、CTNNB1、CUX1、CUX2、CXCL10、CXCL12、CXCL5、CXCL8、CXCR2 、CXCR4、CXXC4、CYB5A、CYB5R3、CYBA、CYCSP5、CYP11B2、CYP19A1、CYP1A1、C YP1A2、CYP1B1、CYP24A1、CYP27B1、CYP2A6、CYP2B6、CYP2B7P1、CYP2C18、CY P2C19、CYP2C8、CYP2C9、CYP2D6、CYP2E1、CYP2J2、CYP2R1、CYP39A1、CYP3A、C YP3A4、CYP3A43、CYP3A5、CYP3A7、CYP4A11、CYP4B1、CYP4F11、CYP4F2、CYP5 1A1、CYP7A1、DAOA、DAPK1、DBH、DCAF4、DCBLD1、DCK、DCP1B、DCTD、DDC、DDHD1 、DDRGK1、DDX20、DDX53、DDX58、DEAF1、DGCR5、DGKH、DGKI、DHFR、DHODH、DIA PH3、DIO1、DIO2、DKK1、DLEU7、DLG5、DLGAP1、DMPK、DNAH12、DNAJB13、DNMT3A 、DOCK4、DOK5、DOT1L、DPP4、DPYD、DPYS、DRD1、DRD2、DRD3、DRD4、DROSHA、DS CAM、DTNBP1、DUSP1、DUX1、DYNC2H1、E2F7、EBF1、ECT2L、EDN1、EGF、EGFR、EG LN3、EHF、EIF2AK4、EIF3A、EIF4E2、ENG、ENOSF1、EPAS1、EPB41、EPHA5、EPHA 6、EPHA8、EPHX1、EPM2A、EPM2AIP1、EPO、ERAP1、ERBB2、ERCC1、ERCC2、ERCC3、 ERCC4、ERCC5、ERCC6L2、EREG、ERICH3、ESR1、ESR2、ETS2、EXO1、F11、F12、F1 3A1、F2、F3、F5、F7、FAAH、FABP1、FABP2、FADS1、FAM19A5、FAM65B、FARS2、FAS 、FASLG、FASTKD3、FAT1、FBXL17、FBXL19、FCAR、FCER1A、FCER1G、FCER2、FCG R2A、FCGR2B、FCGR3A、FDPS、FEN1、FGD4、FGF2、FGF5、FGFBP1、FGFBP2、FGFR2、FGFR4、FHIT、FKBP5、FLOT1、FLT1、FLT3、FLT4、FMO1、FMO2、FMO3、FMO5、FNTB 、FOLH1、FOLR3、FOXC1、FOXP3、FPGS、FSHR、FSIP1、FSTL5、FTO、FYN、FZD3、FZD 4、G6PD、GABRA1、GABRA3、GABRA6、GABRB1、GABRB2、GABRG2、GABRG3、GABRP、 GABRQ、GAD2、GADL1、GAL、GALNT14、GALNT18、GALNT2、GALR1、GAPDHP64、GAPV D1, GATA3, GATA4, GATM, GBP6, GCG, GCKR, GCLC, GDNF, GEMIN4, GFRA2, GGCX. GGH、GHSR、GIPR、GJA1、GLCCI1、GLDC、GLP1R、GLRB、GNAS、GNB3、GNMT、GP1BA、 GP6、GPR1、GPR83、GPX1、GPX3、GPX5、GRIA1、GRIA3、GRID2、GRIK1、GRIK2、GR IK3、GRIK4、GRIN1、GRIN2A、GRIN2B、GRIN3A、GRK4、GRK5、GRM3、GRM7、GSK3B、 GSR、GSTA1、GSTA2、GSTA5、GSTM1、GSTM3、GSTM4、GSTP1、GSTT1、GSTZ1、H19、 HAS3, HCG22, HCP5, HDAC1, HES6, HFE, HIF1A, HLA-A, HLA-B, HLA-C, HLA-DOB. HLA-DPA1、HLA-DPB1、HLA-DPB2、HLA-DQA1、HLA-DQB1、HLA-DRA、HLA-DRB1、 HLA-DRB3, HLA-DRB5, HLA-E, HLA-G, HMGB1, HMGB2, HMGCR, HNF1A, HNF1B, HNF 4A, HNMT, HOMER1, HOTAIR, HOTTIP, HRH1, HRH2, HRH3, HRH4, HS3ST4, HSD11B 1、HSD3B1、HSPA1A、HSPA1L、HSPA5、HSPG2、HTR1A、HTR1B、HTR1D、HTR2A、HTR2 C、HTR3A、HTR3B、HTR5A、HTR6、HTR7、HTRA1、HUS1、HYKK、IBA57、IDO1、IFIT1、 IFNAR1、IFNB1、IFNG、IFNGR1、IFNGR2、IFNL3、IFNL4、IGF1、IGF1R、IGF2BP2、<h2 style=";text-align:left;direction:ltr">IGF2R、IGFBP3、IGFBP7、IKBKG、IKZF3、IL10、IL11、IL12A、IL12B、IL13、IL1 6、IL17A、IL17F、IL17RA、IL18、IL1A、IL1B、IL1RN、IL2、IL21R、IL23R、IL27、 IL2RA、IL2RB、IL3、IL4、IL4R、IL6、IL6R、IL6ST、IL7R、ILKAP、IMPA2、IMPDH 1、IMPDH2、INSIG2、INSR、IP6K2、IRS1、ITGA1、ITGA2、ITGA9、ITGB1、ITGB3、I TGBL1、ITIH3、ITPA、ITPKC、JAK2、KANSL1、KCNE1、KCNH2、KCNH7、KCNIP1、KC NIP4、KCNJ1、KCNJ11、KCNJ6、KCNMA1、KCNMB1、KCNQ1、KCNQ5、KCNT1、KCNT2、K DM4A、KDR、KIAA0391、KIF6、KIR2DL2、KIRREL2、KIT、KL、KLC1、KLC3、KLRC1、 KLRD1、KLRK1、KRAS、KYNU、LAMB3、LARP1B、LCE3B、LCE3C、LDLR、LECT2、LEP、L EPR、LGALS3、LGR5、LIG3、LINC00251、LINC00478、LIPC、LPA、LPHN3、LPIN1、 LPL、LRP1、LRP1B、LRP2、LRP5、LRRC15、LST1、LTA、LTA4H、LTB、LTC4S、LUC7L2 、LYN、LYRM5、MAD1L1、MAFB、MAFK、MALAT1、MAML3、MAN1B1、MAP3K1、MAP3K5、 MAP4K4、MAPK1、MAPK14、MAPT、MC1R、MC4R、MCPH1、MDGA2、MDM2、MDM4、MECP2、 MED12L、MEG3、MET、METTL21A、MEX3C、MGAT4A、MGMT、MIA3、MICA、MICB、MIR1 206、MIR1307、MIR133B、MIR146A、MIR2053、MIR27A、MIR300、MIR423、MIR427 8, MIR449B, MIR492, MIR577, MIR595, MIR604, MIR611, MIR618, MIR7-2, MISP, MLLT3, MLN, MME, MMP1, MMP10, MMP2, MMP3, MMP9, MOB3B, MOCOS, MOV10, MPO,MPZ、MS4A2、MSH2、MSH3、MSH6、MT-RNR1、MTCL1、MTHFD1、MTHFR、MTMR12、MTOR、MTR、MTRF1L、MTRR、MTTP、MUC5B、MUTYH、MVK、MYC、MYLIP、MYOCD、N6AMT、 <h2 style=";text-align:left;direction:ltr">1、NALCN、NANOGP6、NAT1、NAT2、NAV2、NBAS、NBEA、NCF4、NCOA1、NCOA3、NEDD 4、NEDD4L、NEFM、NELFCD、NELL1、NEUROD1、NFATC1、NFATC2、NFE2L2、NFKB1、 NFKBIA, NGF, NGFR, NLGN1, NLRP3, NLRP8, NOD2, NOS1AP, NOS2, NOS3, NPAS3, NPC1L1, NPHS1, NPPA, NPPA-AS1, NQO1, NQO2, NR1D1, NR1H3, NR1I2, NR1I3, NR 3C1, NR3C2, NRAS, NRG1, NRG3, NRP1, NRP2, NRXN1, NT5C1A, NT5C2, NT5C3A, NT5E, NTRK1, NTRK2, NUBPL, NUDT15, NUMA1, OAS1, OASL, OCRL, OPN1SW, OPRD1 OPRK1、OPRM1、OR10AE3P、OR4D6、OR52E2、OR52J3、ORM1、ORM2、ORMDL3、OSMR 、OTOS、OXT、P2RY1、P2RY12、PACSIN2、PADI4、PAPD7、PAPLN、PAPPA2、PARD3B、 PARP11、PAX4、PCK1、PCSK9、PDCD1LG2、PDE4B、PDE4C、PDE4D、PDGFRA、PDGFR B、PDLIM5、PDZRN3、PEAR1、PEMT、PER2、PER3、PGLYRP4、PGR、PHACTR1、PHB2、P HTF1、PI4KA、PICALM、PICK1、PIGB、PIK3CA、PIK3R1、PITPNM2、PKLR、PLA2G4 A、PLAGL1、PLCB1、PLCD3、PLCG1、PLEKHH2、PLEKHN1、PLG、PLXNB3、PMCH、POLA 2、POLG、POLR3G、POMT2、PON1、PON2、POR、POU2F1、POU2F2、POU5F1、PPARA、P PARD、PPARG、PPARGC1A、PPFIA1、PPM1A、PPP1R13L、PPP1R1C、PPP2R5E、PRB2、 PRCP、PRDM1、PRDM16、PRDX4、PRIMPOL、PRKAA1、PRKAA2、PRKCA、PRKCB、PRKC E、PRKCQ、PRKG1、PROC、PROCR、PROM1、PROS1、PROX1、PRRC2A、PRSS53、PSMA4、<h2 style=";text-align:left;direction:ltr">PSMB3P、PSMB4、PSMB8、PSMD14、PSORS1C1、PSORS1C3、PSRC1、PTCHD1、PTEN、 PTGER2、PTGER3、PTGER4、PTGES、PTGFR、PTGIR、PTGS1、PTGS2、PTH、PTH1R、P TPN22、PTPRC、PTPRD、PTPRM、PTPRN2、PYGL、RAB27A、RABEPK、RAC2、RAD18、R AD52、RAF1、RALBP1、RAPGEF5、RARG、RARS、RBFOX1、RBMS3、REEP5、REL、REN、 REPS1、RET、REV1、REV3L、RFK、RGS17、RGS2、RGS4、RGS5、RHBDF2、RHOA、RICT OR、RND1、RNFT2、RORA、RPL13、RRAS2、RRM1,RRM2、RRM2B、RSBN1、RSRP1、RUN X1, RXRA, RYR1, RYR2, RYR3, SACM1L, SCAP, SCARB1, SCGB3A1, SCN10A, SCN1A 、SCN2A、SCN4A、SCN5A、SCN8A、SCN9A、SCNN1B、SCNN1G、SELE、SELP、SEMA3C、S ERPINA3、SERPINA6、SERPINE1、SERPINF1、SERPING1、SETD4、SFRP5、SH2B3、 SH2D5、SH3BP2、SHMT1、SIK3、SIN3A、SKIV2L、SKOR2、SLC10A2、SLC12A3、SLC 12A8、SLC14A2、SLC15A1、SLC15A2、SLC16A5、SLC16A7、SLC17A3、SLC18A2、S LC19A1、SLC1A1、SLC1A2、SLC1A3、SLC1A4、SLC22A1、SLC22A11、SLC22A12、S LC22A16, SLC22A17, SLC22A2, SLC22A3, SLC22A4, SLC22A5, SLC22A6, SLC22A7, SLC22A8, SLC24A4, SLC25A13, SLC25A14, SLC25A27, SLC25A31, SLC26A9 SLC28A1, SLC28A2, SLC28A3, SLC29A1, SLC2A1, SLC2A2, SLC2A9, SLC30A8, SLC30A9, SLC31A1, SLC37A1, SLC39A14, SLC47A1, SLC47A2, SLC5A2, SLC5A7<h2 style=";text-align:left;direction:ltr">SLC6A12、SLC6A2、SLC6A3、SLC6A4、SLC6A5、SLC6A9、SLC7A5、SLC7A8、SLCO1 A2、SLCO1B1、SLCO1B3、SLCO1C1、SLCO2B1、SLCO3A1、SLCO4C1、SLCO6A1、SLIT 1、SMARCAD1、SMYD3、SNAP25、SNORA59B、SNORD68、SOCS3、SOD2、SOD3、SORT1 、SOX10、SP1、SPARC、SPATS2L、SPECC1L、SPG7、SPIDR、SPINK5、SPP1、SPTA1、S QSTM1、SREBF1、SREBF2、SRP19、SRR、ST13、STAT3、STAT4、STAT6、STIM1、STI P1、STK39、STMN1、STMN2、STX1B、STX4、SUGCT、SULT1A1、SULT1A2、SULT1C4、S ULT1E1、SULT2B1、SV2C、SYN3、SYNE3、SZRD1、T、TAAR6、TAC1、TAGAP、TANC1、 TANC2、TAP1、TAP2、TAPBP、TAS2R16、TBC1D1、TBC1D32、TBX21、TBXA2R、TBXAS 1, TCF19, TCF7L2, TCL1A, TDP1, TDRD6, TERT, TET2, TF, TGFB1, TGFBR2, TGFBR3, TH, THBD, THRA, THRB, TIGD1, TK1, TLR2, TLR3, TLR4, TLR5, TLR7, TLR9, TM CC1、TMCO6、TMEFF2、TMEM205、TMEM258、TMEM57、TMPRSS11E、TNF、TNFAIP3、TNFRSF10A、TNFRSF11A、TNFRSF11B、TNFRSF1A、TNFRSF1B、TNFSF10、TNFSF11 、TNFSF13B、TNRC6A、TNRC6B、TOLLIP、TOMM40、TOMM40L、TOP1、TOP2B、TP53、 TPH1、TPH2、TPMT、TRAF1、TRAF3IP2、TRIB3、TRIM5、TRPM6、TSC1、TSPAN5、TTC 6、TUBB1、TUBB2A、TXNRD2、TYMP、TYMS、UBASH3B、UBE2I、UCP2、UCP3、UGGT2、 UGT1A、UGT1A1、UGT1A10、UGT1A3、UGT1A4、UGT1A5、UGT1A6、UGT1A7、UGT1A8、UGT1A9, UGT2B10, UGT2B15, UGT2B17, UGT2B4, UGT2B7, ULK3, UMPS, UPB1, USH2A, USP24, USP5, UST, VAC14, VAS P, VDR, VEGFA, VKORC1, WBP2NL, WBSCR17, WDR7, WIF1, WNK1, WNT5B, WT1, WWOX, XBP1, XDH, XPA, XPC, XPO1, XPO5 , XRCC1, XRCC3, XRCC4, XRCC5, YAP1, YBX1, YEATS4, ZBTB22, ZBTB4, ZCCHC6, ZFP91-CNTF, ZMAT4, ZNF100, ZNF215, ZNF423, ZNF432, ZNF652, ZNF697, ZNF804A, ZNF816, ZNRD1-AS1, and ZSCAN25.
