Deep convolutional neural networks for predicting variant pathogenicity using three-dimensional (3D) protein structures

Abstract

The disclosed technology relates to determining the pathogenicity of variants. In particular, the disclosed technology relates to generating amino acid-specific distance channels for multiple amino acids within a protein. Each amino acid-specific distance channel has voxel-specific distance values ​​for voxels within multiple voxels. A tensor comprises the amino acid-specific distance channels and at least alternative alleles of the protein expressed by the variant. A deep convolutional neural network determines the pathogenicity of the variant based at least partially on processing the tensor. The disclosed technology further augments the tensor with complementary information, such as a reference allele of the protein, evolutionary conservation data regarding the protein, annotation data regarding the protein, and structural confidence data regarding the protein.

Description

Technology Field

[0001] Priority application

[0002] This application claims priority to U.S. Regular Patent Application No. 17 / 468,411 (Agent Document No. ILLM 1037-3 / IP-2051A-US), filed September 7, 2021, with the title "Deep Convolutional Neural Networks to Predict Variant Pathogenicity using Three-Dimensional (3D) Protein Structures," which is a continuation of U.S. Regular Patent Application No. 17 / 232,056 (Agent Document No. ILLM 1037-2 / IP-2051-US), filed April 15, 2021, with the title "Deep Convolutional Neural Networks to Predict Variant Pathogenicity using Three-Dimensional (3D) Protein Structures."

[0003] This application claims priority to U.S. Regular Patent Application No. 17 / 703,935 (Agent Document No. ILLM 1047-2 / IP-2142-US), filed on March 24, 2022, with the title of invention "Multi-channel Protein Voxelization To Predict Variant Pathogenicity Using Deep Convolutional Neural Networks," which in turn claims priority or interest therein to U.S. Provisional Patent Application No. 63 / 175,495 (Agent Document No. ILLM 1047-1 / IP-2142-PRV), filed on April 15, 2021, with the title of invention "Multi-channel Protein Voxelization To Predict Variant Pathogenicity Using Deep Convolutional Neural Networks."

[0004] This application claims priority to U.S. Regular Patent Application No. 17 / 703,958 (Agent Document No. ILLM 1048-2 / IP-2143-US), filed on March 24, 2022, with the title of the invention "Efficient Voxelization For Deep Learning", which in turn claims priority or interest therein to U.S. Provisional Patent Application No. 63 / 175,767 (Agent Document No. ILLM 1048-1 / IP-2143-PRV), filed on April 16, 2021, with the title of the invention "Efficient Voxelization For Deep Learning".

[0005] Priority applications are thus included by reference for all purposes.

[0006] Related applications

[0007] This application relates to a concurrently filed PCT patent application (Agent Document No. ILLM 1037-4 / IP-2051-PCT) with the title of invention "Artificial Intelligence-based Analysis of Protein Three-Dimensional (3D) Structures". The related application is incorporated by reference for all purposes.

[0008] Technology field

[0009] The disclosed technology relates to artificial intelligence type computers and digital data processing systems and corresponding data processing methods and products (i.e., knowledge base systems, inference systems, and knowledge acquisition systems), comprising systems for emulating intelligence; and for inference with uncertainty (e.g., fuzzy logic systems), adaptive systems, machine learning systems, and artificial neural networks. In particular, the disclosed technology relates to analyzing multi-channel voxelized data using deep convolutional neural networks.

[0010] References

[0011] The following is incorporated by reference for all purposes as fully explained in this specification:

[0012] Sundaram, L. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018)]; Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning.

[0013] Cell 176, 535-548 (2019)];

[0014] U.S. Patent Application No. 62 / 573,144, filed on October 16, 2017, with the title "TRAINING A DEEP PATHOGENICITY CLASSIFIER USING LARGE-SCALE BENIGN TRAINING DATA" (Agent Document No. ILLM 1000-1 / IP-1611-PRV);

[0015] U.S. Patent Application No. 62 / 573,149, filed on October 16, 2017, with the title of the invention "PATHOGENICITY CLASSIFIER BASED ON DEEP CONVOLUTIONAL NEURAL NETWORKS (CNNs)" (Agent Document No. ILLM 1000-2 / IP-1612-PRV);

[0016] U.S. Patent Application No. 62 / 573,153, filed on October 16, 2017, with the title of the invention "DEEP SEMI-SUPERVISED LEARNING THAT GENERATES LARGE-SCALE PATHOGENIC TRAINING DATA" (Agent Document No. ILLM 1000-3 / IP-1613-PRV);

[0017] U.S. Patent Application No. 62 / 582,898, filed on November 7, 2017, with the title of the invention "PATHOGENICITY CLASSIFICATION OF GENOMIC DATA USING DEEP CONVOLUTIONAL NEURAL NETWORKS (CNNs)" (Agent Document No. ILLM 1000-4 / IP-1618-PRV);

[0018] U.S. Patent Application No. 16 / 160,903, filed on October 15, 2018, with the title "DEEP LEARNING-BASED TECHNIQUES FOR TRAINING DEEP CONVOLUTIONAL NEURAL NETWORKS" (Agent Document No. ILLM 1000-5 / IP-1611-US);

[0019] U.S. Patent Application No. 16 / 160,986, filed on October 15, 2018, with the title "DEEP CONVOLUTIONAL NEURAL NETWORKS FOR VARIANT CLASSIFICATION" (Agent Document No. ILLM 1000-6 / IP-1612-US);

[0020] U.S. Patent Application No. 16 / 160,968, filed on October 15, 2018, with the title of the invention "SEMI-SUPERVISED LEARNING FOR TRAINING AN ENSEMBLE OF DEEP CONVOLUTIONAL NEURAL NETWORKS" (Agent Document No. ILLM 1000-7 / IP-1613-US); and

[0021] U.S. Patent Application No. 16 / 407,149, filed on May 8, 2019, with the title "DEEP LEARNING-BASED TECHNIQUES FOR PRE-TRAINING DEEP CONVOLUTIONAL NEURAL NETWORKS" (Agent Document No. ILLM 1010-1 / IP-1734-US). Background Technology

[0022] The subject matter discussed in this section should not be presumed to be prior art merely as a result of its reference within this section. Similarly, problems related to the subject matter mentioned in this section or provided as background art should not be assumed to have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, and the approaches themselves may also correspond to embodiments of the claimed technology.

[0023] Genomics, also referred to in a broad sense as functional genomics, aims to characterize the functions of all genomic elements of an organism by using genome-scale analyses such as genome sequencing, transcriptome profiling, and proteomics. Genomics emerged as a data-driven science—it operates by discovering novel properties from the exploration of genome-scale data rather than testing preconceived models and hypotheses. Applications of genomics include finding associations between genotypes and phenotypes, discovering biomarkers for patient stratification, predicting gene functions, and charting biochemically active genomic regions such as transcriptional enhancers.

[0024] Genomic data is too large and too complex to be investigated solely through the visual study of pairwise correlations. Instead, analytical tools are required to support the discovery of unexpected relationships, to derive novel hypotheses and models, and to make predictions. Unlike some algorithms where assumptions and domain expertise are hardcoded, machine learning algorithms are designed to automatically detect patterns in data. Therefore, machine learning algorithms are well-suited for data-driven sciences, and particularly for genomics. However, the performance of machine learning algorithms can depend heavily on how the data is represented—that is, how each variable (also called a feature) is calculated. For example, to classify tumors as malignant or benign from fluorescence microscopy images, a preprocessing algorithm can detect cells, identify cell types, and generate a list of cell counts for each cell type.

[0025] Machine learning models can take estimated cell counts as input features for classifying tumors, which are examples of manually generated features. The central problem is that classification performance depends heavily on the quality and relevance of these features. For instance, relevant visual features, such as cell morphology, distances between cells, or locality within an organ, are not captured in the cell counts, and this incomplete representation of the data can reduce classification accuracy.

[0026] Deep learning, a subcategory of machine learning, addresses this problem by producing end-to-end models that embed the computation of features into the machine learning models themselves. This result was realized through the development of deep neural networks, which are machine learning models involving a series of fundamental operations that compute increasingly complex features by taking the results of preceding operations as input. Deep neural networks can improve prediction accuracy by discovering relevant features of high complexity, such as the cell morphology and spatial organization of cells in the example above. The construction and training of deep neural networks have been made possible by the explosion of data, algorithmic advancements, and substantial increases in computational capacity, particularly through the use of graphical processing units (GPUs).

[0027] The goal of supervised learning is to obtain a model that takes features as input and returns a prediction for a so-called target variable. An example of a supervised learning problem is predicting whether an intron will be splice-out (target) by considering features on RNA, such as the presence of canonical splice site sequences, the location of splicing branches, or intron length. Training a machine learning model refers to learning its parameters, which typically involves minimizing a loss function on the training data for the purpose of making accurate predictions on unseen data.

[0028] For many supervised learning problems in computational biotechnology, input data can be represented as a table having multiple columns or features, each containing potentially useful numerical or categorical data for making predictions. While some input data is naturally represented as features within the table (e.g., temperature or time), other input data (such as deoxyribonucleic acid (DNA) sequences into k-mer counts) must first be transformed using a process called feature extraction to fit the tabulated representation. In the case of intron splicing prediction problems, the presence or absence of standard splice site sequences, the location of splicing branches, and intron lengths may be pre-processed features collected in a tabulated format. Tabulated data is the standard for a wide range of supervised machine learning models, ranging from simple linear models such as logistic regression to more flexible nonlinear models such as neural networks and many others.

[0029] Logistic regression is a binary classifier, that is, a supervised learning model that predicts binary target variables. Specifically, logistic regression predicts the probability of a positive class by calculating a weighted sum of input features mapped to the [0,1] interval using a sigmoid function, a certain type of activation function. The parameters of logistic regression, or other linear classifiers using different activation functions, are weights of the weighted sum. Linear classifiers fail when classes, for example, those of splice-out or non-splice-out introns, cannot be well distinguished by the weighted sum of input features. To improve prediction performance, new input features can be manually added by transforming or combining existing features in new ways, for example, by taking squares or pairwise products.

[0030] Neural networks automatically learn these non-linear feature transformations using hidden layers. Each hidden layer can be thought of as multiple linear models whose outputs are transformed by non-linear activation functions, such as the sigmoid function or the more popular rectified-linear unit (ReLU). Together, these layers organize input features into related complex patterns, which facilitate the task of distinguishing between two classes.

[0031] Deep neural networks utilize many hidden layers, and a layer is considered fully connected when each neuron receives inputs from all neurons in the preceding layer. Neural networks are generally trained using stochastic gradient descent, an algorithm suitable for training models on very large datasets. Implementations of neural networks using modern deep learning frameworks enable rapid prototyping with different architectures and datasets. Fully connected neural networks can be used in numerous genetics applications, such as predicting the percentage of splice-in exons for a given sequence from sequence features like the presence of sequence conservation or splice factor binding motifs; prioritizing potential disease-causing gene variants; and using features such as chromatin marks, gene expression, and evolutionary conservation within a given genomic region. cis - Includes predicting control factors.

[0032] For effective predictions, local dependencies in spatial and longitudinal data must be considered. For example, shuffling DNA sequences or image pixels severely destroys informative patterns. These local dependencies establish spatial or longitudinal data separate from tabulated data, where the ordering of features for this purpose is arbitrary. Consider the problem of classifying genomic regions into bound and unbound by specific transcription factors, where bound regions are defined as high-reliability binding events in chromatin immunoprecipitation followed by ChIP-seq data. Transcription factors bind to DNA by recognizing sequence motifs. Within the sequence k A fully connected layer based on sequence derivation features, such as the number of dimer instances or position weight matrix (PWM) matchings, can be used for this task. Therefore, k Because dimer or PWM instance frequencies are robust to shifting motifs within a sequence, such models can generalize wells for sequences containing the same motif located at different positions. However, they will not recognize patterns where transcription factor binding depends on combinations of multiple motifs with well-defined intervals. Also, possible k - The number of quanta k - It increases exponentially with the length of the dimer, which raises both storage and overfitting problems.

[0033] Convolutional layers are a special form of fully connected layers, where the same fully connected layer is applied locally to all sequence locations, for example, within a 6 bp window. This approach can also be seen as scanning the sequence using multiple PWMs, for example, for the transcription factors GATA1 and TAL1. By using the same model parameters across the locations, the total number of parameters is drastically reduced, and the network can detect motifs at locations that were not visible during training. Each convolutional layer scans the sequence with multiple filters by generating scalar values ​​at all locations that quantify the matching between the filter and the sequence. As in fully connected neural networks, a non-linear activation function (typically ReLU) is applied to each layer. Next, a pooling operation is applied, which aggregates activations from adjacent bins across the location axis, typically taking the maximum or average activation for each channel. Pooling reduces the effective sequence length and makes the signal coarser. Subsequent convolution layers constitute the output of the previous layers and can detect whether the GATA1 motif and the TAL1 motif existed within a certain distance range. Finally, the outputs of the convolution layers can be used as inputs to a fully connected neural network to perform a final prediction task. Thus, different types of neural network layers (e.g., fully connected layers and convolution layers) can be combined within a single neural network.