[0177] In some aspects, the reference molecule (e.g., reference protein) or variant molecule (e.g., missense variant protein) is selected from the group consisting of LMNA, PTEN, TP53, BRCA2, MLH1, MSH2, BRCA1, MSH6, FGFR3, MECP2, CFTR, RET, PTPN11, SCN5A, MYH7, CAV3, PMS2, KRAS, APC, ATM, ARX, DMD, DES, STK11, POLG, NF1, BRAF, TSC1, CDKL5, TSC2, TTN, COL2A1, FMR1, FKTN, KCNQ1, VHL, SLC2A1, FBN1, EPCAM, HRAS, PALB2, RAF1, TNNT2, CEP290, SMAD4, MUTYH, SCN1A, SCN1B, KCNJ2, RYR2, GLA, CDH1, NRAS, FKRP, KCNH2, LDB3, CACNA1A, MYBPC3, FGFR 2, UBE3A, CACNA1C, GJB2, TAZ, SDHB, TNNI3, ACTC1, GAA, TCAP, CHEK2, LAMP2, COL1A1, TTR, DSP, HBB, SDHD, SOS1, NBN, COL1A2, TGFBR2, POMT1, TPM1, FLNA , KCNE1, PCDH19, MAP2K1, CHD7, FOXG1, SDHC, TGFBR1, RYR1, MTHFR, SGCD, CDKN2A, PMP22, POMT2, FH, WT1, EMD, SCN4A, FGFR1, PLP1, PAX6, POMGNT1, TMEM 43, MEN1, PKP2, SLC9A6, RHO, F5, GCK, BRIP1, TRIM32, DSG2, RAD51C, TRPV4, SCN2A, CPT2, KCNE2, GJB6, COL3A1, MAP2K2, NPHP1, DNM2, BMPR1A, PRKAG2, AC ADM, OFD1, MYOT, CASQ2, HEXA, DSC2, MEF2C, HFE, CLN3, PTCH1, CRYAB, JUP, PLN, MED12, ZEB2, FHL1, ABCC8,F2, ACADVL, BAG3, ATP7A, CASR, SCN9A, BSCL2, PDHA1, SHOC2, ETFDH, KCNQ2, HADHA, TNNC1, PRRT2, TPP1, ANO5, COL5A1, ETFB, MPZ, ETFA, ACTA1, PPT1, CASK, STXBP1, ABCD1, KCNJ11, ATRX, GNAS, ABCA4,<h2 style=";text-align:left;direction:ltr">DYSF、ABCC9、TCF4、BLM、SLC22A5、SDHA、MYH6、HCN4、ATP7B、PLA2G6、FANCC、 MYL2、CBS、ANK2、KCNE3、MYL3、CLN5、DCX、PANK2、ALDH7A1、NKX2-5、GBA、TIMM 8A、PNKP、ACTA2、WFS1、MFN2、FOLR1、JAG1、SMN1、SMARCB1、L1CAM、GPC3、KIT 、NSD1、OPA1、DHCR7、NF2、SGCA、MITF、CLRN1、TPM2、SPRED1、MKS1、NIPBL、AGL OTC, RB1, CSRP3, GLB1, TMEM67, CLN6, HNF1B, SMC1A, SCN4B, CACNB2, ACVRL1, DLD, CBL, FXN, ARSA, PSEN1, COL6A3, LAMA2, SMAD3, ENG, PRPS1, ACTN2, TWN K、CAPN3、GDAP1、COL5A2、EYA1、PCDH15、GCH1、SURF1、SGCB、SCN3B、TMEM216 、PITX2、COL6A1、PEX1、MYH11、VCL、NOTCH3、LARGE1、SLC26A4、CLN8、BTD、GAM T, USH2A, MYH9, AR, NPC1, TERT, GABRG2, GCDH, HNF1A, FLNC, IDS, COL6A2, BBS1, RPGR, FLCN, GNE, RPGRIP1L, MEFV, CALM1, CDKN1C, MFSD8, PRPH2, SMPD1, O PHN1, CNTNAP2, BCKDHB, PLOD1, PLEC, CREBBP, SDHAF2, ARHGEF9, AKAP9, RAD51D, NEB, OPA3, MBD5, NPC2, MYO7A, CTSD, VPS13B, GALC, KCNJ5, PAFAH1B1, PY GM, GRN, ASPA, CDK4, PEX7, MET, FBN2, CC2D2A, GARS, NRXN1, PIK3CA, COL11A2, HTT, SLC26A2, SETX, NEXN, TGFB3, SELENON, KCNJ10, CPT1A, HPRT1, ELN, UG T1A1、WAS、OCRL、KCND3、MUT、VCP、HADHB、GPD1L、KCNQ3、SUCLA2、SCO2、FTL、E GR2、PMM2、ALPL、SNTA1、BBS2、G6PC、HADH、PKD2、PKHD1、COQ2、MMACHC、GJB1、BEST1, SGCG, BCKDHA, LDLR, NPHP3, SLC25A20, ACADS, DYNC1H1, KCTD7, MAPT FIG4 TREX1 MMAB PQBP1 GRIN2A COL4A5 MMAA MKKS RPE65 GBE1 NDP HSD17B10, GATA1, APOB, TTC8, SPG7, PDX1, GABRA1, APTX, IKBKAP, NEFL, PEX 6. COL11A1, TBC1D24, TGFB2, CRX, APOE, GUCY2D, PHOX2B, ISPD, ATP1A2, ATP1 3A2, ATL1, SYNE1, ATXN2, SLC6A8, ALMS1, HNF4A, AHI1, ACAD9, PRKAR1A, SNR PN, COL4A1, NOTCH1, SLC25A22, GLDC, ADGRV1, GALT, PEX26, TRDN, PHF6, PNPO KCNT1, MTM1, COX15, SLC4A1, RRM2B, PRSS1, TPM3, BBS10, BAP1, BCS1L, CDH 23 MRE11, PCCA, TBX5, MPL, PAH, SPTAN1, SCN8A, AMT, ASS1, PSEN2, CACNA1S. USH1C, FANCA, CYP21A2, FGD1, PEX12, SLC2A10, WDR62, FAH, GLI3, RUNX1, AN KRD1, GNPTAB, SLC25A4, SERPIN1, RELN, BARD1, RAPSN, DKC1, CSTB, SGCE, F 8. KCNJ8, MYPN, MVK, PEX10, REEP1, CRB1, CHRNA1, RBM20, PCCB, BCOR, NLRP3 HBA1, EPM2A, SKI, GATA2, MYLK, FANCB, TYR, ABCB4, C12orf65, PEX2, LRP5, T TC21B, SLC25A13, HSPB1, HSPB8, MPV17, SPAST, SLC37A4, IQCB1, IDUA, EYA4 KCNA1, PGK1, CYP1B1, WHRN, SMARCA4, TERC, ADSL, DMPK, ATXN1, ATP6AP2, SY NGAP1, RDH12, TARDBP, KMT2D, PRKN, NPHP4, TK2, NHLRC1, GJA1, SUCLG1, GAT A4, NDUFA1, COL4A3, ATXN3, VWF, TH, DBT, KIF1A, MMADHC, MID1, PKD1, AP3B1.<h2 style=";text-align:left;direction:ltr">CHRNA4、DNAJB6、APP、SHH、FA2H、CHRNB2、EDN3、SLC16A2、ELANE、FUS、INS、R PS6KA3、INVS、MYOZ2、TNNT1、ALK、TMEM70、CACNB4、JAK2、CNGB3、SPINK1、AG XT, PAX3, MCOLN1, PEX5, ASPM, DGUOK, IGHMBP2, CFH, SOD1, TUBA1A, DOLK, PROM1, SYN1, HMGCL, KDM5C, RAB39B, DNAJC5, AUH, SHOX, ATXN7, CENPJ, SRPX2, S OX10、CYP2D6、DCTN1、TBX1、ALDOB、ARL6、BBS12、COQ8A、TWIST1、RECQL4、OT X2,PC,DPAGT1,TP63,GP1BA,ARG1,POLD1,SACS,AKT1,PEX3,SMC3,OCA2,CYP 2C19、RMRP、IL2RG、DNAH5、SPG11、NDRG1、COL4A4、FOXC1、BMPR2、MCCC2、MAX 、F9、ERCC6、C9orf72、TYMP、RAI1、AIPL1、MCCC1、SLC25A19、COL9A1、BTK、P3H 1、PDSS2、PCNT、NOTCH2、ATP8B1、ATP1A3、ETHE1、HEXB、SLC25A15、CP、COL9A 2、CHRNA2、CHRNE、CUL4B、DOK7、CHRND、GUSB、SLC19A3、IVD、SH3TC2、EFHC1、 IMPDH1、CRTAP、CYP27A1、HSPD1、SOX2、SDCCAG8、CYP2C9、ALS2、RPS19、GOSR 2、RARS2、GFAP、PEX14、CYP11B1、GMPPB、BBS4、SGSH、GJC2、GLUD1、GATM、TMEM 127、RPGRIP1、PDGFRA、LGI1、MT-ATP6、ADAMTS13、BBS5、WDR45、MTMR2、GATA 6、BBS7、LITAF、POLG2、ABCB11、PRX、ALG2、ABCC6、RNASEH2B、FANCG、ADA、SIL 1、RP2、RASA1、NTRK1、TNFRSF1A、SCNN1B、CHAT、USH1G、FLNB、DNAI1、CFL2、O PTN、NDUFS4、ARL13B、BBS9、TOR1A、LRPPRC、ATPAF2、SAMHD1、TSEN54、NPHS2、<h2 style=";text-align:left;direction:ltr">TSFM、HBA2、GALNS、FKBP14、CHST14、FOXRED1、TRPM4、NHS、RNASEH2A、RNASE H2C、ADGRG1、MT-RNR1、AGK、CEP152、ASL、SNCA、GRIN2B、DTNA、SIX1、CPS1、K IF7, AIFM1, PDHX, NAGLU, MT-TL1, NSDHL, HDAC8, HGSNAT, LRRK2, SBF2, RAB7A, SCNN1G, LRAT, DARS2, KIF5A, RIT1, PCSK9, GFM1, PINK1, NPHS1, ARSB, NDUF S7、POLE、PFKM、SCN2B、IDH2、FBLN5、INPP5E、PDSS1、GABRD、ATP6V0A2、PRIC KLE1、ACAT1、SOX9、CACNA2D1、G6PD、SPG20、SCARB2、NLGN3、ANOS1、NLGN4X、 GABRB3、HAX1、AFG3L2、GJB3、TINF2、KRIT1、GPR143、CDC73、EDNRB、MLYCD、A ARS2、JAK3、SDHAF1、JPH2、NDUFV1、PEX13、PLCB1、ABHD12、PEX16、IRF6、SUMF 1、BSND、DAG1、HLCS、ATR、EGFR、AFF2、EZH2、PEX19、ABCA3、PAK3、NDUFS1、PH YH、PRKCG、TMPO、TULP1、COMP、MPI、MYLK2、HESX1、YARS、BIN1、DPM3、LYST、A ARS、SIX3、ACTG1、C19orf12、PDHBCOQ9、MLC1、NODAL、DPYD、CHM、DPM1,LIPA 、SFTPC、DLAT、VRK1、TUBB2B、ATP6V1B1、HSD17B4、CERKL、EP300、SLC12A3、GA TA3, FANCE, FGD4, CFI, SCN10A, COLQ, COX6B1, FKBP10, EXT1, ADAMTS2, SBDS, CD46, TGIF1, SALL1, ERCC4, KIF1B, SLC17A5, WNK1, KCNA5, ARFGEF2, FANCF ELOVL4、SALL4、CYP7B1、KARS、GRIA3、ALDH5A1、SPR、CLCN1、HCCS、GNS、EIF2 AK3、PUS1、PDE6B、PLOD2、PAX2、DHDDS、WDR19、ALG6、PPARG、VAPB、CHD2、RP1、PSAP、WRN、LMBRD1、INSR、CEBPA、LPIN1、SMS、MT-TK、PARK7、SUFU、UMOD、PRNP、AGA、RAD50、FUCA1、SLC39A13、NDUFA2、ISCU、MT-TS1、SEMA4A、FOXP3、TACO1、LI、 <h2 style=";text-align:left;direction:ltr">G4、AIRE、SRY、KBTBD13、EIF2B5、MT-ND1、IKBKG、DICER1、TRMU、MUSK、SLC25 A3、OTOF、POMK、TBP、RAG2、UPF3B、EDA、RLBP1、RAB3GAP1、LAMB2、CEP41、RAD 21, KDM6A, MCPH1, CABP4, SPATA7, MTRR, LAMA4, EFEMP2, NDUFS8, GALK1, SAG, LCA5, NR2E3, EXT2, GCSH, PPIB, PORCN, EHMT1, CTNNB1, CTNS, TFR2, C3, HCN1 、EIF2B1、SLX4、POU3F4、WDPCP、INF2、LIAS、CHRNB1、ACTB、AP1S2、PHEX、SPT B、NEUROD1、RS1、NPPA、SOX3、FGF23、MAN2B1、DNAH11、ERCC2、DGKE、CCM2、ND UFAF2、EVC、RAG1、HPS1、NDUFS3、NDUFS2、ZIC2、FGF8、LPL、FASTKD2、TCTN2、 CACNA1D、HPS4、CACNA1F、CLCN5、GJA5、SYP、GP1BB、FANCL、ACSL4、IDH1、CLCN KB、CISD2、ROR2、NEU1、GATAD1、MYH3、NDE1、PRPF31、ABCG5、NKX2-1、PGM1、T MEM237、FBP1、CDK5RAP2、NDUFAF5、ZFYVE26、DPM2、PHKA1、MT-ND6、STIL、TU BB3, BICD2, IQSEC2, SPTA1, ITGA7, QDPR, TJP2, PTS, EIF2B3, NOD2, GLRA1, CSF1R, PRF1, ATN1, PAX4, GPSM2, CHMP2B, CFB, EYS, FANCI, ST3GAL3, AGPAT2, P DP1、IL7R、HK1、PNPLA2、RAB27A、DCLRE1C、MC4R、GYS2、B9D1、SCNN1A、ANG、E NPP1、PRPF8、SFTPB、FANCM、AXIN2、LMX1B、NHEJ1、SYNE2、TTC19、PROP1、MAGT 1, COL7A1, FANCD2, FSCN2, NDUFAF1, MT-ND4, KCNJ1, COL12A1, CNGA3, STAT3, TYRP1, NDUFS6, GUCA1B, SLC2A2, SIX5, ADAR, SLC33A1, CCDC39, AMACR, GAN,<h2 style=";text-align:left;direction:ltr">HFE2, B3GLCT, EFNB1, UQCRB, SLC12A6, FGA, HPS3, XRCC2, MTR, C8orf37, ACTN4, EVC2, THAP1, TRPS1, IDH3B, RUNX2, LAMB3, SH2D1A, GDI1, TMC1, DNMT1, P DCD10, MRPS22, LAMA3, TOPORS, CHKB, MTPAP, CYP17A1, POMGNT2, SLC12A1, ZIC3, GLI2, RD3, ALAS2, RPL35A, CNGB1, LDLRAP1, DEPDC5, THBD, DYRK1A, SLC1 9A2、DNAI2、PGAM2、PNKD、ASAH1、WDR35、VKORC1、DOCK8、PHGDH、SLC45A2、GP 9、CCDC78、SPTLC1、IL1RAPL1、SLC35C1、UBE2A、NR0B1、CAVIN1、ACOX1、AGRN、 CA4、COL9A3、CNGA1、LAMC2、DTNBP1、EIF2B2、TTPA、FLVCR1、MYH14、ERBB2、I TGB3,VLDLR、WASHC5、NDUFA11、C2orf71、PTCHD1,NRL、ALDH4A1、RSPH9、ATP5 E、GK、CTDP1、ABL1、TCTN1、ANK1、CTSA、SLC40A1、AKT3、B4GAT1、ZMPSTE24、M ERTK、EIF2B4、ERCC8、NUBPL、PPOX、PDLIM3、PNPLA6、TNXB、PRKG1、FOXH1、CO G7, RPL11, GPHN, ABCG8, PDE6C, B4GALT7, G6PC3, GNA11, CLCN2, NME8, KCNJ13, HEPACAM, SLCO1B1, UQCRQ, NDUFAF4, TMEM138, MT-ND5, NDUFAF3, HMBS, NHP 2、IFITM5、MBTPS2、SMN2、PDE6A、VSX2、MYO6、CPOX、ALG13、CCDC40、ALDH3A2 、NIPA1、TSHR、ZNF423、SQSTM1、MOCS2、L2HGDH、SCO1、TUBB4A、TCOF1、MOCS1、 MTO1、CIB2、HINT1、KIAA2022、ERCC3、PITX3、PRPF3、DNM1L、TCTN3、FHL2、CA 2、GRHPR、PLEKHG5、CDON、KLHL40、TSEN2、SLC1A3、RGR、NEBL、C5orf42、HPS6、<h2 style=";text-align:left;direction:ltr">GFI1、MYCN、LZTR1、BRWD3、TSEN34、F11、SNRNP200、GNAT2、ALG1、TMEM126A、 SP7、KLHL7、TUFM、DLG3、DNAAF2、DNAAF1、VPS13A、NOP10、TMEM5、MCEE、STXB P2、MED25、SHANK3、SLC3A1、TECTA、COX10、CHRNG、RDH5、CDHR1、PHF8、RPL5、 MAOA、GFPT1、RAB3GAP2、CALM2、NAGS、POLR1C、HSD3B2、AMPD1、BUB1B、NEK8、T UBA8、B3GALNT2、FLT3、MATR3、KRT5、GDF6、GREM1、AVPR2、DNAL1、ZDHHC9、CT C1、ALDOA、NR5A1、CYBB、FTSJ1、BLOC1S3、EBP、DCAF17、SPG21、ACAD8、ABCB7、 F12、GLRB、GLIS2、EXOSC3、HUWE1、BMP4、TMIE、GNPTG、RPS26、ITGA2B、LRSAM 1、SLC6A3、ALDH18A1、SERPINC1、KLF11、F7、RPS10、WNT10A、NFIX、MGAT2、ACS F3, RBBP8, CFHR5, COQ6, UBQLN2, CDKN1B, SUOX, FAM126A, COG8, NDUFA10, SMARCE1, ALG8, GSS, EPB42, RPL10, DNAJC19, NAA10, KCNMA1, RPS24, STX11, AL G3、XK、MFRP、TMPRSS3、TSPAN7、SERPINH1、IMPG2、ALG12、SERPINE1、SLC16A 1、TCIRG1、STIM1、ETV6、CLCN7、GDF2、SLC35A1、FAM161A、ARID1B、TMEM231、S LC35A2、NGF、COX4I2、POU1F1、GLIS3、TAF1、PNP、POMC、KIF1BP、BLK、YARS2、 TCN2、UNC13D、HAMP、HOGA1、ACADSB、B4GALT1、MANBA、KAT6B、RSPH4A、ACE、ED AR、WWOX、FARS2、GNAQ、GNPAT、ANKH、ENO3、FRAS1、RANGRF、GALE、TREM2、CD3 D、LEP、TFG、IER3IP1、DYNC2H1、NPM1、KMT2A、CD40LG、PYGL、MT-CYB、DFNB59、<h2 style=";text-align:left;direction:ltr">MRPS16、RTN2、KCNE5、MATN3、TAT、NDUFV2、CDAN1、STS、CAV1、B3GALT6、CTSK 、CALR3、KCNV2、AP4M1、SERPING1、GYS1、HPS5、ST3GAL5、SLC6A5、ARID1A、PRK RA、COG1、COL4A2、EFEMP1、PIK3R2、MTFMT、SEPT9、FOXP1、NDUFAF6、ROM1、KR T14、SLC25A12、SEC23B、TNNI2、CD3E、HPD、PHKB、AIP、FZD4、XPNPEP3、CEP164 、ITGB4、SLMAP、PABPN1、TBCE、GHR、NOG、CACNA2D4、ALG9、FOXL2、TYROBP、TH RB、AP4E1、BDNF、AKT2、DSPP、MPDU1、EDARADD、TPMT、SPTBN2、BLOC1S6、FGF14 、CTSF、PRCD、SRD5A3、PRPF6、TRAPPC11、PHKA2、COCH、AGPS、EARS2、FOXE3、I GBP1、RBP3、PKLR、PIGA、MAT1A、SPTLC2、CEP63、FBXO7、SETBP1、OTOA、RTEL1、 PTF1A、LEPR、SMARCAL1、SCP2、PCBD1、DMP1、MOGS、CNTN1、TNPO3、POLR3A、SL C46A1、FOXI1、MYO15A、KCNQ4、MYOC、PYCR1、APOA5、GRHL2、POR、AICDA、KISS1 R, PRDM16, ARSE, LHFPL5, PDE6G, HARS, SNAI2, VCAN, SMPX, CSF3R, COL17A1, LOXHD1, MTTP, SERPINF1, PROKR2, GNRHR, D2HGDH, B9D2, ZAP70, AP5Z1, CTNNA 3, CSF2RA, SLC34A3, ZNF513, TNFRSF11A, CTRC, RP9, HSPG2, KANSL1, RPS7, TRIOBP, CEL, SHROOM4, SLC7A7, RFT1, ADAMTSL4, ABCA12, ABAT, LPIN2, ERCC5, HGF、PROC、LHX4、ROGDI、ABCA1、DIABLO、ESCO2、PRDM5、PHKG2、FREM1、PRODH 、DIS3L2、RDX、WRAP53、MC1R、ACVR1、ZNF711、IFT80、ACVR2B、EFTUD2、LTBP2、<h2 style=";text-align:left;direction:ltr">MEGF10、RAB18、CLDN14、FLT4、CCT5、SRCAP、ESRRB、PDZD7、NEK1、NR3C2、TBX 20、DNAJB2、FAS、ATXN10、CFHR1、GDF5、PSTPIP1、ARHGEF6、TDP1、GUCA1A、OX CT1、PPP2R2B、AQP2、TRPC6、MARVELD2、FECH、OAT、PEX11B、PRICKLE2、APOC2 、PDGFRB、CACNA1H、LHCGR、SARS2、LRTOMT、COL10A1、XIAP、UNG、MGME1、SLC26 A5, CYBA, PITPNM3, PTH1R, TIMP3, DRD2, PDE6H, ALX4, TXNRD2, OBSl1, ORC1, GH1, CSPP1, LFTY2, CCDC50, ABCD4, DIAPH1, CDH3, CCHD10, PAX8, GDNF, MT- CO1, HARS2, HTRA1, BMP1, MSRB3, ZDHHC15, CAVIN4, AP4S1, CFHR3, ACADL, NDUFA9, MSX1, MYO3A, CYP11B2, CTF1, MAK, AP4B1, IFT122, ABHD5, MARS, A2ML1 CHST3、CYLD、GDF1、XPA、MT-TH、TPRN、MT-TQ、POU4F3、XPC、GRIN1、GIPC3、CY P27B1,POLR1D、LHX3、TGFB1、TOR1AIP1、CNBP、GM2A、DDHD2、TRPM1、BCKDK、D NAAF3、HSD11B2、ADAM9、CLCNKA、NDUFB3、LAS1L、MAGI2、ANKRD11、NMNAT1、Z FYVE27、DNMT3A、PROK2、SMARCA2、GFER、POLR3B、NDUFA12、PLCE1、STRA6、EMX 2, HMGCS2, ASCL1, COMT, PROS1, KCNC3, ILK, FGB, C10orf11, ILDR1, ANKRD26, GRXCR1, SZT2, HNRNPDL, KIF11, FGG, DDC, TTBK2, FREM2, ZNF469, TUSC3, TFA P2A, DLL3, CLIC2, GDF3, MT-TS2, CYP3A5, AHCY, LDHA, SLC52A3, PRKCSH, ACY1, ACO2, KCNK3, AMER1, WNT1, MARS2, NYX, VPS35, UROS, COG6, REN, AVP, MTORTBX3, RBM10, PFN1, TPO, MYBPC1, SERPINB6, PTPRC, H19, ABCB6, WNT7A, MYO5A, CCDC88C, ATP6V0A4, OSTM1, SRD5A2, CDT1, DFNA5, ESPN, MYF6, USB1, DDOST, CRYM, , APOA1, ATXN8OS, AGTR2, SLC17A8, MSX2, DST, LTBP4, KLHL3, AAAS, RFX6, LBR, CYP3A4, F13A1, R AX2, RAC2, PREPL, ERLIN2, ANK3, NFU1, LRP4, TNFRSF13B, TNFSF11, SNAP29, LAMC3, RBM8A, ORC6, GRM6, COG5, ORC4, PDYN, CRELD1, SLC5A7, ITGA3, SPINK5, WNT4, ENAM, C1QTNF5, PDK3, HTRA2, GNB4, WNK4, COG4, MT-TI, HSPB3, MT-TL2, HCFC1, POT1, ICOS, SIGMAR1, ATP2A1, GNAT1, SOS2, CTSC, and are encoded by "clinical testing genes" selected from the group consisting of FOXP2, TMEM165, CXCR4, SH3BP2, TACR3, CFC1, ABCC2, DNAJC6, DHODH, CPA6, AK2, HOXD13, VPS45, PLOD3, KRT1, MT-ATP8, DNAAF5, TGM1, TSPAN12, IFT172, CD2AP, MRPL3, LIFR, RIMS1, CNNM4, CDC6, F10, FOXC2, STAT5B, PIK3R1, ORAI1, ZNF81, ZFP57, CYP24A1, GLE1, COL18A1, TIA1, RPL26, GNAO1, LCAT, VDR, ANO10, TNNT3, LZTFL1, COL4A6, and SHANK2.