[0034] Convolutional neural networks (CNNs) can predict various molecular phenotypes based solely on DNA sequences. Applications include classifying transcription factor binding sites and predicting molecular phenotypes such as chromatin features, DNA contact maps, DNA methylation, gene expression, translation efficiency, RBP binding, and microRNA (miRNA) targets. In addition to predicting molecular phenotypes from sequences, CNNs can be applied to more technical tasks traditionally handled by manually written bioinformatics pipelines. For example, CNNs can predict guide RNA specificity, denoise ChIP-seq, improve Hi-C data resolution, predict the origin of experiments from DNA sequences, and identify gene variants. CNNs have also been employed to model long-range dependencies in the genome. Although interacting regulatory elements may be located at a distance on the unfolded linear DNA sequence, these elements are often proximal to the actual 3D chromatin structure. Therefore, modeling molecular phenotypes from linear DNA sequences can be improved by allowing for long-range dependencies and enabling the model to implicitly learn solar terms of 3D configurations, such as promoter-enhancer looping, despite the rough approximation of chromatin. This is achieved by using extended convolutions with receptive fields of up to 32 kb. Extended convolutions also allow splice sites to be predicted from the sequence using a receptive field of 10 kb, thereby enabling the integration of gene sequences across distances as long as typical human introns (Reference [Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning.[Refer to Cell 176, 535-548 (2019)]).

[0035] Different types of neural networks can be characterized by their parameter sharing schemes. For example, fully connected layers do not share parameters, whereas convolutional layers impose translation invariance by applying the same filters at every position of their input. Recurrent neural networks (RNNs) are an alternative to convolutional neural networks for processing sequential data, such as DNA sequences or time series, which implement different parameter sharing schemes. Recurrent neural networks apply the same operation to each sequence element. The operation takes the memory of the previous sequence element and the new input as inputs. It updates the memory and selectively emits an output that is passed to subsequent layers or used directly as model predictions. By applying the same model to each sequence element, recurrent neural networks are invariant with respect to position indices in the processed sequence. For example, a recurrent neural network can detect open read frames in a DNA sequence regardless of their position within the sequence. This task requires the recognition of a specific series of inputs, such as an in-frame stop codon following a start codon.

[0036] The primary advantage of recurrent neural networks over convolutional neural networks is that, theoretically, they can pass information through infinitely long sequences via memory. Additionally, recurrent neural networks can naturally process sequences of widely varying lengths, such as mRNA sequences. However, convolutional neural networks combined with various tricks (e.g., extended convolutions) can achieve performance comparable to or even better than recurrent neural networks for sequence modeling tasks such as audio synthesis and machine translation. Recurrent neural networks can aggregate the outputs of convolutional neural networks to predict single-cell DNA methylation states, RBP binding, transcription factor binding, and DNA accessibility. Furthermore, because recurrent neural networks apply sequential operations, they cannot be easily parallelized and are therefore much slower to compute than convolutional neural networks.

[0037] While each human possesses a unique genetic code, most of the human genetic code is common to all humans. In some cases, the human genetic code may contain outliers called genetic variants, which may be common among individuals in relatively small groups of humans. For example, a specific human protein may contain amino acids of a specific sequence, whereas variants of that protein may differ by just one amino acid within the same specific sequence.

[0038] Gene variants are pathogenic and can lead to diseases. While most of these variants have been exhausted from genomes by natural selection, the ability to identify which variants are potentially pathogenic can help researchers focus on these variants to gain an understanding of the corresponding diseases and their diagnosis, treatment, or cure. The clinical interpretation of millions of human gene variants remains unclear. Some of the most frequent pathogenic variants are single nucleotide missense mutations that alter the amino acids of proteins. However, not all missense mutations are pathogenic.

[0039] Models capable of directly predicting molecular phenotypes from biological sequences can be used as in silico perturbation tools to probe associations between genetic and phenotypic variations, and have emerged as novel methods for identifying quantitative trait loci and prioritizing variants. These approaches are particularly important when the majority of variants identified by genome-wide association studies of complex phenotypes are non-coding, which makes it difficult to estimate their effects and contributions to the phenotypes. Furthermore, linkage imbalances result in blocks of variants being co-inherited, causing difficulties in accurately identifying individual causal variants. Therefore, sequence-based deep learning models, which can be used as interrogation tools to evaluate the impact of such variants, offer a promising approach for discovering potential drivers of complex phenotypes. One example involves indirectly predicting the effects of short insertions or deletions (indels) and non-coding single nucleotide variants from differences between two variants in terms of transcription factor binding, chromatin accessibility, or gene expression predictions. Another example involves predicting the generation of novel splice sites from the sequence or quantitative effects of gene variants on splicing.

[0040] End-to-end deep learning approaches for predicting variant effects are applied to predict the pathogenicity of missense variants from sequence-conserving data and protein sequences (see the literature referred to herein as "PrimateAI" [Sundaram, L. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018)]). PrimateAI uses deep neural networks trained on variants of known pathogenicity through data augmentation using cross-species information. In particular, PrimateAI uses sequences of wild-type and variant proteins to compare differences and determine the pathogenicity of mutations using trained deep neural networks. Such an approach utilizing protein sequences for pathogenicity prediction is promising because it can avoid the circularity problem and overfitting to prior knowledge. However, compared to the appropriate number of data required to effectively train deep neural networks, the number of clinical data available in ClinVar is relatively small. To overcome this lack of data, PrimateAI uses common human variants and variants from primates as positive data, while simulated variants based on trinucleotide context were used as unlabeled data.

[0041] PrimateAI outperforms previous methods when trained directly on sequence alignments. PrimateAI learns important protein domains, conserved amino acid positions, and sequence dependencies directly from training data consisting of approximately 120,000 human samples. PrimateAI substantially surpasses the performance of other variant pathogenicity prediction tools in distinguishing benign and pathogenic neomutations in candidate developmental disorder genes and reproducing prior knowledge from ClinVar. These results suggest that PrimateAI represents a significant advance for variant classification tools capable of reducing reliance on clinical reports for prior knowledge.

[0042] The core of protein biology is understanding how structural elements generate observed functions. The abundance of protein structure data enables the development of computational methods to systematically derive the rules governing structural-functional relationships. However, the performance of these methods depends significantly on the choice of protein structure representation.

[0043] Protein regions are microenvironments within a protein structure distinguished by their structural or functional roles. A region can be defined by its three-dimensional (3D) location and the local neighborhoods surrounding that location where the structure or function exists. The core of rational protein engineering is understanding how the structural arrangement of amino acids generates functional characteristics within protein regions. Determining the structural and functional roles of individual amino acids within a protein provides engineers with information to assist in modifying protein functions. Identifying functionally or structurally important amino acids allows for focused engineering efforts, such as site-induced mutagenesis, to alter targeted protein functional properties. Alternatively, this knowledge can help avoid engineering designs that would negate desired functions.

[0044] Since it is established that structure is much more conserved than sequence, the increase in protein structure data provides an opportunity to systematically study the underlying patterns governing structural-functional relationships using data-driven approaches. A fundamental aspect of random computational protein analysis is how protein structure information is represented. The performance of machine learning methods often depends more on the choice of data representation than on the machine learning algorithm employed. Good representations efficiently capture the most critical information, while poor representations generate noisy distributions lacking underlying patterns.

[0045] The recent success of protein structure overload and deep learning algorithms provides an opportunity to develop tools for automatically extracting task-specific representations of protein structures. Consequently, an opportunity arises to predict variant pathogenicity using multi-channel voxelized representations of 3D protein structures as input to deep neural networks. Brief explanation of the drawing

[0046] In the drawings, similar reference numerals generally refer to similar parts across different drawings. Additionally, the drawings are not necessarily to scale; instead, they are generally emphasized to illustrate the principles of the disclosed technology. In the following description, various embodiments of the disclosed technology are described with reference to the drawings below. FIG. 1 is a flowchart illustrating the process of a system for determining the pathogenicity of variants according to various embodiments of the disclosed technology. FIG. 2 schematically illustrates an exemplary reference amino acid sequence of a protein and an alternative amino acid sequence of a protein according to one embodiment of the disclosed technology. FIG. 3 illustrates the amino acid-specific classification of amino acid atoms within the reference amino acid sequence of FIG. 2 according to one embodiment of the disclosed technology. FIG. 4 illustrates the amino acid-specific properties of the 3D atomic coordinates of alpha-carbon atoms classified in FIG. 3 in amino acid units, according to one embodiment of the disclosed technology. FIG. 5 schematically illustrates a process for determining distance values ​​per voxel according to one embodiment of the disclosed technology. FIG. 6 illustrates an example of 21 amino acid-specific distance channels according to one embodiment of the disclosed technology. FIG. 7 is a schematic diagram of a distance channel tensor according to one embodiment of the disclosed technology. FIG. 8 illustrates one-hot encodings of a reference amino acid and an alternative amino acid from FIG. 2 according to one embodiment of the disclosed technology. FIG. 9 is a schematic diagram of a voxelized one-hot encoded reference amino acid and a voxelized one-hot encoded variant / alternative amino acid according to one embodiment of the disclosed technology. FIG. 10 schematically illustrates a connection process that connects the distance channel tensor and reference allele tensor of FIG. 7 on a voxel-by-voxel basis, according to one embodiment of the disclosed technology. FIG. 11 schematically illustrates a connection process that connects the distance channel tensor of FIG. 7, the reference allele tensor of FIG. 10, and an alternative allele tensor on a voxel-by-voxel basis, according to one embodiment of the disclosed technology. FIG. 12 is a flowchart illustrating the process of a system for determining the pan-amino acid conservation frequencies of the nearest atoms and assigning them to voxels (voxelization) according to one embodiment of the disclosed technology. FIG. 13 illustrates a voxel-to-closest amino acid according to one embodiment of the disclosed technology. FIG. 14 illustrates an exemplary multiple sequence alignment of reference amino acid sequences spanning 99 types according to one embodiment of the disclosed technology. FIG. 15 illustrates an example of determining a pan-amino acid conservation frequency sequence for a specific voxel according to one embodiment of the disclosed technology. FIG. 16 illustrates the respective pan-amino acid conservation frequencies determined for each voxel using the position frequency logic described in FIG. 15, according to one embodiment of the disclosed technology. FIG. 17 illustrates evolutionary profiles per voxel of voxelized voxels according to one embodiment of the disclosed technology. FIG. 18 illustrates an example of an evolutionary profile tensor according to one embodiment of the disclosed technology. FIG. 19 is a flowchart illustrating the process of a system for determining the conservation frequencies per amino acid of the nearest atoms and assigning them to voxels (voxelization) according to one embodiment of the disclosed technology. FIG. 20 illustrates various examples of voxelized annotation channels connected to a distance channel tensor according to one embodiment of the disclosed technology. FIG. 21 illustrates different combinations and permutations of input channels that may be provided as inputs to a pathogenicity classifier for determining the pathogenicity of a target variant, according to one embodiment of the disclosed technology. FIG. 22 illustrates different methods for calculating disclosed distance channels according to various embodiments of the disclosed technology. FIG. 23 illustrates different examples of evolutionary channels according to various embodiments of the disclosed technology. FIG. 24 illustrates different examples of annotation channels according to various embodiments of the disclosed technology. FIG. 25 illustrates different examples of structural reliability channels according to various embodiments of the disclosed technology. FIG. 26 illustrates an exemplary processing architecture of a pathogenicity classifier according to one embodiment of the disclosed technology. FIG. 27 illustrates an exemplary processing architecture of a pathogenicity classifier according to one embodiment of the disclosed technology. Figures 28, 29, 30, and 31 use PrimateAI as a benchmark model to demonstrate the classification superiority of the disclosed PrimateAI 3D over PrimateAI. FIGS. 32a and FIGS. 32b illustrate a disclosed efficient voxelization process according to various embodiments of the disclosed technology. FIG. 33 illustrates a method of associating atoms with voxels containing atoms according to one embodiment of the disclosed technology. FIG. 34 illustrates generating a voxel-atom mapping from an atomic-voxel mapping to identify the nearest atoms on a voxel-by-voxel basis, according to one embodiment of the disclosed technology. FIGS. 35a and FIGS. 35b illustrate a method in which the disclosed efficient voxelization has a runtime complexity of O (#atoms) versus O (#atoms*#voxels) without using the disclosed efficient voxelization. FIG. 36 illustrates an exemplary computer system that can be used to implement the disclosed technology. Specific details for implementing the invention

[0047] The following discussion is presented to enable any person skilled in the art to manufacture and use the disclosed technology, and is provided in relation to specific applications and requirements thereof. Various variations of the disclosed embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the disclosed technology. Accordingly, the disclosed technology is not intended to be limited to the illustrated embodiments, but should be conformed to the broadest category consistent with the principles and features disclosed herein.