[0178] IV. Computer Systems and Readable Devices The methods described herein, e.g., conformational dynamics generation (e.g., MD simulation), dimensionality reduction, ML model training, or any combination thereof, may be implemented using a computer system due to the computational complexity involved. Thus, in some aspects, the disclosure provides a system, comprising: (a) a memory; and (b) at least one processor, coupled to the memory, for: (i) generating (e.g., using MD) conformational dynamics of a reference molecule; (ii) processing the simulated conformational dynamics of the reference molecule to generate a low-dimensional representation of the conformational landscape of the reference molecule; (iii) extracting conformational states from the low-dimensional conformational landscape of the reference molecule structure to seed a simulation of a variant molecule (e.g., a variant protein); and (iv) detecting at least one missense in the simulated conformation of the reference molecule in the extracted conformational state. and (v) generating conformational dynamics of the variant molecule or set thereof by introducing missense mutations; (vi) processing the conformational dynamics of the variant molecule or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecule or set thereof; and (vii) training a machine learning predictive model using the low-dimensional representation of the conformational landscape of the reference molecule and the variant molecule using clinical data as training labels, wherein the predictive model classifies the effect of variants (e.g., missense mutations) in the reference molecule.
[0179] The disclosure also provides a tangible computer-readable device having instructions stored thereon that, when executed by at least one computing device, cause the at least one computing device to: (a) generate (e.g., using MD) a conformational dynamics of a reference molecule (e.g., a reference protein); (b) process the conformational dynamics of the reference molecule to generate a low-dimensional representation of the conformational landscape of the reference molecule; (c) extract states from the low-dimensional conformational landscape of the reference molecule to seed a simulation of a variant molecule; and (d) seed the simulated conformation of the reference molecule in the extracted states with at least one variant ( (e) generating a variant molecule or a set of variant molecules (e.g., by introducing a variant (e.g., a missense mutation)); (e) generating conformational dynamics of the variant molecule or the set of variant molecules (e.g., using MD); (f) processing the conformational dynamics of the variant molecule or the set of variant molecules to generate a low-dimensional representation of the conformational landscape of the variant molecule or the set of variant molecules; and (g) training a machine learning predictive model using the low-dimensional representation of the conformational landscape of the reference molecule and the variant molecule using clinical data as training labels, where the predictive model classifies the effect of the variant (e.g., a protein missense mutation) in the reference molecule.
[0180] The present disclosure provides a computer system for implementing the methods described herein, such as ML prediction models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants), and / or generating such ML prediction models. A computer system is an apparatus or device that includes one or more processors, memory, input / output interfaces, and network interfaces. In some aspects, a computer system includes hardware elements electrically coupled via a bus, including a processor, an input device, an output device, a storage device, a computer-readable storage medium reader, a communication system, processing acceleration (e.g., a DSP or special purpose processor), and a memory.
[0181] The computer-readable storage medium reader may further be coupled to a computer-readable storage medium, which combination generically represents a storage medium, memory, etc. for containing computer-readable information on a temporary and / or more permanent basis, in addition to remote, local, fixed and / or removable storage devices, which may include storage devices, memories, and / or any other such accessible system resources. In some aspects, the memory is communicatively coupled to one or more processors. The one or more processors are configured to execute computer program instructions for implementing the methods described herein. The computer program instructions are stored on and / or provided by the memory. Examples of computers include personal computers, laptops, computer clusters, supercomputers, and any other type of computer.
[0182] Memory is an example of a computer-readable medium. It may include volatile storage devices, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (Flash RAM) in the computer-readable medium. Computer-readable media include permanent, non-permanent, removable, and non-removable media. Computer-readable commands, data structures, and program modules are examples of information stored on a computer-readable medium. Examples of computer storage media include phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), flash memory, compact disk read-only memory (CD-ROM), digital multifunction disks (DVD), and magnetic tape. Computer-readable media does not include temporary or transitory computer-readable media such as modulated data signals and carrier waves.
[0183] Computer processors include a number of components, including multi-core CPUs, multi-CPU computers and computer clusters, many-core CPUs, general purpose graphical processor units (gpGPUs), and other computer processors.
[0184] In embodiments, the methods described herein, e.g., ML predictive models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants), and / or the step of generating such ML predictive models, are performed using a computer processor with sufficient computational power to perform the simulations and analysis. Currently, molecular simulations and machine learning are performed on parallel architectures such as multi-core CPUs, multi-CPU systems, and general-purpose GPUs. In embodiments, a computer system for performing the methods described herein, e.g., ML predictive models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants), and / or the step of generating such ML predictive models, includes sufficient memory for computation and storage, such as disks, disk arrays, solid-state drives (SSDs), at a speed sufficient for the simulations and analysis.
[0185] A single architecture may be utilized to implement one or more servers, which may be further configured according to currently desired protocols, protocol variations, extensions, and the like. However, it will be apparent to one of ordinary skill in the art that aspects may be fully utilized according to more specific application requirements. Customized hardware may also be utilized, and / or particular elements may be implemented in hardware, software, or both. Additionally, connections to other computing devices, such as network input / output devices (not shown), may be utilized, although it should be understood that wired, wireless, modem, and / or other connections or connections to other computing devices may also be utilized.
[0186] In one embodiment, the system further comprises one or more devices for providing input data to the one or more processors. The system further comprises a memory for storing the dataset of ranked data elements. In another embodiment, the device for providing input data comprises a detector for detecting features of the data elements, such as, for example, a fluorescent plate reader, a mass spectrometer, a gene chip reader, or a gene sequencer.
[0187] The system may further include a database management system. A user's request or query may be formatted in an appropriate language understood by the database management system, which processes the query and extracts relevant information from the training set database. The system may be connectable to a network to which a network server and one or more clients are connected. The network may be a local area network (LAN) or a wide area network (WAN), as known in the art. Preferably, the server includes the necessary hardware to execute a computer program product (e.g., software) to access database data for processing user requests. The system may communicate with an input device for providing data related to data elements (e.g., values of expressions) to the system.
[0188] Some aspects described herein may be implemented to include a computer program product. The computer program product may include a computer readable medium having computer readable program code embodied therein for executing an application program on a computer with a database. As used herein, a "computer program product" refers to a set of instructions organized in the form of natural language or programming language statements contained in a physical medium of any nature (e.g., written, electronic, magnetic, optical, or other) and usable in conjunction with a computer or other automated data processing system. Such programming language statements, when executed by a computer or data processing system, cause the computer or data processing system to operate according to the particular content of the statements.
[0189] Computer program products include, but are not limited to, source and object code programs and / or test or data libraries embedded in a computer readable medium. Furthermore, computer program products that enable a computer system or data processing device to operate in a preselected manner may be provided in several forms, including, but not limited to, original source code, assembly code, object code, machine language, encrypted or compressed versions of the foregoing, and any equivalents. In one aspect, a computer program product is provided for implementing a therapeutic, diagnostic, prognostic, or monitoring method disclosed herein, for example, to determine whether to administer a particular therapy based on the classification of a variant protein by the ML predictive model disclosed herein.
[0190] In some aspects, a computer program product includes a computer readable medium and embodies program code executable by a processor of a computing device or system, program code for performing at least one of the computer-implemented methods disclosed herein, or an ML predictive model generated according to the methods disclosed herein. In some aspects, the computer readable medium may include a library of pre-computed MD structures, pre-computed models, pre-computed variant pathogenicity scores, or any combination thereof.
[0191] While various aspects have been described as methods or apparatus, it should be understood that the aspects may be implemented through code coupled to a computer (e.g., code resident on or accessible by a computer). For example, software and databases may be utilized to implement many of the methods described above. Thus, in addition to aspects that are accomplished by hardware, it should also be noted that these aspects may be accomplished using an article of manufacture consisting of a computer usable medium having computer readable program code embodied therein, which enables the functionality disclosed in this description. Thus, the aspects are also desirably considered to be protected by this patent in their program code means.
[0192] Further, some aspects may be code stored in virtually any type of computer readable memory, including but not limited to RAM, ROM, magnetic media, optical media, or magneto-optical media. Even more generally, some aspects may be implemented in software or hardware, or any combination thereof, including but not limited to software running on a general purpose processor, microcode, PLA, or ASIC.
[0193] It is also contemplated that some aspects may be accomplished as a computer signal embodied in a carrier wave, as well as a signal propagated over a transmission medium (e.g., electrical and optical). Thus, various types of information discussed above may be formatted in structures, such as data structures, and transmitted as an electrical signal over a transmission medium or stored in a computer-readable medium.
[0194] V. Companion Diagnostics The methods disclosed herein or the generated predictive models described herein may be provided as a companion diagnostic available, for example, via a web server, to inform a clinician or patient about potential treatment options. The methods disclosed herein may include taking or otherwise obtaining a biological sample, performing an analytical method (e.g., applying an ML predictive model disclosed herein) to classify the variant (e.g., a mutation in a variant protein), for example, as pathogenic, and based on the assignment, providing a suitable treatment for administration to the patient.
[0195] VI. Accuracy and Clinical Performance Metrics of ML Predictive Models In general, predictive models developed for medical diagnosis should be validated and their accuracy judged as reliable by physicians. Multiple performance metrics can be used to validate ML predictive models, such as accuracy, RO curve, kappa statistic, MAE, and RMSE.
[0196] The accuracy of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein can be determined by a 95% confidence interval (CI). In general, a prediction model is considered to have good accuracy if its 95% CI does not overlap with 1. In some embodiments, the 95% CI of the predictive model is at least about 1.08, at least about 1.10, at least about 1.12, at least about 1.14, at least about 1.15, at least about 1.16, at least about 1.17, at least about 1.18, at least about 1.19, at least about 1.20, at least about 1.21, at least about 1.22, at least about 1.23, at least about 1.24, at least about 1.25, at least about 1.26, at least about 1.27, at least about 1.28, at least about 1.29, at least about 1.30, at least about 1.31, at least about 1.32, at least about 1.33, at least about 1.34, or at least about 1.35, or more. The 95% CI of the predictive model can be at least about 1.14, at least about 1.15, at least about 1.16, at least about 1.20, at least about 1.21, at least about 1.26, or at least about 1.28. The 95% CI of the predictive model can be less than about 1.75, less than about 1.74, less than about 1.73, less than about 1.72, less than about 1.71, less than about 1.70, less than about 1.69, less than about 1.68, less than about 1.67, less than about 1.66, less than about 1.65, less than about 1.64, less than about 1.63, less than about 1.62, less than about 1.61, less than about 1.60, less than about 1.59, less than about 1.58, less than about 1.57, less than about 1.56, less than about 1.55, less than about 1.54, less than about 1.53, less than about 1.52, less than about 1.51, less than about 1.50, or lower. The 95% CI of the predictive model may be less than about 1.61, less than about 1.60, less than about 1.59, less than about 1.58, less than about 1.56, 1.55, or less than 1.53.The 95% CI of the predictive model can be about 1.10 to 1.70, about 1.12 to about 1.68, about 1.14 to about 1.62, about 1.15 to about 1.61, about 1.15 to about 1.59, about 1.16 to about 1.160, about 1.19 to about 1.55, about 1.20 to about 1.54, about 1.21 to about 1.53, about 1.26 to about 1.63, about 1.27 to about 1.61, or about 1.28 to about 1.60.
[0197] In some embodiments, the accuracy of an ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein depends on the difference in the 95%CI range (e.g., the difference between the high and low values of the 95%CI interval). In general, ML prediction models with a large difference in the range of the 95%CI interval are considered to be more variable and less accurate than ML prediction models with a small difference in the range of the 95%CI interval. In some embodiments, an ML predictive model is considered to be more accurate if the 95% CI range difference is less than about 0.60, less than about 0.55, less than about 0.50, less than about 0.49, less than about 0.48, less than about 0.47, less than about 0.46, less than about 0.45, less than about 0.44, less than about 0.43, less than about 0.42, less than about 0.41, less than about 0.40, less than about 0.39, less than about 0.38, less than about 0.37, less than about 0.36, less than about 0.35, less than about 0.34, less than about 0.33, less than about 0.32, less than about 0.31, less than about 0.30, less than about 0.29, less than about 0.28, less than about 0.27, less than about 0.26, less than about 0.25, or less. The difference in the 95% CI range of the ML prediction model may be less than about 0.48, less than about 0.45, less than about 0.44, less than about 0.42, less than about 0.40, less than about 0.37, less than about 0.35, less than about 0.33, or less than about 0.32. In some embodiments, the difference in the 95% CI range of the ML prediction model is about 0.25 to about 0.50, about 0.27 to about 0.47, or about 0.30 to about 0.45. In some embodiments, the sensitivity of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein is at least about 60%. In some embodiments, the sensitivity is at least about 65%. In some embodiments, the sensitivity is at least about 70%. In some embodiments, the sensitivity is at least about 75%. In some embodiments, the sensitivity is at least about 80%. In some embodiments, the sensitivity is at least about 85%. In some embodiments, the sensitivity is at least about 90%. In some embodiments, the sensitivity is at least about 95%.In some embodiments, the ML prediction model for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein is clinically significant. In some embodiments, the clinical significance of the ML prediction model is determined by the AUC value. To be clinically significant, the AUC value is at least about 0.5, at least about 0.55, at least about 0.6, at least about 0.65, at least about 0.7, at least about 0.75, at least about 0.8, at least about 0.85, at least about 0.9, or at least about 0.95. The clinical significance of the prediction model can be determined by percent accuracy. For example, a predictive model is determined to be clinically significant if the accuracy of the predictive model is at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 72%, at least about 75%, at least about 77%, at least about 80%, at least about 82%, at least about 84%, at least about 86%, at least about 88%, at least about 90%, at least about 92%, at least about 94%, at least about 96%, or at least about 98%. In other aspects, the clinical significance of the ML predictive models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein is determined by the median fold difference (MDF) value. To be clinically significant, the MDF value is at least about 0.8, at least about 0.9, at least about 1.0, at least about 1.1, at least about 1.2, at least about 1.3, at least about 1.4, at least about 1.5, at least about 1.6, at least about 1.7, at least about 1.9, or at least about 2.0. In some embodiments, the MDF value is 1.1 or greater. In other embodiments, the MDF value is 1.2 or greater. Alternatively, or in addition, the clinical significance of the ML prediction model is determined by the P value of a t-test.In some embodiments, to be clinically significant, the t-test P value is less than about 0.070, less than about 0.065, less than about 0.060, less than about 0.055, less than about 0.050, less than about 0.045, less than about 0.040, less than about 0.035, less than about 0.030, less than about 0.025, less than about 0.020, less than about 0.015, less than about 0.010, less than about 0.005, less than about 0.004, or less than about 0.003. The t-test P value may be less than about 0.050. Alternatively, the t-test P value is less than about 0.010. In some embodiments, the clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein is determined by clinical outcome. For example, different clinical outcomes may have different minimum or maximum thresholds for AUC values, MDF values, t-test P values, and accuracy values that determine whether a predictive model is clinically significant. In another example, an ML predictive model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein is considered clinically significant if the t-test P value is less than about 0.08, less than about 0.07, less than about 0.06, less than about 0.05, less than about 0.04, less than about 0.03, less than about 0.02, less than about 0.01, less than about 0.005, less than about 0.004, less than about 0.003, less than about 0.002, or less than about 0.001. In some embodiments, the performance of an ML predictive model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein is based on odds ratios.ML prediction models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein provide odds ratios of at least about 1.30, at least about 1.31, at least about 1.32, at least about 1.33, at least about 1.34, at least about 1.35, at least about 1.36, at least about 1.37, at least about 1.38, at least about 1.39, at least about 1.40, at least about 1.41, at least about 1.42, at least about 1.43, at least about 1.44, at least about 1.45, at least about 1.46, at least about 1.47, at least about 1.48, at least about 1.49, at least about 2.00, at least about 2.01, at least about 2.02, at least about 2.03, at least about 