[0048] Detailed descriptions of various embodiments will be better understood when read in conjunction with the attached drawings. To the extent that the drawings illustrate functional block diagrams of various embodiments, the functional blocks do not necessarily represent divisions between hardware circuits. Thus, for example, one or more of the functional blocks (e.g., modules, processors, or memories) may be implemented in a single piece of hardware (e.g., blocks of general-purpose signal processors or random access memory, hard disks, etc.) or in multiple pieces of hardware. Similarly, programs may be standalone programs, integrated as subroutines into an operating system, functions within an installed software package, and so on. It will be understood that the various embodiments are not limited to the arrangements and means depicted in the drawings.

[0049] The processing engines and databases of the drawings, designated as modules, may be implemented in hardware or software and do not need to be divided into exactly the same blocks as depicted in the drawings. Some of the modules may also be implemented on different processors, computers, or servers, or distributed among a number of different processors, computers, or servers. It will also be understood that some of the modules may be combined with those depicted in the drawings and operated in parallel or in a different order without affecting the achieved function. The modules within the drawings may also be conceived as flowchart steps in a method. A module also does not necessarily have its entire code placed adjacent to memory; parts of the code may be separated from other parts of the code, with code from other modules or other functions placed in between.

[0050] Protein structure-based pathogenicity determination

[0051] FIG. 1 is a flowchart illustrating the process (100) of a system for determining the pathogenicity of variants. In step (102), the sequence accessor (104) of the system accesses reference and alternative amino acid sequences. In 112, the 3D structure generator (114) of the system generates 3D protein structures for the reference amino acid sequence. In some embodiments, the 3D protein structures are homology models of human proteins. In one embodiment, the so-called SwissModel homology modeling pipeline provides a public repository of predicted human protein structures. In another embodiment, the so-called HHpred homology modeling predicts the structure of a target protein from template structures using a tool called a modeler.

[0052] Proteins are represented in 3D space as sets of atoms and their coordinates. Amino acids can contain various atoms, such as carbon atoms, oxygen (O) atoms, nitrogen (N) atoms, and hydrogen (H) atoms. Atoms can be further classified into side chain atoms and backbone atoms. Backbone carbon atoms are alpha-carbons (C α ) atoms and beta-carbon (C β It can include ) atoms.

[0053] In step (122), the coordinate classifier (124) of the system classifies the 3D atomic coordinates of the 3D protein structures by amino acid. In one embodiment, the amino acid-based classification involves assigning the 3D atomic coordinates to 21 amino acid categories (including stop or gap amino acid categories). In one example, the amino acid-based classification of alpha-carbon atoms may enumerate the alpha-carbon atoms under each of the 21 amino acid categories. In another example, the amino acid-based classification of beta-carbon atoms may enumerate the beta-carbon atoms under each of the 21 amino acid categories.

[0054] In another example, the classification of oxygen atoms by amino acid can list oxygen atoms under each of the 21 amino acid categories. In another example, the classification of nitrogen atoms by amino acid can list nitrogen atoms under each of the 21 amino acid categories. In another example, the classification of hydrogen atoms by amino acid can list hydrogen atoms under each of the 21 amino acid categories.

[0055] Those skilled in the art will understand that, in various embodiments, the amino acid classification may include a subset of 21 amino acid categories and a subset of different atomic elements.

[0056] In step (132), the voxel grid generator (134) of the system instantiates a voxel grid. The voxel grid may have any resolution, for example, 3x3x3, 5x5x5, 7x7x7, etc. The voxels within the voxel grid may be of any size, for example, 1 angstrom (Å) on each side, 2 Å on each side, 3 Å on each side, etc. A person skilled in the art will understand that these exemplary dimensions refer to cubic dimensions because the voxels are cubes. Also, a person skilled in the art will understand that these exemplary dimensions are not limited and that the voxels may have any cubic dimensions.

[0057] In step (142), the voxel grid centerer (144) of the system centers the voxel grid at the reference amino acid that experiences the target variant at the amino acid level. In one embodiment, the voxel grid is centered at the atomic coordinates of a specific atom of the reference amino acid that experiences the target variant, for example, the 3D atomic coordinates of the alpha-carbon atom of the reference amino acid that experiences the target variant.

[0058] Street channels

[0059] Voxels within a voxel grid may have multiple channels (or features). In one embodiment, voxels within a voxel grid have multiple distance channels (e.g., 21 distance channels for each of the 21 amino acid categories (including stop or gap amino acid categories)). In step (152), a distance channel generator (154) of the system generates amino acid-specific distance channels for the voxels within the voxel grid. Distance channels are generated independently for each of the 21 amino acid categories.

[0060] For example, consider the alanine (A) amino acid category. For example, consider further that the voxel grid is 3x3x3 in size and has 27 voxels. Then, in one embodiment, the alanine distance channel contains 27 distance values ​​for each of the 27 voxels within the voxel grid. The 27 distance values ​​in the alanine distance channel are measured from the respective centers of the 27 voxels within the voxel grid to their respective nearest atoms within the alanine amino acid category.

[0061] In one example, the alanine amino acid category contains only alpha-carbon atoms, and thus, the closest atoms are those alanine alpha-carbon atoms closest to each of the 27 voxels in the voxel grid. In another example, the alanine amino acid category contains only beta-carbon atoms, and thus, the closest atoms are those alanine beta-carbon atoms closest to each of the 27 voxels in the voxel grid.

[0062] In another example, the alanine amino acid category contains only oxygen atoms, and therefore, the closest atoms are the alanine oxygen atoms closest to each of the 27 voxels within the voxel grid. In another example, the alanine amino acid category contains only nitrogen atoms, and therefore, the closest atoms are the alanine nitrogen atoms closest to each of the 27 voxels within the voxel grid. In another example, the alanine amino acid category contains only hydrogen atoms, and therefore, the closest atoms are the alanine hydrogen atoms closest to each of the 27 voxels within the voxel grid.

[0063] Similar to the alanine distance channel, the distance channel generator (154) generates a distance channel (i.e., a set of distance values ​​per voxel) for each of the remaining amino acid categories. In other embodiments, the distance channel generator (154) generates distance channels for only a subset of 21 amino acid categories.

[0064] In other embodiments, the selection of the nearest atoms is not limited to a specific atom type. That is, within a target amino acid category, the nearest atom to a specific voxel is selected regardless of the atomic element of the nearest atom and the distance value to the specific voxel calculated to be included in the distance channel for the target amino acid category.

[0065] In other embodiments, distance channels are generated on an atomic element basis. Instead of having distance channels for amino acid categories, or in addition to them, distance values ​​may be generated for atomic element categories regardless of the amino acid to which the atoms belong. For example, consider that the atoms of the amino acids in a reference amino acid sequence span seven atomic elements: carbon, oxygen, nitrogen, hydrogen, calcium, iodine, and sulfur. Subsequently, the voxels within the voxel grid are configured to have seven distance channels, and thus, each of the seven distance channels has 27 voxel-specific distance values ​​that specify distances only to the nearest atoms within the corresponding atomic element category. In other embodiments, distance channels may be generated for only a subset of the seven atomic elements. In other embodiments, the atomic element categories and the generation of distance channels are for the same atomic element, e.g., alpha-carbon (C α ) atoms and beta-carbon (C β ) It can be further stratified by variations of atoms.

[0066] In other embodiments, distance channels are generated on an atomic type basis, for example, distance channels for only side chain atoms and distance channels for only backbone atoms may be generated.

[0067] The nearest atoms can be searched within a predefined maximum scan radius (e.g., 6 angstroms (Å)) from the voxel centers. Additionally, multiple atoms may be closest to the same voxel within the voxel grid.

[0068] Distances are calculated between the 3D coordinates of the voxel centers and the 3D atomic coordinates of the atoms. Additionally, distance channels are generated as a voxel grid centered at the same location (e.g., centered at the 3D atomic coordinates of the alpha-carbon atom of the reference amino acid experiencing the target variant).

[0069] The distances may be Euclidean distances. Additionally, the distances may be parameterized by atomic size (or atomic influence) (e.g., by using the Lennard-Jones potential and / or Van der Waals atomic radius of the corresponding atom). Furthermore, the distance values ​​may be normalized by the maximum scan radius, or by the maximum observed distance value of the nearest atom within the target amino acid category, target atomic element category, or target atomic type category. In some embodiments, the distances between voxels and atoms are calculated based on the polar coordinates of the voxels and atoms. The polar coordinates are parameterized by the angles between the voxels and atoms. In one embodiment, this angle information is used to generate angle channels for the voxels (i.e., independent of the distance channels). In some embodiments, the angles between the nearest atom and neighboring atoms (e.g., backbone atoms) may be used as features encoded in the voxels.

[0070] Reference alleles and alternative allele channels

[0071] The voxels within the voxel grid may also have reference allele and alternative allele channels. In step (162), the one-hot encoder (164) of the system generates a reference one-hot encoding of the reference amino acid in the reference amino acid sequence and an alternative one-hot encoding of the alternative amino acid in the alternative amino acid sequence. The reference amino acid undergoes a target variant. The alternative amino acid is a target variant. The reference amino acid and the alternative amino acid are located at the same position in the reference amino acid sequence and the alternative amino acid sequence, respectively. The reference amino acid sequence and the alternative amino acid sequence have the same positional amino acid composition with one exception. The exception is a position where the reference amino acid is in the reference amino acid sequence and the alternative amino acid is in the alternative amino acid sequence.

[0072] In step (172), the linker (174) of the system links the amino acid-specific distance channels with reference and alternative one-hot encodings. In another embodiment, the linker (174) links the atomic element-specific distance channels with reference and alternative one-hot encodings. In yet another embodiment, the linker (174) links the atomic type-specific distance channels with reference and alternative one-hot encodings.

[0073] In step (182), the runtime logic (184) of the system determines the pathogenicity of a target variant by processing the linked amino acid-based / atomic element-based / atomic type-based distance channels and reference and alternative one-hot encodings through a pathogenicity classifier (pathogenicity determination engine), which is ultimately inferred as the pathogenicity determination of the base nucleotide variant generating the target variant at the amino acid level. The pathogenicity classifier is trained using labeled datasets of positive and pathogenic variants, for example, using a backpropagation algorithm. Further details regarding the labeled datasets of positive and pathogenic variants and exemplary architectures and training of the pathogenicity classifier can be found in the jointly owned U.S. Patent Applications No. 16 / 160,903; No. 16 / 160,986; No. 16 / 160,968; and No. 16 / 407,149.

[0074] FIG. 2 schematically illustrates a reference amino acid sequence (202) of protein (200) and an alternative amino acid sequence (212) of protein (200). Protein (200) contains N amino acids. The positions of the amino acids within protein (200) are labeled 1, 2, 3...N. In the illustrated example, position 16 is the position that experiences an amino acid variant (214) (mutation) caused by a base nucleotide variant. For example, in the reference amino acid sequence (202), position 1 has the reference amino acid phenylalanine (F), position 16 has the reference amino acid glycine (G) (204), and position N (e.g., the last amino acid of the sequence (202)) has the reference amino acid leucine (L). Although not illustrated for clarity, the remaining positions within the reference amino acid sequence (202) contain various amino acids in an order specific to protein (200). The alternative amino acid sequence (212) is identical to the reference amino acid sequence (202) except for the variant (214) at position 16, which contains the alternative amino acid alanine (A) (214) instead of the reference amino acid glycine (G) (204).

[0075] FIG. 3 illustrates the amino acid-specific classification of atoms of amino acids within a reference amino acid sequence (202), also referred to herein as "atomic classification (300)." Among the 20 natural amino acids listed in column (302), certain types of amino acids may be repeated in the protein. That is, certain types of amino acids may occur more than once in the protein. The protein may also have some undetermined amino acids categorized by the 21st-stop or gap amino acid category. The right column of FIG. 3 shows alpha-carbons (C) from different amino acids. α ) Contains a count of atoms.

[0076] Specifically, FIG. 3 shows the alpha-carbon (C) of an amino acid within a reference amino acid sequence (202). αFigure 3 illustrates the classification of atoms by amino acid. Column (308) of Figure 3 lists the total number of alpha-carbon atoms observed for the reference amino acid sequence (202) in each of the 21 amino acid categories. For example, column (308) lists 11 alpha-carbon atoms observed for the alanine (A) amino acid category. Since each amino acid has only one alpha-carbon atom, this means that alanine occurs 11 times in the reference amino acid sequence (202). In another example, arginine (R) occurs 35 times in the reference amino acid sequence (202). The total number of alpha-carbon atoms across the 21 amino acid categories is 828.