2.04, at least about 2.05, at least about 2.06, at least about 2.07, at least about 2.08, at least about 2.09, at least about 30, at least about 31, at least about 32, at least about 33, at least about 34, at least about 35, at least about 36, at least about 37, at least about 38, at least about 39, at least about 40, at least about 41, at least about 42, at least about 43, at least about 44, at least about 45, at least about 46, at least about 47, at least about 48, at least about 49, at least about 50, at least about 51, at least about 52, at least about 53, at least about 54, at least about 55, at least about 56, at least about 57, at A ML prediction model may be considered to have good performance if its odds ratio is at least about 1.42, at least about 1.43, at least about 1.44, at least about 1.45, at least about 1.46, at least about 1.47, at least about 1.48, at least about 1.49, at least about 1.50, at least about 1.52, at least about 1.55, at least about 1.57, at least about 1.60, at least about 1.62, at least about 1.65, at least about 1.67, at least about 1.70, or more. In some embodiments, the odds ratio of the ML prediction model is at least about 1.33. The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be based on a univariate analysis odds ratio P value (uvaORPval). The univariate analysis odds ratio P value (uvaORPval) of the prediction model may be about 0 to about 0.4. The univariate analysis odds ratio P value (uvaORPval) of the prediction model may be about 0 to about 0.3. The univariate analysis odds ratio P value (uvaORPval) of the prediction model may be about 0 to about 0.2. The univariate analysis odds ratio P value (uvaORPval) of the ML prediction model may be 0.25 or less, about 0.22 or less, about 0.21 or less, about 0.20 or less, about 0.19 or less, about 0.18 or less, about 0.17 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12 or less, or about 0.11 or less. The univariate odds ratio P-value (uvaORPval) of an ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein can be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less.The univariate analysis odds ratio P value (uvaORPval) of the ML prediction model may be about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less. The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be based on the multivariate analysis odds ratio P value (mvaORPval). The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model may be about 0 to about 1. The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model may be about 0 to about 0.9. The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model may be about 0 to about 0.8. The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model may be about 0.90 or less, about 0.88 or less, about 0.86 or less, about 0.84 or less, about 0.82 or less, or about 0.80 or less. The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model may be about 0.78 or less, about 0.76 or less, about 0.74 or less, about 0.72 or less, about 0.70 or less, about 0.68 or less, about 0.66 or less, about 0.64 or less, about 0.62 or less, about 0.60 or less, about 0.58 or less, about 0.56 or less, about 0.54 or less, about 0.52 or less, or about 0.50 or less. The multivariate analysis odds ratio P-value (mvaORPval) of the ML predictive model can be about 0.48 or less, about 0.46 or less, about 0.44 or less, about 0.42 or less, about 0.40 or less, about 0.38 or less, about 0.36 or less, about 0.34 or less, about 0.32 or less, about 0.30 or less, about 0.28 or less, about 0.26 or less, about 0.25 or less, about 0.22 or less, about 0.21 or less, about 0.20 or less, about 0.19 or less, about 0.18 or less, about 0.17 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12 or less, or about 0.11 or less. The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model can be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less.The multivariate analysis odds ratio P value (mvaORPval) of the ML prediction model may be about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less. The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be based on the Kaplan-Meier P value (KM P value). The Kaplan-Meier P value (KM P value) of the ML prediction model may be about 0 to about 0.8. The Kaplan-Meier P value (KM P value) of the ML prediction model may be about 0 to about 0.7. The Kaplan-Meier P-value (KM P-value) of the ML predictive model can be about 0.80 or less, about 0.78 or less, about 0.76 or less, about 0.74 or less, about 0.72 or less, about 0.70 or less, about 0.68 or less, about 0.66 or less, about 0.64 or less, about 0.62 or less, about 0.60 or less, about 0.58 or less, about 0.56 or less, about 0.54 or less, about 0.52 or less, or about 0.50 or less. The Kaplan-Meier P-value (KM P-value) of the ML predictive model can be about 0.48 or less, about 0.46 or less, about 0.44 or less, about 0.42 or less, about 0.40 or less, about 0.38 or less, about 0.36 or less, about 0.34 or less, about 0.32 or less, about 0.30 or less, about 0.28 or less, about 0.26 or less, about 0.25 or less, about 0.22 or less, about 0.21 or less, about 0.20 or less, about 0.19 or less, about 0.18 or less, about 0.17 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12 or less, or about 0.11 or less. The Kaplan-Meier P-value (KM P-value) of the ML prediction model may be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less. The Kaplan-Meier P-value (KM P-value) of the ML prediction model may be about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less. The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be based on the survival AUC value (survAUC). The survival rate AUC value (survAUC) of the ML prediction model may be about 0 to 1. The survival rate AUC value (survAUC) of the ML prediction model may be about 0 to about 0.9. The survival rate AUC value (survAUC) of the ML prediction model may be about 1 or less, about 0.98 or less, about 0.96 or less, about 0.94 or less, about 0.92 or less, about 0.90 or less, about 0.88 or less, about 0.86 or less, about 0.84 or less, about 0.82 or less, or about 0.80 or less. The survival AUC value (survAUC) of the ML predictive model can be about 0.80 or less, about 0.78 or less, about 0.76 or less, about 0.74 or less, about 0.72 or less, about 0.70 or less, about 0.68 or less, about 0.66 or less, about 0.64 or less, about 0.62 or less, about 0.60 or less, about 0.58 or less, about 0.56 or less, about 0.54 or less, about 0.52 or less, or about 0.50 or less. The survival AUC value (survAUC) of the ML predictive model can be about 0.48 or less, about 0.46 or less, about 0.44 or less, about 0.42 or less, about 0.40 or less, about 0.38 or less, about 0.36 or less, about 0.34 or less, about 0.32 or less, about 0.30 or less, about 0.28 or less, about 0.26 or less, about 0.25 or less, about 0.22 or less, about 0.21 or less, about 0.20 or less, about 0.19 or less, about 0.18 or less, about 0.17 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12 or less, or about 0.11 or less.The survival AUC value (survAUC) of the ML prediction model may be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less. The survival AUC value (survAUC) of the ML prediction model may be about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less. The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein may be based on univariate analysis hazard ratio P value (uvaHRPval). The univariate analysis hazard ratio P value (uvaHRPval) of the ML prediction model may be about 0 to about 0.4. The univariate analysis hazard ratio P value (uvaHRPval) of the ML prediction model may be about 0 to about 0.3. The univariate analysis hazard ratio P value (uvaHRPval) of the ML prediction model may be about 0.40 or less, about 0.38 or less, about 0.36 or less, about 0.34 or less, or about 0.32 or less. The univariate analysis hazard ratio P value (uvaHRPval) of the ML prediction model may be about 0.30 or less, about 0.29 or less, about 0.28 or less, about 0.27 or less, about 0.26 or less, about 0.25 or less, about 0.24 or less, about 0.23 or less, about 0.22 or less, about 0.21 or less, or about 0.20 or less. The ML prediction model univariate analysis hazard ratio P value (uvaHRPval) may be about 0.19 or less, about 0.18 or less, about 0.17 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12 or less, or about 0.11 or less. The ML prediction model univariate analysis hazard ratio P value (uvaHRPval) may be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less. The univariate analysis hazard ratio P-value (uvaHRPval) of the ML predictive model may be about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less.The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be based on a multivariate analysis hazard ratio P value (mvaHRPval). The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0 to about 1. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0 to about 0.9. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 1 or less, about 0.98 or less, about 0.96 or less, about 0.94 or less, about 0.92 or less, about 0.90 or less, about 0.88 or less, about 0.86 or less, about 0.84 or less, about 0.82 or less, or about 0.80 or less. The multivariate hazard ratio P-value (mvaHRPval) of the ML predictive model can be about 0.80 or less, about 0.78 or less, about 0.76 or less, about 0.74 or less, about 0.72 or less, about 0.70 or less, about 0.68 or less, about 0.66 or less, about 0.64 or less, about 0.62 or less, about 0.60 or less, about 0.58 or less, about 0.56 or less, about 0.54 or less, about 0.52 or less, or about 0.50 or less. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML predictive model is about 0.48 or less, about 0.46 or less, about 0.44 or less, about 0.42 or less, about 0.40 or less, about 0.38 or less, about 0.36 or less, about 0.34 or less, about 0.32 or less, about 0.30 or less, about 0.28 or less, about 0.26 or less, about 0.25 or less, about 0.22 or less, about 0.21 or less, about 0.20 or less, about 0.19 or less, about 0.18 or less, about 0.17 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12, or about 0.11 or less. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less.The clinical significance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be based on a multivariate analysis hazard ratio P value (mvaHRPval). The multivariate analysis hazard ratio P value (mvaHRPval) of the prediction model may be about 0 to about 0.60. The significance of the ML prediction model may be based on a multivariate analysis hazard ratio P value (mvaHRPval). The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0 to about 0.50. The significance of the prediction model may be based on a multivariate analysis hazard ratio P value (mvaHRPval). The multivariate analysis hazard ratio P-value (mvaHRPval) of the ML predictive model can be about 0.50 or less, about 0.47 or less, about 0.45 or less, about 0.43 or less, about 0.40 or less, about 0.38 or less, about 0.35 or less, about 0.33 or less, about 0.30 or less, about 0.28 or less, about 0.25 or less, about 0.22 or less, about 0.20 or less, about 0.18 or less, about 0.16 or less, about 0.15 or less, about 0.14 or less, about 0.13 or less, about 0.12 or less, about 0.11, or about 0.10 or less. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0.10 or less, about 0.09 or less, about 0.08 or less, about 0.07 or less, about 0.06 or less, about 0.05 or less, about 0.04 or less, about 0.03 or less, about 0.02 or less, or about 0.01 or less. The multivariate analysis hazard ratio P value (mvaHRPval) of the ML prediction model may be about 0.01 or less, about 0.009 or less, about 0.008 or less, about 0.007 or less, about 0.006 or less, about 0.005 or less, about 0.004 or less, about 0.003 or less, about 0.002 or less, or about 0.001 or less. The performance of an ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be determined by the AUC value, and improvement in performance may be determined by the difference between the AUC value of the ML prediction model disclosed herein and the AUC value of a current prediction model.In some embodiments, an ML predictive model disclosed herein is superior to a current predictive model if the AUC value of the ML predictive model disclosed herein is at least about 0.05, at least about 0.06, at least about 0.07, at least about 0.08, at least about 0.09, at least about 0.10, at least about 0.11, at least about 0.12, at least about 0.13, about 0.14, at least about 0.15, at least about 0.16, at least about 0.17, at least about 0.18, at least about 0.19, at least about 0.20, at least about 0.022, at least about 0.25, at least about 0.27, at least about 0.30, at least about 0.32, at least about 0.35, at least about 0.37, at least about 0.40, at least about 0.42, at least about 0.45, at least about 0.47, or at least about 0.50 or more greater than the AUC value of the current predictive model. In some embodiments, the AUC value of the ML prediction model disclosed herein is at least about 0.10 greater than the AUC value of the current prediction model. In some embodiments, the AUC value of the ML prediction model disclosed herein is at least about 0.13 greater than the AUC value of the current prediction model. In some embodiments, the AUC value of the ML prediction model disclosed herein is at least about 0.18 greater than the AUC value of the current prediction model. The performance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein may be determined by an odds ratio, and the improvement in performance may be determined by comparing the odds ratio of the ML prediction model disclosed herein with the odds ratio of the current prediction model. The comparison of the performance of two or more prediction models may generally be based on a comparison of the absolute value of the (1 odds ratio) of a first prediction model with the absolute value of the (1 odds ratio) of a second prediction model. In general, a predictive model with a larger absolute value of (1 odds ratio) can be considered to have better performance compared to a predictive model with a smaller absolute value of (1 odds ratio).