[0077] FIG. 4 illustrates the amino acid-specific properties of the 3D atomic coordinates of alpha-carbon atoms of a reference amino acid sequence (202) based on the atomic classification (300) of FIG. 3. This is referred to herein as "atomic coordinate bucketing" (400). In FIG. 4, lists (404 to 440) table the 3D atomic coordinates of alpha-carbon atoms bucketed into each of the 21 amino acid categories.

[0078] In the illustrated embodiment, the bucketing (400) of FIG. 4 follows the sorting (300) of FIG. 3. For example, in FIG. 3, the alanine amino acid category has 11 alpha-carbon atoms, and thus, in FIG. 4, the alanine amino acid category has 11 3D atomic coordinates of the corresponding 11 alpha-carbon atoms from FIG. 3. This sorting-bucketing logic also flows from FIG. 3 to FIG. 4 for other amino acid categories. However, this sorting-bucketing logic is for conceptual purposes only, and in other embodiments, the disclosed technique may not need to perform sorting (300) and bucketing (400) to position the nearest atoms per voxel, and may perform fewer, additional, or different steps. For example, in some embodiments, the disclosed technology may locate the nearest atoms by voxel by using a classification and search algorithm that returns the nearest atoms by voxel from one or more databases in response to a search query configured to accommodate classification criteria (e.g., by amino acid, by atomic element, by atomic type), a predefined maximum scan radius, and types of distances (e.g., Euclidean, Mahalanobis, normalized, denormalized). In various embodiments of the disclosed technology, a plurality of classification and search algorithms from current or future fields of technology may be similarly used by those skilled in the art to locate the nearest atoms by voxel.

[0079] In FIG. 4, 3D atomic coordinates are represented by orthogonal coordinates x, y, z, but any type of coordinate system, such as spherical or cylindrical coordinates, may be used, and the claimed subject matter is not limited in this respect. In some embodiments, one or more databases may contain information regarding the 3D atomic coordinates of alpha-carbon atoms and other atoms of amino acids within a protein. Such databases may be searchable by a specific protein.

[0080] As discussed above, voxels and voxel grids are 3D entities. However, for clarity, the drawings depict voxels and voxel grids in a two-dimensional (2D) format, and the description discusses this. For example, a 3x3x3 voxel grid of 27 voxels is depicted and described herein as a 3x3 2D pixel grid having 9 2D pixels. Those skilled in the art will understand that the 2D format is intended for figurative purposes only and to cover 3D counterparts (i.e., 2D pixels represent 3D voxels, and a 2D pixel grid represents a 3D voxel grid). Furthermore, the drawings are not to scale. For example, voxels of size 2 angstroms (Å) are depicted using a single pixel.

[0081] Calculate distance per voxel

[0082] FIG. 5 schematically illustrates a process for determining voxel-specific distance values, also referred to herein as "voxel-specific distance calculation (500)". In the illustrated example, voxel-specific distance values ​​are calculated only for the alanine (A) distance channel. However, the same distance calculation logic may be executed for each of the 21 amino acid categories to generate 21 amino acid-specific distance channels, and may be further extended to other atomic types such as beta-carbon atoms and other atomic elements such as oxygen, nitrogen, and hydrogen, as discussed above in relation to FIG. 1. In some embodiments, atoms are randomly rotated before distance calculation to make the training of the pathogenicity classifier invariant to atomic orientation.

[0083] In FIG. 5, the voxel grid (522) has nine voxels (514) identified by indices (1, 1), (1, 2), (1, 3), (2, 1), (2, 2), (2, 3), (3, 1), (3, 2), and (3, 3). The voxel grid (522) is centered, for example, at the 3D atomic coordinates (532) of the alpha-carbon atom of the glycine (G) amino acid at position 16 in the reference amino acid sequence (202), because, as discussed above in relation to FIG. 2, in an alternative amino acid sequence (212), position 16 experiences a variant in which the glycine (G) amino acid is mutated into an alanine (A) amino acid. Additionally, the center of the voxel grid (522) coincides with the center of the voxel (2, 2).

[0084] The centered voxel grid (522) is used to calculate voxel-specific distances for each of the 21 amino acid-specific distance channels. For example, starting with the alanine (A) distance channel, the distances between the 3D coordinates of the centers of the 9 voxels (514) and the 3D atomic coordinates (402) of the 11 alanine alpha-carbon atoms are measured to locate the nearest alanine alpha-carbon atom for each of the 9 voxels (514). Subsequently, 9 distance values ​​for the 9 distances between the 9 voxels (514) and their respective nearest alanine alpha-carbon atoms are used to construct the alanine distance channel. The generated alanine distance channel arranges the 9 alanine distance values ​​in the same order as the 9 voxels (514) within the voxel grid (522).

[0085] The above process is executed for each of the 21 amino acid categories. For example, a centered voxel grid (522) is similarly used to calculate the arginine (R) distance channel, so that the distances between the 3D coordinates of the respective centers of the 9 voxels (514) and the 3D atomic coordinates (404) of the 35 arginine alpha-carbon atoms are measured to locate the nearest arginine alpha-carbon atom to each of the 9 voxels (514). Subsequently, 9 distance values ​​for the 9 distances between the 9 voxels (514) and their respective nearest arginine alpha-carbon atoms are used to construct the arginine distance channel. The generated arginine distance channel arranges the 9 arginine distance values ​​in the same order as the 9 voxels (514) within the voxel grid (522). The 21 amino acid-specific distance channels are encoded by voxel to form a distance channel tensor.

[0086] Specifically, in the illustrated example, the distance (512) is the center of the voxel (1, 1) of the voxel grid (522) and Cα in the list (402). A5It is between the nearest alpha-carbon (Cα) atoms. Therefore, the value assigned to the voxel (1, 1) is the distance (512). In another example, Cα A4 The atom is the Cα atom closest to the center of voxel (1, 2). Therefore, the value assigned to voxel (1, 2) is the center of voxel (1, 2) and Cα A4 It is the distance between atoms. In another example, Cα A6 The atom is the Cα atom closest to the center of voxel (2, 1). Therefore, the value assigned to voxel (2, 1) is the center of voxel (2, 1) and Cα A6 It is the distance between atoms. In another example, Cα A6 The atom is also the Cα atom closest to the centers of voxels (3, 2) and (3, 3). Therefore, the value assigned to voxel (3, 2) is the center of voxel (3, 2) and Cα A6 It is the distance between, and the value assigned to voxel (3, 3) is the center of voxel (3, 3) and Cα A6 It is the distance between atoms. In some embodiments, the distance values ​​assigned to the voxels (514) may be normalized distances. For example, the distance value assigned to the voxel (1, 1) may be the distance (512) divided by the maximum distance (502) (a predefined maximum scan radius). In some embodiments, the nearest atomic distances may be Euclidean distances and the nearest atomic distances may be normalized by dividing the Euclidean distances by the maximum nearest atomic distance (e.g., the maximum distance (502)).

[0087] As previously described, for amino acids having alpha-carbon atoms, the distances may be the nearest alpha-carbon atom distances from the corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding amino acid. Additionally, for amino acids having beta-carbon atoms, the distances may be the nearest beta-carbon atom distances from the corresponding voxel centers to the nearest beta-carbon atoms of the corresponding amino acid. Similarly, for amino acids having backbone atoms, the distances may be the nearest backbone atom distances from the corresponding voxel centers to the nearest backbone atoms of the corresponding amino acid. Similarly, for amino acids having side chain atoms, the distances may be the nearest side chain atom distances from the corresponding voxel centers to the nearest side chain atoms of the corresponding amino acid. In some embodiments, the distances may additionally / alternatively include distances to a second, third, fourth nearest atom, etc.

[0088] Distance channels by amino acid

[0089] FIG. 6 illustrates an example of 21 amino acid-specific distance channels (600). Each column of FIG. 6 corresponds to its respective one of the 21 amino acid-specific distance channels (602 to 642). Each amino acid-specific distance channel contains distance values ​​for each of the voxels (514) of the voxel grid (522). For example, the amino acid-specific distance channel (602) for alanine (A) contains distance values ​​for its respective one of the voxels (514) of the voxel grid (522). As mentioned above, the voxel grid (522) is a 3D grid of volume 3x3x3 and contains 27 voxels. Likewise, although FIG. 6 illustrates the voxels (514) in two dimensions (e.g., 9 voxels of a 3x3 grid), each amino acid-specific distance channel may contain 27 voxel-specific distance values ​​for a 3x3x3 voxel grid.

[0090] Directional encoding

[0091] In some embodiments, the disclosed technique uses a directional parameter to specify the directional orientation of a reference amino acid within a reference amino acid sequence (202). In some embodiments, the disclosed technique uses a directional parameter to specify the directional orientation of an alternative amino acid within an alternative amino acid sequence (212). In some embodiments, the disclosed technique uses a directional parameter to specify the location of a protein (200) experiencing a target variant at the amino acid level.

[0092] As discussed above, all distance values ​​of the 21 amino acid-specific distance channels (602 to 642) are measured from their respective nearest atoms to the voxels (514) within the voxel grid (522). These nearest atoms originate from one of the reference amino acids in the reference amino acid sequence (202). These originating reference amino acids containing the nearest atoms may be classified into two categories: (1) such originating reference amino acid preceding the variant experience reference amino acid (204) in the reference amino acid sequence (202), and (2) such originating reference amino acid following the variant experience reference amino acid (204) in the reference amino acid sequence (202). The originating reference amino acid in the first category may be called the preceding reference amino acid. The originating reference amino acid in the second category may be called the following reference amino acid.

[0093] A directional parameter is applied to such distance values ​​in 21 amino acid-specific distance channels (602 to 642) measured from such nearest atoms derived from the preceding reference amino acid. In one embodiment, the directional parameter is multiplied by such distance values. The directional parameter can be any number, e.g. -1.

[0094] As a result of the application of directional parameters, 21 amino acid-specific distance channels (600) contain some distance values ​​that indicate to the pathogenicity classifier which end of the protein (200) is the start end and which end is the end end. This also enables the pathogenicity classifier to reconstruct the protein sequence from 3D protein structure information supplied by the distance channels and reference and allele channels.

[0095] Distance Channel Tensor

[0096] FIG. 7 is a schematic diagram of a distance channel tensor (700). The distance channel tensor (700) is a voxelized representation of the amino acid-specific distance channels (600) from FIG. 6. In the distance channel tensor (700), 21 amino acid-specific distance channels (602 to 642) are connected by voxel, similar to the RGB channels of a color image. The voxelized dimension of the distance channel tensor (700) is 21x3x3x3 (where 21 represents 21 amino acid categories and 3x3x3 represents a 3D voxel grid with 27 voxels); FIG. 7 is a 2D depiction of the dimension of 21x3x3.

[0097] One-hot encodings

[0098] FIG. 8 illustrates one-hot encodings (800) of a reference amino acid (204) and an alternative amino acid (214). In FIG. 8, the left column is the one-hot encoding (802) of the reference amino acid glycine (G) (204), where 1 is for the glycine amino acid category and 0s are for all other amino acid categories. In FIG. 8, the right column is the one-hot encoding (804) of the variant / alternative amino acid alanine (A) (214), where 1 is for the alanine amino acid category and 0s are for all other amino acid categories.

[0099] FIG. 9 is a schematic diagram of a voxelized one-hot encoded reference amino acid (902) and a voxelized one-hot encoded variant / alternative amino acid (912). The voxelized one-hot encoded reference amino acid (902) is a voxelized representation of the one-hot encoding (802) of the reference amino acid glycine (G) (204) from FIG. 8. The voxelized one-hot encoded alternative amino acid (912) is a voxelized representation of the one-hot encoding (804) of the variant / alternative amino acid alanine (A) (214) from FIG. 8. The voxelized dimension of the voxelized one-hot encoded reference amino acid (902) is 21x1x1x1 (where 21 indicates 21 amino acid categories); FIG. 9 is a 2D depiction of the dimension 21x1x1. Similarly, the voxelized dimension of an alternative one-hot encoded amino acid (912) is 21x1x1x1 (where 21 represents 21 amino acid categories); FIG. 9 is a 2D depiction of the dimension 21x1x1.

[0100] Reference allele tensor

[0101] FIG. 10 schematically illustrates a linking process (1000) that links the distance channel tensor (700) and the reference allele tensor (1004) of FIG. 7 on a voxel-by-voxel basis. The reference allele tensor (1004) is a voxel-by-voxel aggregation (repetition / cloning / replication) of the voxelized one-hot encoded reference amino acid (902) from FIG. 9. That is, multiple copies of the voxelized one-hot encoded reference amino acid (902) are linked to each other on a voxel-by-voxel basis according to the spatial arrangement of the voxels (514) within the voxel grid (522), and thus, the reference allele tensor (1004) has a corresponding copy of the voxelized one-hot encoded reference amino acid (910) for each of the voxels (514) within the voxel grid (522).

[0102] The linking process (1000) generates a linked tensor (1010). The voxelized dimension of the reference allele tensor (1004) is 21x3x3x3 (where 21 represents 21 amino acid categories and 3x3x3 represents a 3D voxel grid with 27 voxels); FIG. 10 is a 2D depiction of the reference allele tensor (1004) having a dimension of 21x3x3. The voxelized dimension of the linked tensor (1010) is 42x3x3x3; FIG. 10 is a 2D depiction of the linked tensor (1010) having a dimension of 42x3x3.

[0103] Alternative allele tensor

[0104] FIG. 11 schematically illustrates a linking process (1100) that links the distance channel tensor (700) of FIG. 7, the reference allele tensor (1004) of FIG. 10, and the alternative allele tensor (1104) on a voxel-by-voxel basis. The alternative allele tensor (1104) is a voxel-by-voxel aggregation (repetition / cloning / replication) of the voxelized one-hot encoded alternative amino acid (912) from FIG. 9. That is, multiple copies of the voxelized one-hot encoded alternative amino acid (912) are linked to each other on a voxel-by-voxel basis according to the spatial arrangement of the voxels (514) within the voxel grid (522), and thus, the alternative allele tensor (1104) has a corresponding copy of the voxelized one-hot encoded alternative amino acid (910) for each of the voxels (514) within the voxel grid (522).

[0105] The linking process (1100) generates a linked tensor (1110). The voxelized dimension of the alternative allele tensor (1104) is 21x3x3x3 (where 21 represents 21 amino acid categories and 3x3x3 represents a 3D voxel grid with 27 voxels); FIG. 11 is a 2D depiction of the alternative allele tensor (1104) having a dimension of 21x3x3. The voxelized dimension of the linked tensor (1110) is 63x3x3x3; FIG. 11 is a 2D depiction of the linked tensor (1110) having a dimension of 63x3x3.

[0106] In some embodiments, runtime logic (184) processes a tensor (1110) linked through a pathogenicity classifier to determine the pathogenicity of a variant / alternative amino acid alanine (A) (214), which is ultimately inferred as the pathogenicity determination of the base nucleotide variant that produces the variant / alternative amino acid alanine (A) (214).

[0107] Evolutionary conservation channels

[0108] Predicting the functional outcomes of variants relies, at least in part, on the assumption that important amino acids for a protein family are conserved through evolution due to negative selection (i.e., amino acid changes at these sites were previously harmful) and that mutations at these sites increase the likelihood of pathogenicity (causing disease) in humans. Generally, homologous sequences of target proteins are collected and aligned, and a metric of conservation is calculated based on the weighted frequencies of different amino acids observed at target sites within the alignment.

[0109] Accordingly, the disclosed technique links the distance channel tensor (700), the reference allele tensor (1004), and the alternative allele tensor (1104) to evolutionary channels. One example of evolutionary channels is pan-amino acid conserved frequencies. Another example of evolutionary channels is per-amino acid conserved frequencies.

[0110] In some embodiments, evolutionary channels are constructed using position weight matrices (PWMs). In other embodiments, evolutionary channels are constructed using position specific frequency matrices (PSFMs). In yet other embodiments, evolutionary channels are constructed using computational tools such as SIFT, PolyPhen, and PANTHER-PSEC. In still other embodiments, evolutionary channels are conservation channels based on evolutionary preservation. Conservation is related to preservation, as it also reflects the effect of negative selection that acted to prevent evolutionary change at a given site within the protein.

[0111] Pan-amino acid evolution profiles

[0112] FIG. 12 is a flowchart illustrating a process (1200) of a system for determining the pan-amino acid conservation frequencies of the nearest atoms and assigning them to voxels (voxelization) according to one embodiment of the disclosed technology. FIG. 12, FIG. 13, FIG. 14, FIG. 15, FIG. 16, FIG. 17, and FIG. 18 are discussed simultaneously.

[0113] In step (1202), a similar sequence finder (1204) of the system extracts amino acid sequences that are similar (homologous) to the reference amino acid sequence (202). Similar amino acid sequences can be selected from a number of species, such as primates, mammals, and vertebrates.

[0114] In step (1212), the aligner (1214) of the system aligns the reference amino acid sequence (202) with similar amino acid sequences by position, that is, the aligner (1214) performs multiple sequence alignment. FIG. 14 illustrates an exemplary multiple sequence alignment (1400) of the reference amino acid sequence (202) across 99 species. In some embodiments, the multiple sequence alignment (1400) may be divided to generate, for example, a first position frequency matrix (1402) for primates, a second position frequency matrix (1412) for mammals, and a third position frequency matrix (1422) for primates. In other embodiments, a single position frequency matrix is ​​generated across 99 species.

[0115] In step (1222), the system's pan-amino acid conservation frequency calculator (1224) uses multiple sequence alignment to determine the pan-amino acid conservation frequencies of reference amino acids within the reference amino acid sequence (202).

[0116] In step (1232), the nearest atom finder (1234) of the system finds the nearest atoms for the voxels (514) within the voxel grid (522). In some embodiments, the search for the nearest atoms per voxel may not be limited to any specific amino acid category or atom type. That is, the nearest atoms per voxel may be selected across amino acid categories and amino acid types, insofar as they are the nearest atoms to their respective voxel centers. In other embodiments, the search for the nearest atoms per voxel may be limited to only specific atom categories, such as only specific atomic elements like oxygen, nitrogen, and hydrogen, or only alpha-carbon atoms, or only beta-carbon atoms, or only side chain atoms, or only backbone atoms.

[0117] In step (1242), the amino acid selector (1244) of the system selects such a reference amino acid within the reference amino acid sequence (202) containing the nearest atoms identified in step (1232). Such a reference amino acid may be referred to as the nearest reference amino acid. FIG. 13 illustrates an example of locating the nearest atoms (1302) for the voxels (514) within the voxel grid (522) and mapping the nearest reference amino acid (1312) containing the nearest atoms (1302) for the voxels (514) within the voxel grid (522), respectively. This is identified in FIG. 13 as the "voxel-nearest amino acid mapping (1300)".

[0118] In step (1252), the voxelizer (1254) of the system voxels the pan-amino acid conservation frequencies of the nearest reference amino acid. FIG. 15 illustrates an example of determining a pan-amino acid conservation frequency sequence for a first voxel (1, 1) in a voxel grid (522), also referred to herein as “determining evolutionary profile per voxel (1500)”.

[0119] Referring to FIG. 13, the nearest reference amino acid mapped to the first voxel (1, 1) is the aspartic acid (D) amino acid at position 15 in the reference amino acid sequence (202). Subsequently, a multiple sequence alignment of the reference amino acid sequence (202) with, for example, 99 homologous amino acid sequences of 99 species is analyzed at position 15. Such position-specific and interspecies analysis indicates how many instances of amino acids are found from each of the 21 amino acid categories at position 15 across 100 aligned amino acid sequences (i.e., reference amino acid sequence (202) + 99 homologous amino acid sequences).

[0120] In the example illustrated in FIG. 15, the aspartate (D) amino acid is found at position 15, which is 96 of the 100 aligned amino acid sequences. Thus, the aspartate amino acid category (1504) is assigned a pan-amino acid conservation frequency of 0.96. Similarly, in the illustrated example, the valine (V) amino acid is found at position 15, which is 4 of the 100 aligned amino acid sequences. Thus, the valine amino acid category (1514) is assigned a pan-amino acid conservation frequency of 0.04. Since no instances of amino acids from other amino acid categories are detected at position 15, the remaining amino acid categories are assigned a pan-amino acid conservation frequency of 0. In this way, each of the 21 amino acid categories is assigned its own pan-amino acid conservation frequency that can be encoded in the pan-amino acid conservation frequency sequence (1502) for the first voxel (1, 1).

[0121] FIG. 16 illustrates the respective pan-amino acid conservation frequencies (1612 to 1692) determined for each of the voxels (514) within the voxel grid (522) using the positional frequency logic described in FIG. 15, also referred to herein as "voxel-evolution profile mapping (1600)".

[0122] Subsequently, the evolutionary profiles per voxel (1602) are used by the voxelizer (1254) to generate the evolutionary profiles per voxel (1700) shown in FIG. 17. Often, each of the voxels (514) within the voxel grid (522) has a different pan-amino acid conservation frequency sequence and, consequently, a different evolutionary profile per voxel, because the voxels are regularly mapped to different nearest atoms and, consequently, to different nearest reference amino acids. Of course, when two or more voxels have the same nearest atom and, by this, the same pan-amino acid conservation frequency sequence and the same evolutionary profile per voxel are assigned to each of the two or more voxels.

[0123] FIG. 18 illustrates an example of an evolutionary profile tensor (1800) in which voxelized voxel-specific evolutionary profiles (1700) are voxel-specifically connected to each other according to the spatial arrangement of voxels (514) within a voxel grid (522). The voxelized dimension of the evolutionary profile tensor (1800) is 21x3x3x3 (where 21 represents 21 amino acid categories and 3x3x3 represents a 3D voxel grid having 27 voxels); FIG. 18 is a 2D depiction of the evolutionary profile tensor (1800) having a dimension of 21x3x3.

[0124] In step (1262), the connector (174) connects the evolution profiles tensor (1800) with the distance channel tensor (700) by voxel. In some embodiments, the evolution profiles tensor (1800) is connected with the connector tensor (1110) by voxel to create an additional connected tensor (not shown) of dimension 84x3x3x3.

[0125] In step (1272), runtime logic (184) determines the pathogenicity of the target variant by processing an additional linked tensor of dimension 84x3x3x3 through a pathogenicity classifier, which is ultimately inferred as the pathogenicity determination of the base nucleotide variant that generates the target variant at the amino acid level.

[0126] Amino acid sugar evolution profiles

[0127] FIG. 19 is a flowchart illustrating the process (1900) of a system for determining the conservation frequencies per amino acid of the nearest atoms and assigning them to voxels (voxelization). In FIG. 19, steps (1202, 1212) are the same as in FIG. 12.

[0128] In step (1922), the system’s conserved frequency calculator (1924) uses multiple sequence alignment to determine the conserved frequencies per amino acid of the reference amino acid within the reference amino acid sequence (202).

[0129] In step (1932), the nearest atom finder (1934) of the system finds 21 nearest atoms across each of the 21 amino acid categories for each of the voxels (514) within the voxel grid (522). Each of the 21 nearest atoms is different from one another because they are selected from different amino acid categories. This leads to the selection of 21 unique nearest reference amino acids for a specific voxel, which in turn leads to the generation of 21 unique positional frequency matrices for a specific voxel, and ultimately leads to the determination of 21 unique conservation frequencies per amino acid for a specific voxel.

[0130] In step (1942), the amino acid selector (1944) of the system selects, for each of the voxels (514) in the voxel grid (522), 21 reference amino acids in the reference amino acid sequence (202) containing the 21 closest atoms identified in step (1932). Such reference amino acids may be called the closest reference amino acids.

[0131] In step (1952), the voxelizer (1954) of the system voxels the conserved amino acid per amino acid of the 21 closest reference amino acids identified for a specific voxel in step 1942). The 21 closest reference amino acids must be located at 21 different positions within the reference amino acid sequence (202) because they correspond to different basic closest atoms. Thus, for a specific voxel, 21 position frequency matrices for the 21 closest reference amino acids can be generated. The 21 position frequency matrices can be generated across multiple species in which homologous amino acid sequences are aligned positionally with the reference amino acid sequence (202), as discussed above in relation to FIGS. 12 through 15.

[0132] Next, using 21 positional frequency matrices, 21 positional conservation scores can be calculated for the 21 closest reference amino acids identified for a specific voxel. These 21 positional conservation scores form pan-amino acid conservation frequencies for a specific voxel, similar to the pan-amino acid conservation frequency sequence (1502) in FIG. 12, except that the sequence (1502) has many zero (0) entries; whereas each element (feature) within the amino acid conservation frequency sequence has a certain value (e.g., a floating-point number) because the 21 closest reference amino acids across the 21 amino acid categories must have different positions that yield different positional frequency matrices and different amino acid conservation frequencies.

[0133] The above process is performed for each of the voxels (514) within the voxel grid (522), and the generated voxel-specific amino acid conservation frequencies are voxelized, tensorized, linked, and processed for pathogenicity determination similar to the pan-amino acid conservation frequencies discussed in relation to FIGS. 12 to 18.

[0134] Comment channels

[0135] FIG. 20 illustrates various examples of voxelized annotation channels (2000) connected to a distance channel tensor (700). In some embodiments, the voxelized annotation channels are one-hot markers for different protein annotations, for example, whether an amino acid (residue) is part of a transmembrane region, signal peptide, active site, or any other binding site, or whether the residue is subject to posttranslational modifications, PathRatio (see Pei P, Zhang A: A Topological Measurement for Weighted Protein Interaction Network. CSB 2005, 268-278). Additional examples of annotation channels may be found in the specific embodiments section below and in the claims.