[0198] In some embodiments, the performance of the ML prediction model for diagnosing and / or predicting the phenotypic impact of a variant (e.g., a missense variant) disclosed herein is based on a comparison of odds ratios and 95% confidence intervals (CI). For example, a first prediction model may have a larger absolute value (1 odds ratio) than a second prediction model, but the 95% CI of the first prediction model may overlap with 1 (e.g., lower accuracy), while the 95% CI of the second prediction model does not overlap with 1. In this example, the accuracy of the first prediction model is lower than the accuracy of the second prediction model, so the second prediction model is considered to be superior to the first prediction model. In another example, the first prediction model may be superior to the second prediction model based on a comparison of odds ratios, but the difference in the 95% CI of the first prediction model is at least about 2 times larger than the 95% CI of the second prediction model. In this case, the second prediction model is considered to be superior to the first prediction model.
[0199] In some embodiments, the ML predictive models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein are more accurate than current predictive models. The ML predictive models disclosed herein are more accurate than current predictive models if the 95% CI range of the ML predictive model disclosed herein does not span or overlap 1, and the 95% CI range of the current predictive model or clinical variable spans or overlaps 1. In some embodiments, the ML predictive models for diagnosing and / or predicting the phenotypic impact of variants (e.g., missense variants) disclosed herein are more accurate than current predictive models known in the art. In some embodiments, the ML prediction models disclosed herein are more accurate than current prediction models known in the art if the difference in the 95% CI range of the ML prediction models disclosed herein is about 0.70, about 0.60, about 0.50, about 0.40, about 0.30, about 0.20, about 0.15, about 0.14, about 0.13, about 0.12, about 0.10, about 0.09, about 0.08, about 0.07, about 0.06, about 0.05, about 0.04, about 0.03, or about 0.02 times smaller than the difference in the 95% CI range of the current prediction models. The ML prediction models disclosed herein are more accurate than current prediction models if the difference in the 95% CI range of the ML prediction models disclosed herein is about 0.20 to about 0.04 times smaller than the difference in the 95% CI range of the current prediction models. EXAMPLES
[0200] Example 1 Predictive model of the MSH2 (MutS homolog 2) / MSH6 (MutS homolog 6) heterodimer In one particular example, the native MSH2 / MSH6 heterodimer was simulated using MD to generate conformational dynamics, which were then used to determine the native conformational landscape. From the conformational dynamics, a low-dimensional representation was determined, which was then used to identify 30 metastable conformational states. See FIG. 6. From each of these conformations, 45 variants were introduced into the MSH2 protein and simulations were run for each variant. Each variant was simulated 10 times in each of the 30 states for between 1 and 25 ns, a total of less than 1 microsecond per variant. In comparison, the native MSH2 / MSH6 heterodimer was simulated for approximately 5 microseconds to determine the conformational landscape.
[0201] Simulations of each variant were processed to reduce dimensionality by first extracting structural features of each conformation in each trajectory for each state (either backbone dihedral angles (φ, ψ angles), connectivity at each amino acid position defined by the number of amino acids in the vicinity of 15 nm or inter-residue distances for amino acid pairs closer than 18 nm). The dimensionality of the data was then further reduced using PCA or tICA to a final dimensionality of 64 features per state.
[0202] Model training tested how much trajectory data was used in the modeling to determine the minimum amount of simulation time required. Three dataset definitions were used: one repeat of 10 ns for 4 states (40 ns per variant), one repeat of 5 ns for 17 states (85 ns per variant), or two repeats of 3 ns for 11 states (66 ns per variant). See Figure 8. This represented a dramatic reduction in the simulation time required per variant compared to the approximately 5 microseconds of simulation required to determine the conformational landscape of the native MSH2 / MSH6 heterodimer.
[0203] In summary, five dimensionality reduction methods (dihedral angles, tlCA; connectivity, tlCA; dihedral angles, PCA; contacts, PCA; and connectivity, PCA) and three dataset definitions (10ns, 1 iteration, 4 states; 5ns, 1 iteration, 17 states; 3ns, 2 iterations, 11 states) were used for hyperparameter tuning of feature quantification methods and dataset definitions.
[0204] The reduced dimensionality data was then modeled using a sklearn pipeline for hyperparameter tuning. 25% of the data was reserved as a test set, and the remaining data was split into a 5-fold cross-validation set. The cross-validation set was trained using a random forest model, hyperparameters that tuned the number of estimators, minimum number of samples to split a node, maximum number of features per tree, maximum depth of a tree, and minimum samples per leaf. Hyperparameter tuning was scored against F1 accuracy score. Performance on the cross-validation set was used to evaluate the featurization methods and dataset definition. A hold-out test set was used to evaluate model generalization after selecting the top models based on cross-validation performance.
[0205] Nearly all combinations of feature quantification methods, dataset definitions, and model hyperparameter tuning achieved high performance accuracy (>0.8). As shown in FIG. 9, the performance of the best performing model (LSS model) generated according to the disclosed method was compared to the performance of the same variants in sequence-based and FoldX-based models for variant selection, and when compared to the true labels, the overall LSS model predictions of these variants were closer to the true labels. The best performing model (LSS model) from the cross-validation set (using amino acid connectivity, tICA dimensionality reduction) achieved model performance on the hold-out test set that was better than the performance of sequence-based and FoldX-based models. As shown in FIG. 10, the performance of the prediction model for MSH2 / MSH6 heterodimer developed according to the disclosed method (LSS model) was better than alternative methods developed using, for example, FoldX, Polyphen, or process scikit (a computational predictor developed by INVITAE™), as measured by the area under the receiver operating characteristic (ROC) curve. The LSS model achieved a model performance of 0.997 AUC, while FoldX, Polyphen, and process scikit achieved performances of 0.919, 0.900, and 0.900 AUC, respectively.
[0206] VII. Terminology In order that this disclosure may be more readily understood, certain terms are first defined. As used in this disclosure, unless expressly provided herein, each of the following terms shall have the meaning set forth below. Additional definitions are set forth throughout this disclosure.
[0207] Unless the context clearly dictates otherwise, the singular forms "a," "an," and "the" include plural referents. The terms "a" (or "an"), as well as "one or more" and "at least one," may be used interchangeably herein. In certain aspects, the terms "a" or "an" refer to "single." In other aspects, the terms "a" or "an" include "two or more" or "plural."
[0208] Furthermore, the term "and / or" as used herein should be understood as a specific disclosure of each of two particular features or components with or without the other. Thus, the term "and / or" used in phrases such as "A and / or B" herein is intended to include A and B, A or B, A (alone), and B (alone). Similarly, the term "and / or" used in phrases such as "A, B, and / or C" is intended to encompass each of the following aspects: A, B, and C; A, B, or C; A or C; A or B; B or C; A and C; A and B; B and C; A (alone); B (alone); and C (alone).
[0209] As described herein, any concentration range, percentage range, ratio range, or integer range, unless otherwise indicated, should be understood to include any integer value within the recited range, and fractions thereof, where appropriate (such as tenths and hundredths of integers).
[0210] Wherever an embodiment is described herein with the term "comprising," it is understood that alternative analogous embodiments described with the terms "consisting of" and / or "consisting essentially of" are also provided.
[0211] Units, prefixes, and symbols are expressed in the form accepted by the International System of Units (SI). The headings provided herein are not limitations of the various aspects of the present disclosure and may be understood by reference to the specification as a whole. Thus, the terms defined are more fully defined by reference to the specification as a whole.
[0212] The terms "about," "comprising essentially of," or "consisting essentially of" refer to a value or composition that is within an acceptable error range for a particular value or composition as determined by one of ordinary skill in the art, which depends in part on how the value or composition is measured or determined, i.e., the limitations of the measurement system. For example, "about," "comprising essentially of," or "consisting essentially of" can mean within one or more than one standard deviation, in accordance with practices in the art. Alternatively, "about," "comprising essentially of," or "consisting essentially of" can mean a range of up to 10%. Additionally, particularly with respect to biological systems or processes, the term can mean up to an order of magnitude or up to 5-fold the value. Where a specific value or composition is provided in this disclosure and claims, unless otherwise indicated, the meaning of "about," "comprising essentially of," or "consisting essentially of" should be considered within an acceptable range of error for that particular value or composition.
[0213] As used herein, the term "approximately" when applied to one or more values of interest refers to a value similar to a stated reference value. In certain embodiments, the term "approximately" refers to a range of values that falls within 10%, 9%, 8%, 7%, 6%, 5%, 4%, 3%, 2%, 1%, or less in either direction (greater or less) of the stated reference value, unless otherwise stated or clear from the context (except where such number exceeds 100% of possible values).
[0214] It should be understood that the Detailed Description section, and not the Overview and Summary sections, are intended to be used to interpret the embodiments. The Overview and Summary sections may describe one or more, but not all, example embodiments of the invention as contemplated by the inventor(s), and thus are not intended to limit the invention and the accompanying embodiments in any manner.
[0215] The present invention has been described above using functional building blocks illustrating the implementation of certain functions and relationships thereof. The boundaries of these functional building blocks have been arbitrarily defined herein for the convenience of description. Alternative boundaries can be defined so long as the certain functions and relationships thereof are appropriately performed.
[0216] The foregoing description of the specific embodiments fully reveals the general nature of the present invention, so that others, by applying knowledge within the skill of the art, can easily modify and / or adapt such specific embodiments to various applications without departing from the general concept of the present invention and without undue experimentation. Such adaptations and modifications are therefore intended to be within the meaning and range of equivalents of the disclosed embodiments, based on the teaching and guidance presented herein. It should be understood that the terms or terminology used herein are for the purpose of description and not of limitation, as the terms or terminology used herein would be interpreted by those skilled in the art in light of the teaching and guidance.
[0217] The breadth and scope of the present invention should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following embodiments and their equivalents.
[0218] The contents of all references (including literature references, patents, patent applications, and websites) that may be cited throughout this disclosure, as well as the references cited therein, are expressly incorporated herein by reference in their entirety for any purpose.
Claims
1. 1. A computer-implemented method for high-throughput classification of variant effects, comprising: (i) generating conformational dynamics of a reference molecular structure (e.g., using molecular dynamics (MD)); (ii) processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; and (iii) extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; and (iv) generating a variant molecular structure or a set thereof by introducing at least one variant into the extracted conformational state of the reference molecular structure; (v) generating (e.g., using MD) conformational dynamics of said variant molecular structure or set thereof; (vi) processing the conformational dynamics of the variant molecular structure or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or set thereof; (vii) training a machine learning predictive model using the reduced dimensional representation of the conformational landscapes of the reference and variant molecular structures using clinical data as training labels, wherein the predictive model classifies the variant effects in the reference molecular structures.
2. generating a conformational dynamics of the reference molecular structure or the variant molecular structure; (a) obtaining a reference molecular structure or a variant molecular structure; (b) generating conformational dynamics using the obtained reference molecular structure or variant protein structure.
3. 3. The method of claim 2, wherein the reference molecular structure or variant molecular structure is obtained from the Protein Data Bank (PDB) or AlphaFold.
4. 3. The method of claim 2, wherein the conformational dynamics are generated using GROMACS, OpenMM, NAMD, Amber, or LAMPS.
5. The method of claim 2 , wherein the generation of the conformational dynamics is parallelized.
6. The method of claim 1 , wherein the reference molecular structure is a three-dimensional structure of a wild-type protein.
7. 7. The method of any one of claims 1 to 6, wherein each of step (i) and / or step (ii) of claim 1 independently comprises at least about 10,000, at least about 20,000, at least about 30,000, at least about 40,000, at least about 50,000, at least about 60,000, at least about 70,000, at least about 80,000, at least about 90,000, at least about 100,000, at least about 200,000, at least about 300,000, at least about 400,000, at least about 500,000, at least about 600,000, at least about 700,000, at least about 800,000, at least about 900,000, or at least about 1 million simulated structural conformations comprising said conformational dynamics.