[0136] The voxelized annotation channels may be arranged by voxel so that the voxels may have the same annotation sequence as the voxelized reference allele and alternative allele sequences (e.g., annotation channels (2002, 2004, 2006)), or the voxels may have their own annotation sequences as the voxelized evolutionary profiles (1700) (e.g., annotation channels (2012, 2014, 2016) (as shown in different colors)).

[0137] The annotation channels are voxelized, tensorized, linked, and processed for pathogenicity determination similar to the pan-amino acid conservation frequencies discussed in relation to Figures 12 to 18.

[0138] Structural reliability channels

[0139] The disclosed technique can also associate various voxelized structural confidence channels with a distance channel tensor (700). Some examples of structural confidence channels include: GMQE scores (provided by SwissModel); B-factors; a column of temperature factors of homology models (indicating how well residues in a protein structure satisfy (physical) constraints); a normalized number aligning the template protein to the residue closest to the center of the voxel (alignments provided by HHpred, e.g., the voxel is closest to residues in three template structures, indicating that the aligned feature has a value of 3 / 6 = 0.5 among six template structures); minimum, maximum, and average TM scores; and predicted TM scores of template protein structures aligned with the residue closest to the voxel (continuing the above example, assuming three template structures have TM scores of 0.5, 0.5, and 1.5, the minimum is 0.5, the mean is 2 / 3, and the maximum is 1.5). TM scores can be provided by HHpred for each protein template. Additional examples of structure confidence channels can be found in the specific embodiments section below and in the claims.

[0140] The voxelized structural confidence channels may be arranged by voxel so that the voxels may have the same structural confidence sequences, such as voxelized reference allele and alternative allele sequences, or the voxels may have their own structural confidence sequences, such as voxelized evolutionary profiles (1700).

[0141] Structural reliability channels are voxelized, tensorized, linked, and processed for pathogenicity determination similar to the pan-amino acid conservation frequencies discussed in relation to Figures 12 to 18.

[0142] Pathogenicity classifier

[0143] FIG. 21 illustrates different combinations and permutations of input channels that may be provided as inputs (2102) to a pathogenicity classifier (2108) for determining the pathogenicity (2106) of a target variant. One of the inputs (2102) may be distance channels (2104) generated by a distance channel generator (2272). FIG. 22 illustrates different methods for calculating distance channels (2104). In one embodiment, distance channels (2104) are generated based on distances (2202) between voxel centers and atoms across multiple atomic elements regardless of amino acid. In some embodiments, the distances (2202) are normalized by a maximum scan radius to generate normalized distances (2202a). In another embodiment, distance channels (2104) are generated based on distances (2212) between voxel centers and alpha-carbon atoms in amino acid units. In some embodiments, the distances (2212) are normalized by the maximum scan radius to generate normalized distances (2212a). In another embodiment, the distance channels (2104) are generated based on the distances (2222) between the voxel centers and the beta-carbon atoms in amino acid units. In some embodiments, the distances (2222) are normalized by the maximum scan radius to generate normalized distances (2222a). In another embodiment, the distance channels (2104) are generated based on the distances (2232) between the voxel centers and the side chain atoms in amino acid units. In some embodiments, the distances (2232) are normalized by the maximum scan radius to generate normalized distances (2232a). In another embodiment, the distance channels (2104) are generated based on the distances (2242) between the voxel centers and the backbone atoms in amino acid units. In some embodiments, the distances (2242) are normalized by the maximum scan radius to produce normalized distances (2242a).In another embodiment, distance channels (2104) are generated based on distances (2252) (one feature) between voxel centers and their respective nearest atoms, regardless of atomic type and amino acid type. In another embodiment, distance channels (2104) are generated based on distances (2262) (one feature) between voxel centers and atoms from non-standard amino acids. In some embodiments, distances between voxels and atoms are calculated based on polar coordinates of voxels and atoms. Polar coordinates are parameterized by angles between voxels and atoms. In one embodiment, this angle information is used to generate angle channels for voxels (i.e., independent of distance channels). In some embodiments, angles between the nearest atom and neighboring atoms (e.g., backbone atoms) can be used as features encoded in voxels.

[0144] Another of the inputs (2102) may be a feature (2114) indicating missing atoms within a specified radius.

[0145] Another of the inputs (2102) may be a one-hot encoding (2124) of a reference amino acid. Another of the inputs (2102) may be a one-hot encoding (2134) of a variant / alternative amino acid.

[0146] Another of the inputs (2102) may be evolutionary channels (2144) generated by the evolutionary profile generator (2372) illustrated in FIG. 23. In one embodiment, the evolutionary channels (2144) may be generated based on pan-amino acid conservation frequencies (2302). In another embodiment, the evolutionary channels (2144) may be generated based on pan-amino acid conservation frequencies (2312).

[0147] Another of the inputs (2102) may be a feature (2154) indicating a missing residue or a missing evolution profile.

[0148] Another of the inputs (2102) may be annotation channels (2164) generated by an annotation generator (2472), as illustrated in FIG. 24. In one embodiment, the annotation channels (2164) may be generated based on molecular processing annotations (2402). In another embodiment, the annotation channels (2164) may be generated based on region annotations (2412). In yet another embodiment, the annotation channels (2164) may be generated based on site annotations (2422). In yet another embodiment, the annotation channels (2164) may be generated based on amino acid modification annotations (2432). In yet another embodiment, the annotation channels (2164) may be generated based on secondary structure annotations (2442). In yet another embodiment, the annotation channels (2164) may be generated based on laboratory information annotations (2452).

[0149] Another of the inputs (2102) may be structural reliability channels (2174) generated by a structural reliability generator (2572) as illustrated in FIG. 25. In one embodiment, structural reliability (2174) may be generated based on global model quality estimations (GMQEs) (2502). In another embodiment, structural reliability (2174) may be generated based on qualitative model energy analysis (QMEAN) scores (2512). In yet another embodiment, structural reliability (2174) may be generated based on temperature factors (2522). In yet another embodiment, structural reliability (2174) may be generated based on mold modeling scores (2542). Examples of mold modeling scores (2542) include minimum mold modeling scores (2542a), average mold modeling scores (2542b), and maximum mold modeling scores (2542c).

[0150] Those skilled in the art will understand that any permutation and combination of input channels may be connected as inputs to be processed through a pathogenicity classifier (2108) for determining the pathogenicity (2106) of a target variant. In some embodiments, only a subset of input channels may be connected. Input channels may be connected in any order. In one embodiment, input channels may be connected into a single tensor by a tensor generator (input encoder) (2110). This single tensor may then be provided as an input to a pathogenicity classifier (2108) for determining the pathogenicity (2106) of a target variant.

[0151] In one embodiment, the pathogenicity classifier (2108) uses convolutional neural networks (CNNs) having multiple convolutional layers. In another embodiment, the pathogenicity classifier (2108) uses recurrent neural networks (RNNs), such as long short-term memory networks (LSTMs), bi-directional LSTMs (Bi-LSTMs), and gated recurrent units (GRUs). In yet another embodiment, the pathogenicity classifier (2108) uses both CNNs and RNNs. In yet another embodiment, the pathogenicity classifier (2108) uses graph convolutional neural networks that model dependencies of graph-structured data. In yet another embodiment, the pathogenicity classifier (2108) uses variational autoencoders (VAEs). In another embodiment, the pathogenicity classifier (2108) uses generative adversarial networks (GANs). In another embodiment, the pathogenicity classifier (2108) may also be a self-attention-based language model, such as that implemented by transducers and BERTs, for example.

[0152] In other embodiments, the pathogenicity classifier (2108) may use 1D convolutions, 2D convolutions, 3D convolutions, 4D convolutions, 5D convolutions, extended or atrous convolutions, transposition convolutions, depth-separable convolutions, point-separable convolutions, 1x1 convolutions, group convolutions, flattened convolutions, spatial and cross-channel convolutions, shuffled group convolutions, spatially separable convolutions, and deconvolutions. It can use one or more loss functions, such as logistic regression / log loss, multi-class cross-entropy / softmax loss, binary cross-entropy loss, mean-squared error loss, L1 loss, L2 loss, smooth L1 loss, and Huber loss. It can use any parallelism, efficiency, and compression schemes, such as TFRecords, compressed encoding (e.g., PNG), sharding, parallel calls for map transformation, batching, prefetching, model parallelism, data parallelism, and synchronous / asynchronous stochastic gradient descent (SGD). It may include upsampling layers, downsampling layers, circular connections, gates and gated memory units (e.g., LSTM or GRU), residual blocks, residual connections, highway connections, skip connections, peephole connections, activation functions (e.g., non-linear transformation functions such as Rectifying Linear Unit (ReLU), Leaky ReLU, Exponential Linear Unit (ELU), Sigmoid and Hyperbolic Tangent (tanh)), batch normalization layers, regularization layers, dropout, pooling layers (e.g., Max or Average Pooling), Global Average Pooling layers, decay mechanisms, and Gaussian Error Linear Units.

[0153] The pathogenicity classifier (2108) is trained using backpropagation-based gradient update techniques. Exemplary gradient descent techniques that can be used to train the pathogenicity classifier (2108) include stochastic gradient descent, batch gradient descent, and mini-batch gradient descent. Some examples of gradient descent optimization algorithms that can be used to train the pathogenicity classifier (2108) are Momentum, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, Adam, AdaMax, Nadam, and AMSGrad. In other embodiments, the pathogenicity classifier (2108) may be trained by unsupervised learning, semi-supervised learning, self-learning, reinforcement learning, multitask learning, multimodal learning, transfer learning, knowledge distillation, etc.

[0154] FIG. 26 illustrates an exemplary processing architecture (2600) of a pathogenicity classifier (2108) according to one embodiment of the disclosed technology. The processing architecture (2600) comprises a cascade of processing modules (2606, 2610, 2614, 2618, 2622, 2626, 2630, 2634, 2638, 2642), each of which may include 1D convolutions (1x1x1 CONV), 3D convolutions (3x3x3 CONV), ReLU nonlinearity, and batch normalization (BN). Other examples of processing modules include full connection (FC) layers, dropout layers, flattening layers, and a final softmax layer that generates exponentially normalized scores for target variants belonging to the positive and pathogenic classes. In FIG. 26, "64" indicates the number of convolution filters applied by a specific processing module. In FIG. 26, the size of the input voxel (2602) is 15x15x15x8. FIG. 26 also illustrates the respective volumetric dimensions of the intermediate inputs (2604, 2608, 2612, 2616, 2620, 2624, 2628, 2632, 2636, 2640) generated by the processing architecture (2600).

[0155] FIG. 27 illustrates an exemplary processing architecture (2700) of a pathogenicity classifier (2108) according to one embodiment of the disclosed technology. The processing architecture (2700) includes a cascade of processing modules (2708, 2714, 2720, 2726, 2732, 2738, 2744, 2750, 2756, 2762, 2768, 2774, 2780), such as 1D convolutions (CONV 1D), 3D convolutions (CONV 3D), ReLU nonlinearity, and batch normalization (BN). Other examples of processing modules include dense layers, dropout layers, flattening layers, and a final softmax layer that generates exponentially normalized scores for target variants belonging to the positive and pathogenic classes. In FIG. 27, "64" and "32" indicate the number of convolution filters applied by a specific processing module. In FIG. 27, the size of the input voxel (2704) supplied by the input layer (2702) is 7x7x7x108. FIG. 27 also illustrates the respective volumetric dimensions of intermediate inputs (2710, 2716, 2722, 2728, 2734, 2740, 2746, 2752, 2758, 2764, 2770, 2776, 2782) generated by the processing architecture (2700) and intermediate outputs (2706, 2712, 2718, 2724, 2730, 2736, 2742, 2748, 2754, 2760, 2766, 2772, 2778, 2784).

[0156] Those skilled in the art will understand that other current and future artificial intelligence, machine learning, and deep learning models, datasets, and training techniques can be integrated into the disclosed variant pathogenicity classifier without departing from the spirit of the disclosed technology.

[0157] Performance results as objective indicators of originality and non-obviousness

[0158] The variant pathogenicity classifier disclosed herein performs pathogenicity predictions based on 3D protein structures and is referred to as "PrimateAI 3D". "Primate AI" is a jointly owned and previously disclosed variant pathogenicity classifier that generates pathogenicity prediction-based protein sequences. Further details regarding PrimateAI can be found in jointly owned U.S. Patent Applications No. 16 / 160,903; No. 16 / 160,986; No. 16 / 160,968; and No. 16 / 407,149, and in the literature [Sundaram, L. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018)].

[0159] Figures 28, 29, 30, and 31 use PrimateAI as a benchmark model to demonstrate the classification superiority of PrimateAI 3D over PrimateAI. The performance results in Figures 28, 29, 30, and 31 are generated from a classification task that accurately distinguishes benign variants from pathogenic variants across multiple validation sets. PrimateAI 3D is trained on training sets that are different from the multiple validation sets. PrimateAI 3D is trained on common human variants and variants from primates, which are used as benign datasets, but simulates variants based on trinucleotide contexts, which are used as unlabeled or pseudo-pathogenic datasets.

[0160] New Developmental Delay Disorder (new DDD) is one example of a validation set used to compare the classification accuracy of Primate AI 3D against Primate AI. The new DDD validation set labels variants from individuals with DDD as pathogenic and labels the same variants from healthy kin of individuals with DDD as benign. A similar labeling scheme is used with the autism spectrum disorder (ASD) validation set shown in Fig. 31.

[0161] BRCA1 is another example of a validation set used to compare the classification accuracy of Primate AI 3D against Primate AI. The BRCA1 validation set labels synthetically generated reference amino acid sequences that simulate the protein of the BRCA1 gene as positive variants, and synthetically modified allele amino acid sequences that simulate the protein of the BRCA1 gene as pathogenic variants. A similar labeling scheme is used with different validation sets for the TP53 gene, the TP53S3 gene and its variants, and other genes and their variants, as illustrated in Fig. 31.

[0162] Figure 28 identifies the performance of the benchmark PrimateAI model with horizontal bars (labeled "PAI") and the performance of the disclosed PrimateAI 3D model with horizontal bars (labeled "ens10_7x7x7x2_hhpred_evo+alt"). The horizontal bars labeled "ens10_7x7x7x2_hhpred_evo+alt_paisum" depict pathogenicity predictions derived by combining the respective pathogenicity predictions of the disclosed PrimateAI 3D model and the benchmark PrimateAI model. In the legend, "ens10" indicates an ensemble of 10 PrimateAI 3D models, each trained with a different seed training dataset and randomly initialized with different weights and biases. Additionally, "7x7x7x2" depicts the size of the voxel grid used to encode input channels during the training of the ensemble of 10 PrimateAI 3D models. For a given variant, an ensemble of 10 PrimateAI 3D models generates 10 pathogenicity predictions, which are subsequently combined (e.g., by averaging) to generate a final pathogenicity prediction for the given variant. This logic is similarly applied to ensembles of different group sizes.

[0163] Additionally, in Figure 28, the y-axis represents different validation sets, and the x-axis represents p-values. Larger p-values, i.e., longer horizontal bars, indicate greater accuracy in distinguishing benign variants from pathogenic variants. As demonstrated by the p-values ​​in Figure 28, PrimateAI 3D outperforms PrimateAI across most validation sets (the only exception being the tp53s3_A549 validation set). That is, the horizontal bars for PrimateAI 3D (labeled as "ens10_7x7x7x2_hhpred_evo+alt") are consistently longer than the horizontal bars for PrimateAI (labeled as "PAI").

[0164] In addition, in Fig. 28, the "mean" category along the y-axis calculates the average of the p-values ​​determined for each of the validation sets. In the mean category as well, PrimateAI 3D outperforms PrimateAI.

[0165] In Fig. 29, PrimateAI is represented by horizontal bars (labeled "PAI"), an ensemble of 20 PrimateAI 3D models trained on a 3x3x3 voxel grid is represented by horizontal bars labeled "ns20_3x3x3x2_evo+alt", an ensemble of 10 PrimateAI 3D models trained on a 7x7x7 voxel grid is represented by horizontal bars labeled "ens10_7x7x7x2_evo+alt", an ensemble of 20 PrimateAI 3D models trained on a 7x7x7 voxel grid is represented by horizontal bars labeled "ens20_7x7x7x2_evo+alt", and an ensemble of 20 PrimateAI 3D models trained on a 17x17x17 voxel grid is It is represented by horizontal bars labeled as "ens20_17xl 7x1 7x2_evo+alt".

[0166] Additionally, in Figure 29, the y-axis represents different validation sets, and the x-axis represents p-values. As before, larger p-values, i.e., longer horizontal bars, indicate greater accuracy in distinguishing benign variants from pathogenic variants. As demonstrated by the p-values ​​in Figure 20, different configurations of PrimateAI 3D outperform PrimateAI across most validation sets. That is, the horizontal bars representing ensembles of multiple PrimateAI 3D models (labeled as "ns20_3x3x3x2_evo+alt", "ens10_7x7x7x2_evo+alt", "ens20_7x7x7x2_evo+alt", and "ens20_17x1 7x1 7x2_evo+alt") are generally longer than the horizontal bars for PrimateAI (labeled as "PAI").

[0167] In addition, in Fig. 29, the "mean" category along the y-axis calculates the average of the p-values ​​determined for each of the validation sets. In the mean category as well, different configurations of PrimateAI 3D outperform PrimateAI.

[0168] In Fig. 30, the vertical bars with darker shades represent PrimateAI ("PrimateAI (v1)"), and the vertical bars with lighter shades represent PrimateAI 3D ("PrimateAI 3D"). In Fig. 30, the y-axis represents p-values, and the x-axis represents different validation sets. In Fig. 30, without exception, PrimateAI 3D consistently outperforms PrimateAI across all validation sets. That is, the vertical bars for PrimateAI 3D are always longer than the vertical bars for PrimateAI.

[0169] Figure 31 identifies the performance of the benchmark PrimateAI model with vertical bars labeled "PAI-v1_plain" and the performance of the disclosed PrimateAI 3D model with vertical bars labeled "PAI-3D-orig_plain". The vertical bars labeled "PAI-3D-orig_paisum" depict pathogenicity predictions derived by combining the respective pathogenicity predictions of the disclosed PrimateAI 3D model and the benchmark PrimateAI model. In Figure 31, the y-axis represents p-values, and the x-axis represents different validation sets.

[0170] As evidenced by the p-values ​​in Figure 31, PrimateAI 3D outperforms PrimateAI across most validation sets (the only exception being the tp53s3_A549_p53NULL_Nutlin-3 validation set). That is, the vertical bars for PrimateAI 3D are consistently longer than the vertical bars for PrimateAI.

[0171] In addition, in Fig. 31, a separate "average" chart calculates the average of the p-values ​​determined for each of the validation sets. In the average chart as well, PrimateAI 3D outperforms PrimateAI.

[0172] Average statistics can be biased by outliers. To address this, a separate "method rank" chart is also shown in Fig. 31. Higher ranks indicate poorer classification accuracy. In the method rank chart as well, PrimateAI 3D outperforms PrimateAI, which has all counts of 3, by having more counts of lower ranks 1 and 2.

[0173] In FIGS. 28 to 31, it is also evident that combining PrimateAI 3D with PrimateAI produces excellent classification accuracy. That is, a protein can be supplied to PrimateAI as an amino acid sequence to produce a first output, and the same protein can be supplied to PrimateAI 3D as a 3D, voxelized protein structure to produce a second output, and the first and second outputs can be aggregated, combined, or analyzed to produce a final pathogenicity prediction for the variants experienced by the protein.

[0174] Efficient voxelization

[0175] FIG. 32 is a flowchart illustrating an efficient voxelization process (3200) that efficiently identifies the nearest atoms in voxel units.

[0176] Now, distance channels are discussed again. As discussed above, the reference amino acid sequence (202) may contain different types of atoms, such as alpha-carbon atoms, beta-carbon atoms, oxygen atoms, nitrogen atoms, hydrogen atoms, etc. Thus, as discussed above, distance channels may be arranged by the nearest alpha-carbon atoms, nearest beta-carbon atoms, nearest oxygen atoms, nearest nitrogen atoms, nearest hydrogen atoms, etc. For example, in FIG. 6, each of the nine voxels (514) has 21 amino acid-specific distance channels for the nearest alpha-carbon atoms. FIG. 6 can be further extended so that each of the nine voxels (514) also has 21 amino acid-specific distance channels for the nearest beta-carbon atoms, and each of the nine voxels (514) also has the nearest general atomic distance channel for the nearest atom regardless of the type of atom and the type of amino acid. In this way, each of the 9 voxels (514) can have 43 distance channels.

[0177] Now, the number of distance calculations required to identify the nearest atoms in voxel units for inclusion in distance channels is discussed. Consider the example of Fig. 3, which illustrates a total of 828 alpha-carbon atoms distributed across 21 amino acid categories. In Fig. 6, to calculate the amino acid-specific distance channels (602 to 642), that is, to determine 189 distance values, the distances from each of the 9 voxels (514) to each of the 828 alpha-carbon atoms are measured, resulting in 9 * 828 = 7,452 distance calculations. In the 3D case of 27 voxels, this results in 27 * 828 = 22,356 distance calculations. When 828 beta-carbon atoms are also included, this number increases to 27 * 1656 = 44,712 distance calculations.

[0178] This means that, as illustrated in Fig. 35a, the runtime complexity for identifying the nearest atoms at the voxel level for single protein voxelization is O(#atoms * #voxels). Additionally, the runtime complexity for single protein voxelization increases to O(#atoms * #voxels * #attributes) when distance channels are calculated across various attributes (e.g., different features or channels per voxel, such as annotation channels and structural confidence channels).

[0179] Consequently, distance calculations become the most computationally intensive part of the voxelization process, potentially consuming valuable computational resources from critical runtime tasks such as model training and model inference. For example, consider the case of training a model on a training dataset of 7,000 proteins. Generating distance channels for multiple voxels across multiple amino acids, atoms, and attributes entails more than 100 voxels per protein, which can result in approximately 800,000 voxels in a single training iteration (epoch). Running training for 20 to 40 epochs while rotating atomic coordinates in each epoch can result in up to 32 million voxels.

[0180] In addition to high computational costs, the data size for 32 million voxels is too large to fit into main memory (e.g., >20 TB for a 15x15x15 voxel grid). Considering repeated training runs for parameter optimization and ensemble learning, the memory footprint of the voxelization process becomes too large to be stored on disk, causing the voxelization process to become part of model training rather than a precomputation step.

[0181] The disclosed technology provides an efficient voxelization process that achieves up to approximately 100x speedup compared to a runtime complexity of O(#atoms * #voxels). The disclosed efficient voxelization process reduces the runtime complexity for single-protein voxelization to O(#atoms). For different features or channels per voxel, the disclosed efficient voxelization process reduces the runtime complexity for single-protein voxelization to O(#atoms * #attributes). As a result, the voxelization process becomes as fast as model training, shifting the computational bottleneck from voxelization back to computing neural network weights on processors such as GPUs, ASICs, TPUs, FPGAs, CGRAs, etc.

[0182] In some embodiments of the disclosed efficient voxelization process involving large voxel grids, the runtime complexity for single protein voxelization is O(#atoms + voxels) and O(#atoms * #attributes + voxels) for cases of different features or channels per voxel. The "+ voxels" complexity is observed when the number of atoms is infinitesimal compared to the number of voxels, for example, when there is one atom in a 100x100x100 voxel grid (i.e., 1 million voxels per atom). In such a scenario, the runtime is dominated by the overhead of a huge number of voxels, for example, to allocate memory for 1 million voxels, initialize 1 million voxels to zero, and so on.

[0183] Now, details of the disclosed efficient voxelization process are discussed. FIGS. 32a, FIGS. 32b, FIGS. 33, FIGS. 34, and FIGS. 35b are discussed simultaneously.

[0184] Starting from FIG. 32a, in step (3202), each atom (e.g., each of the 828 alpha-carbon atoms and each of the 828 beta-carbon atoms) is associated with an atom-containing voxel (e.g., one of the nine voxels (514)). The term “containing” refers to the 3D atomic coordinates of the atom being located within the voxel. The atom-containing voxel is also referred to herein as an “atom-containing voxel.”

[0185] FIGS. 32b and FIGS. 33 describe a method for selecting a voxel containing a specific atom. FIG. 33 uses 2D atomic coordinates as a representative of 3D atomic coordinates. Note that the voxel grid (522) is regularly spaced from each of the voxels (514) having the same step size (e.g., 1 angstrom (Å) or 2 Å).

[0186] Also, in FIG. 33, the voxel grid (522) has indices [0, 1, 2] along a first dimension (e.g., x-axis) and indices [0, 1, 2] along a second dimension (e.g., y-axis). Also, in FIG. 33, each voxel (514) within the voxel grid (522) is identified by voxel indices [voxel 0, voxel 1, ..., voxel 8] and by voxel center indices [(1, 1), (1, 2), ..., (3, 3)].

[0187] In addition, in FIG. 33, the center coordinates of the voxel centers along the first dimension, i.e., the first dimension voxel coordinates, are identified. In addition, in FIG. 33, the center coordinates of the voxel centers along the second dimension, i.e., the second dimension voxel coordinates, are identified.

[0188] First, in step (3202a) (step 1 of FIG. 33), the 3D atomic coordinates (1.7456, 2.14323) of a specific atom are quantized to produce quantized 3D atomic coordinates (1.7, 2.1). Quantization can be achieved by rounding or truncation of bits.