8. 2. The method of claim 1 , wherein processing the conformational dynamics to generate a reduced-dimensional representation of the conformational landscape of the reference molecular structure comprises extracting structural features from the conformational dynamics.
9. 9. The method of claim 8, wherein the structural features are selected from the group consisting of bond angles, inter-residue distances, residue variations, surface accessibility, domain relative orientation, binding site interactions, and any combination thereof.
10. 10. The method of claim 8 or 9, wherein said extraction of structural features results in at least a 20-fold, 30-fold, 40-fold, or 50-fold reduction in structural features.
11. The method of any one of claims 8 to 10, wherein processing the conformational dynamics further comprises dimensionality reduction.
12. The method of claim 11 , wherein the dimensionality reduction is obtained using Principal Component Analysis (PCA) or an autoencoder.
13. 13. The method of claim 11 or 12, wherein the dimensionality reduction results in at least a 50-fold, 60-fold, 70-fold, 80-fold, 90-fold, or 100-fold reduction in structural features.
14. 14. The method of any one of claims 8 to 13, wherein the reduced dimensional representation of the conformational landscape identifies conformations in which the reference molecular structure and / or variant molecular structure are in metastable conformational states.
15. The method of claim 1 , wherein the low-dimensional representation comprises motional modes that characterize conformational changes associated with molecular function.
16. 17. The method of any one of claims 1 to 16, wherein extracting metastable conformational states from the low-dimensional conformational landscape comprises clustering conformational states based on their representation in a low-dimensional space.
17. 17. The method of claim 16, wherein clustering the conformational states comprises k-means clustering, hierarchical clustering, density-based clustering, DBSCAN, spectral clustering, Gaussian mixture models, or any combination thereof.
18. 18. The method of claim 16 or 17, further comprising prioritizing the clustered conformational states.
19. 20. The method of claim 18, wherein prioritizing the clustered conformational states is based on cluster properties.
20. 20. The method of claim 19, wherein the cluster properties are selected from the group consisting of distance from a conformational landscape centroid, cluster occupancy, cluster distribution statistics, structural properties of clusters, thermodynamics of clusters, or any combination thereof.
21. 21. The method of claim 20, wherein the structural characteristics are selected from the group consisting of bond angles, inter-residue distances, surface accessibility, domain relative orientation, binding site interactions, structural similarity to known protein conformations, and any combination thereof.
22. 22. The method of claim 21 , wherein the bond angles comprise φ-ψ dihedral angles, side chain χ angles, or a combination thereof.
23. 21. The method of claim 20, wherein the cluster distribution statistic is selected from the group consisting of Silhouette score, Elbow score, Calinski-Harabasz index, Rand index, mutual information, homogeneity, completeness, V-scale, Davies-Bouldin index, and any combination thereof.
24. 24. The method of any one of claims 1 to 23, wherein extracting metastable conformational states from the low-dimensional conformational landscape comprises extracting representative conformational states.
25. 25. The method of claim 24, wherein the representative conformational state is the conformational state closest to the center of gravity of all conformational states of the cluster.
26. 2. The method of claim 1, wherein at least one variant is a pathogenic mutation.
27. 2. The method of claim 1, wherein at least one variant is a non-pathogenic (benign) mutation.
28. 10. The method of claim 1, wherein the variant molecular structure or set thereof comprises variants observed or observable in a medical genetic test.
29. 2. The method of claim 1, wherein the variant molecular structure or set thereof comprises a pathogenic variant.
30. The method of claim 1 , wherein the variant molecular structure or set thereof comprises a non-pathogenic (benign) variant.
31. The method of claim 1 , wherein the variant molecular structure or set thereof comprises a variant of unknown significance.
32. The method of claim 1 , wherein the variant molecular structure or set thereof is generated using a backbone-dependent rotamer library.
33. 33. The method of claim 32, comprising identifying side-chain conformations (rotamers) with existing backbone dihedral angles of said variant molecular structures at variant positions in said backbone-dependent rotamer library.
34. 34. The method of claim 33, further comprising testing whether introduction of a rotamer-defined scaffold-dependent rotamer library causes steric clashes with nearby residues.
35. 35. The method of claim 34, wherein testing is carried out until a rotamer that minimizes steric clashes is identified.
36. 35. The method of claim 34, further comprising performing minimization and equilibration simulations to regularize the conformational geometry of the variant molecular structure and reduce the energetic impact of introducing the variant into the variant molecular structure.
37. The method of claim 1 , wherein the predictive model is generated using machine learning.
38. 2. The method of claim 1, wherein the clinical data used as training labels includes biomarker status (e.g., the presence or absence or expression level of a particular biomarker), biometric data, lifestyle related data, response to treatment, symptoms of a disease or condition, protein expression data, type of treatment administered, dosage, dosage regimen, route of administration, presence or absence of concomitant therapy, response to therapy, age, weight, sex, ethnicity, ClinVar submissions, INVITAE™ clinical interpretations, data from other experimental or computational models, or any combination thereof.
39. 2. The method of claim 1, wherein the classification of the variant effect in the reference molecular structure comprises calculating a predicted pathogenicity probability.
40. The method of claim 1 , wherein the machine learning predictive model is generated using logistic regression.
41. The method of claim 1 , wherein the machine learning predictive model is generated using a random forest.
42. The method of claim 1 , wherein the machine learning predictive model is generated using an artificial neural network.
43. 1. A computer-implemented method for scoring the clinical significance of a variant, comprising: (a) generating conformational dynamics of a reference molecular structure (e.g., using MD); (b) processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; and (c) extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; and (d) generating a variant molecular structure or a set thereof by introducing at least one variant into the extracted metastable conformational state of the reference molecular structure; (e) generating (e.g., using MD) conformational dynamics of said variant molecular structure or set thereof; (f) processing the conformational dynamics of the variant molecular structure or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or set thereof; (g) training a machine learning predictive model using the reduced dimensional representation of the conformational landscapes of the reference and variant molecular structures using clinical data as training labels, wherein the predictive model outputs a score of the clinical significance of the variant.
44. 1. A system comprising: (i) a memory; and (ii) at least one processor, coupled to the memory; i. generating conformational dynamics of a reference molecular structure (e.g., using MD); ii. processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; iii. Extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; iv. generating a variant molecular structure or a set thereof by introducing at least one variant into the extracted metastable conformational state of the reference molecular structure; v. generating (e.g., using MD) conformational dynamics of said variant molecular structure or set thereof; vi. processing the conformational dynamics of the variant molecular structure or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or set thereof; and vii. at least one processor configured to train a machine learning predictive model using the reduced dimensional representation of the conformational landscapes of the reference and variant molecular structures using clinical data as training labels, wherein the predictive model classifies variant effects in the reference molecular structures.
45. A tangible computer readable device having instructions stored thereon that, when executed by at least one computing device, cause the at least one computing device to: i. generating conformational dynamics of a reference molecular structure (e.g., using MD); ii. processing the conformational dynamics of the reference molecular structure to generate a low-dimensional representation of a conformational landscape of the reference molecular structure; iii. Extracting metastable conformational states from the low-dimensional conformational landscape of the reference molecular structure to seed a simulation of a variant molecular structure; iv. generating a variant molecular structure or a set thereof by introducing at least one variant into said extracted metastable conformational state of a reference protein structure; v. generating (e.g., using MD) conformational dynamics of said variant molecular structure or set thereof; vi. processing the conformational dynamics of the variant molecular structure or set thereof to generate a low-dimensional representation of the conformational landscape of the variant molecular structure or set thereof; vii. A tangible computer readable device configured to perform operations including: training a machine learning predictive model using the reduced dimensional representation of the conformational landscapes of the reference and variant molecular structures using clinical data as training labels, wherein the predictive model classifies variant effects in the reference molecular structures.
46. 46. The method of any one of claims 1 to 43, the system of claim 44, or the tangible computer readable device of claim 45, wherein the classification of the variants in the variant molecular structure can be used to (i) treat a patient, (ii) select a patient for treatment, (iii) initiate treatment, (iv) discontinue treatment, (v) interrupt treatment, (vi) modify treatment, or (vii) any combination thereof.
47. 46. A method of (i) treating a patient, (ii) selecting a patient for treatment, (iii) initiating treatment, (iv) stopping treatment, (v) interrupting treatment, (vi) modifying treatment, or (vii) any combination thereof, comprising applying the method of any one of claims 1 to 43, the system of claim 44, or the tangible computer readable device of claim 45.
48. 46. A personalized medicine treatment comprising a therapeutic agent capable of treating a disease or disorder caused by a variant identified as pathogenic by the method of any one of claims 1 to 43, or by using the system of claim 44 or the tangible computer readable device of claim 45.
49. 1. A computer-implemented method for generating a reduced-dimensional representation of a conformational landscape of a molecular structure, the method comprising: (i) extracting structural features from the conformational landscape of the molecular structure; and (ii) reducing the dimensionality of the conformational landscape.
50. 50. The method of claim 49, wherein the structural features are selected from the group consisting of bond angles, inter-residue distances, residue variation, surface accessibility, domain relative orientation, binding site interactions, and any combination thereof.
51. 51. The method of claim 50, wherein extracting the structural feature results in at least a 20-fold, 30-fold, 40-fold, or 50-fold reduction in the structural feature.
52. 50. The method of claim 49, wherein the reducing dimensionality comprises using principal component analysis (PCA) or an autoencoder.
53. 50. The method of claim 49, wherein the reducing dimensionality results in at least a 50-fold, 60-fold, 70-fold, 80-fold, 90-fold, or 100-fold reduction in structural features.
54. 1. A computer-implemented method for identifying metastable conformational states of a molecular structure, the method comprising clustering a low-dimensional representation of the conformational landscape of the molecular structure.
55. 55. The method of claim 54, wherein clustering the conformational states comprises k-means clustering, hierarchical clustering, density-based clustering, DBSCAN, spectral clustering, Gaussian mixture models, or any combination thereof.
56. 56. The method of claim 54 or 55, further comprising prioritizing the clustered conformational states.
57. 57. The method of claim 56, wherein prioritizing the clustered conformational states is based on cluster properties.
58. 58. The method of claim 57, wherein the cluster properties are selected from the group consisting of distance from a conformational landscape centroid, cluster occupancy, cluster distribution statistics, structural properties of clusters, thermodynamics of clusters, or any combination thereof.
59. 59. The method of claim 58, wherein the structural characteristics are selected from the group consisting of bond angles, inter-residue distances, surface accessibility, domain relative orientation, binding site interactions, structural similarity to known protein conformations, and any combination thereof.
60. 60. The method of claim 59, wherein the bond angles comprise φ-ψ dihedral angles, side chain χ angles, or a combination thereof.
61. 59. The method of claim 58, wherein the cluster distribution statistic is selected from the group consisting of Silhouette score, elbow score, Calinski-Harabasz index, Rand index, mutual information, homogeneity, completeness, V-scale, Davies-Bouldin index, and any combination thereof.
62. 1. A computer-implemented method for high-throughput classification of variant effects in a molecular structure, the method comprising: (i) generating a low-dimensional representation of a conformational landscape of the molecular structure; and (ii) identifying metastable conformational states from the low-dimensional representation of the conformational landscape of the molecular structure.
63. 63. The method of claim 62, wherein generating a reduced dimensional representation of the conformational landscape of the molecule comprises: (a) extracting structural features from the conformational landscape of the molecular structure; and (b) reducing the dimensionality of the conformational landscape.
64. 63. The method of claim 62, wherein identifying metastable conformational states from the reduced-dimensional representation of the conformational landscape of the molecular structure comprises clustering the reduced-dimensional representation of the conformational landscape of the molecular structure.