[0189] Next, in step (3202b) (step 2 of FIG. 33), the voxel coordinates (or voxel centers or voxel center coordinates) of the voxels (514) are assigned to 3D atomic coordinates quantized in dimensional units. For the first dimension, the quantized atomic coordinate 1.7 is assigned to voxel 1 because it covers first-dimensional voxel coordinates in the range of 1 to 2 and is centered at 1.5 in the first dimension. Note that voxel 1 has an index 1 along the first dimension, in contrast to having an index 0 along the second dimension.

[0190] In the case of the second dimension, starting from voxel 1, the voxel grid (522) is traversed along the second dimension. This results in the quantized atomic coordinate 2.5 being assigned to voxel 7, because it covers second-dimensional voxel coordinates in the range of 2 to 3 and is centered at 2.5 in the second dimension. Note that voxel 7 has index 2 along the second dimension, in contrast to having index 1 along the first dimension.

[0191] Next, in step (3202c) (step 3 of FIG. 33), dimension indices corresponding to the assigned voxel coordinates are selected. That is, for voxel 1, index 1 is selected along the first dimension, and for voxel 7, index 2 is selected along the second dimension. Those skilled in the art will understand that the above steps can be similarly executed to select dimension indices along the third dimension in the case of the third dimension.

[0192] Next, in step (3202d) (step 4 of FIG. 33), an accumulated sum is generated based on positional weighting of dimension indices selected by the squares of the radix. The general idea behind positional numbering systems is that numerical values ​​are expressed through increasing squares of the radix (or base), for example, binary is radix 2, ternary is radix 3, octal is radix 8, and hexadecimal is radix 16. This is often referred to as a weighted numbering system because each position is weighted by the square of the radix. The set of valid digits for a positional numbering system is equal in size to the radix of the system. For example, in a decimal system there are 10 digits from 0 to 9, and in a ternary system there are 3 digits from 0, 1, and 2. In a radix system, the largest significant digit is 1 less than the radix (therefore, in any radix system, 8 is not a smaller significant digit than 9). Any decimal integer can be accurately represented in any other integral-based system, and vice versa.

[0193] Returning to the example in Fig. 33, the selected dimension indices 1 and 2 are converted into a single integer by multiplying them positionally by their respective squares of base 3 and summing the results of the positional multiplications. Here, base 3 is chosen because 3D atomic coordinates have three dimensions (however, Fig. 33 illustrates only 2D atomic coordinates with two dimensions for simplification).

[0194] Since index 2 is located at the rightmost bit (i.e., the least significant bit), it is multiplied by 3 to the power of 0 to yield 2. Since index 1 is located at the second rightmost bit (i.e., the second least significant bit), it is multiplied by 3 to the power of 1 to yield 3. This results in a cumulative sum of 5.

[0195] Next, in step (3202e) (step 5 of FIG. 33), a voxel index of a voxel containing a specific atom is selected based on the accumulated sum. That is, the accumulated sum is interpreted as a voxel index of a voxel containing a specific atom.

[0196] In step (3212), after each atom is associated with an atom-containing voxel, each atom is further associated with one or more voxels in the neighborhood of the atom-containing voxel, also referred to herein as “neighbor voxels.” Neighbor voxels may be selected based on being within a predefined radius (e.g., 5 angstroms (Å)) of the atom-containing voxel. In other embodiments, neighbor voxels may be selected based on being adjacent to the atom-containing voxel (e.g., top, bottom, right, left neighbor voxels). The generated associations associating each atom with the atom-containing voxel and neighbor voxels are encoded in an atom-voxel mapping (3402), also referred to herein as an element-cell mapping. In one example, a first alpha-carbon atom is associated with a first subset of voxels (3404) including the atom-containing voxel and neighbor voxels for the first alpha-carbon atom. In another example, the second alpha-carbon atom is associated with a second subset of voxels (3406) including an atom-containing voxel and neighboring voxels for the second alpha-carbon atom.

[0197] Note that no distance calculations are performed to determine the atomic-containing voxels and neighboring voxels. The atomic-containing voxels are selected by a spatial arrangement of voxels that allows the assignment of quantized 3D atomic coordinates to corresponding regularly spaced voxel centers within a voxel grid (without using any distance calculations). Additionally, neighboring voxels are selected (again without using any distance calculations) by being spatially adjacent to the atomic-containing voxels within the voxel grid.

[0198] In step (3222), each voxel is mapped to the atoms to which it was associated in steps (3202, 3212). In one embodiment, this mapping is encoded in a voxel-atom mapping (3412), which is generated based on the atomic-voxel mapping (3402) (e.g., by applying a voxel-based classification key on the atomic-voxel mapping (3402)). The voxel-atom mapping (3412) is also referred to herein as "cell-element mapping". In one example, the first voxel is mapped in steps (3202, 3212) to a first subset of alpha-carbon atoms (3414) including the alpha-carbon atoms associated with the first voxel. In another example, the second voxel is mapped to a second subset of alpha-carbon atoms (3416) including alpha-carbon atoms associated with the second voxel in steps (3202, 3212).

[0199] In step (3232), for each voxel, the distances between the atoms mapped to the voxel in step (3222) and the voxel are calculated. Step (3232) has a runtime complexity of O(#atoms) because the distance to a specific atom is measured only once from each voxel to which that specific atom is uniquely mapped in the voxel-atom mapping (3412). This is the case when no neighboring voxels are considered. If there are no neighbors, the implied constant factor in big-O notation is 1. If there are neighbors, the big-O notation is equal to the number of neighbors + 1, since the number of neighbors is constant for each voxel, and thus, the runtime complexity of O(#atoms) is maintained. In contrast, in FIG. 35a, the distances to a specific atom are measured redundantly as many times as there are voxels (e.g., 27 distances to a specific atom due to 27 voxels).

[0200] In FIG. 35b, based on the voxel-atom mapping (3412), each voxel is mapped to its respective subset of 828 atoms, as exemplified by the respective ellipses for each voxel (excluding distance calculations for neighboring voxels). Each subset generally does not overlap, with some exceptions. As shown in FIG. 35b by the yellow overlap between the prime symbol "'" and the ellipses, there is some minor overlap when multiple atoms are mapped to the same voxel. This minimal overlap has an additive effect on the runtime complexity of O(#atoms) and no multiplicative effect. This overlap is the result of considering neighboring voxels after determining the voxel containing the atom. If there are no neighboring voxels, there may be no overlap because the atom is associated with only one voxel. However, if neighbors are considered, each neighbor can potentially be associated with the same atom (unless there is another atom of the same amino acid closer).

[0201] In step (3242), for each voxel, the atom closest to the voxel is identified based on the distances calculated in step (3232). In one embodiment, this identification is encoded in a voxel-closest atom mapping (3422), also referred to herein as "cell-closest element mapping." In one example, the first voxel is mapped to the second alpha-carbon atom as its closest alpha-carbon atom (3424). In another example, the second voxel is mapped to the 31st alpha-carbon atom as its closest alpha-carbon atom (3426).

[0202] In addition, as voxel-specific distances are calculated using the technique discussed above, the atomic types and amino acid types of atoms are categorized, and the corresponding distance values ​​are stored to generate categorized distance channels.

[0203] Once the distances to the nearest atoms are identified using the technique discussed above, these distances can be encoded into distance channels for voxelization and subsequent processing by the pathogenicity classifier (2108).

[0204] computer system

[0205] FIG. 36 illustrates an exemplary computer system (3600) that may be used to implement the disclosed technology. The computer system (3600) includes at least one central processing unit (CPU) (3672) that communicates with a number of peripheral devices via a bus subsystem (3655). These peripheral devices may include, for example, a storage subsystem (3610) including memory devices and a file storage subsystem (3636), user interface input devices (3638), user interface output devices (3676), and a network interface subsystem (3674). The input and output devices allow user interaction with the computer system (3600). The network interface subsystem (3674) provides an interface to external networks that includes an interface to corresponding interface devices in other computer systems.

[0206] In one embodiment, the pathogenicity classifier (2108) is linked to the storage subsystem (3610) and user interface input devices (3638) so as to be communicable.

[0207] User interface input devices (3638) may include keyboards; pointing devices such as mice, trackballs, touchpads, or graphic tablets; scanners; touch screens integrated within a display; audio input devices such as voice recognition systems and microphones; and other types of input devices. Generally, the use of the term “input device” is intended to include all possible types of devices and methods for inputting information into the computer system (3600).

[0208] User interface output devices (3676) may include non-visual displays such as display subsystems, printers, fax machines, or audio output devices. The display subsystem may include flat devices such as LED displays, cathode ray tubes (CRTs), liquid crystal displays (LCDs), projection devices, or some other mechanism for generating visible images. The display subsystem may also provide non-visual displays such as audio output devices. Generally, the use of the term "output device" is intended to include all possible types of devices and methods for outputting information from the computer system (3600) to a user or to another machine or computer system.

[0209] The storage subsystem (3610) stores programming and data configurations that provide the functionality of some or all of the modules and methods described herein. These software modules are generally executed by processors (3678).

[0210] The processors (3678) may be graphics processing units (GPUs), field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), and / or coars-grain reconfigurable architectures (CGRAs). The processors (3678) may be hosted by deep learning cloud platforms such as Google Cloud Platform™, Xilinx™, and Cirrascale™. Examples of processors (3678) include Google’s Tensor Processing Unit (TPU)™, rackmount solutions such as the GX4 Rackmount Series™, GX36 Rackmount Series™, NVIDIA DGX-1™, Microsoft’s Stratix V FPGA™, Graphcore’s Intelligent Processor Unit (IPU)™, Qualcomm’s Zeroth Platform™ with Snapdragon processors™, NVIDIA’s Volta™, NVIDIA’s DRIVE PX™, NVIDIA’s JETSON TX1 / TX2 MODULE™, Intel’s Nirvana™, Movidius VPU™, Fujitsu DPI™, ARM’s DynamicIQ™, IBM TrueNorth™, Lambda GPU servers with Testa V100s™, etc.

[0211] The memory subsystem (3622) used in the storage subsystem (3610) may include a plurality of memories, including a main random access memory (RAM) (3632) for storing instructions and data during program execution and a read-only memory (ROM) (3634) for storing fixed instructions. The file storage subsystem (3636) may provide permanent storage for program and data files and may include a hard disk drive, a floppy disk drive, a CD-ROM drive, an optical drive, or a removable media cartridge with associated removable media. Modules implementing the functions of the specified embodiments may be stored in other machines accessible by the file storage subsystem (3636) within the storage subsystem (3610) or by the processor.

[0212] The bus subsystem (3655) provides a mechanism for various components and subsystems of the computer system (3600) to communicate with each other as intended. Although the bus subsystem (3655) is schematically illustrated as a single bus, alternative embodiments of the bus subsystem may use multiple buses.

[0213] The computer system (3600) itself may be of various types, including personal computers, portable computers, workstations, computer terminals, network computers, televisions, mainframes, server farms, widely distributed loosely networked sets of computers, or any other data processing systems or user devices. Due to the ever-changing nature of computers and networks, the description of the computer system (3600) depicted in FIG. 36 is intended only as a specific example to illustrate preferred embodiments of the invention. Many other configurations of the computer system (3600) having more or fewer components than the computer system depicted in FIG. 36 are possible.

[0214] Specific implementation examples 1

[0215] The following embodiments may be implemented as a system, method, or manufactured article. One or more features of the embodiments may be combined with the basic embodiment. Non-mutually exclusive embodiments are taught to be combinable. One or more features of the embodiments may be combined with other embodiments. The present invention periodically reminds the user of these options. The omission of citations repeating these options from some embodiments should not be construed as limiting the combinations taught in the foregoing sections—these citations are thereby incorporated by reference into each of the following embodiments.

[0216] Although the disclosed technology uses 3D data as input, in other embodiments, it may similarly use 1D data, 2D data (e.g., pixels and 2D atomic coordinates), 4D data, 5D data, etc.

[0217] In some embodiments, the system includes a memory for storing amino acid-specific distance channels for a plurality of amino acids in a protein. Each amino acid-specific distance channel has voxel-specific distance values ​​for voxels within a plurality of voxels. The voxel-specific distance values ​​specify the distance from corresponding voxels within a plurality of voxels to atoms of corresponding amino acids within a plurality of amino acids. The system further includes a pathogenicity determination engine configured to process a tensor comprising amino acid-specific distance channels and alternative alleles of the protein expressed by the variant. The pathogenicity determination engine may also be configured to determine the pathogenicity of the variant based at least partially on the tensor.

[0218] In some embodiments, the system further includes a distance channel generator that centers a voxel grid of voxels on the alpha-carbon atoms of each residue of an amino acid. The distance channel generator can center a voxel grid on the alpha-carbon atoms of a specific amino acid residue located at a variant amino acid in a protein.

[0219] The system may be configured to encode the directionality of an amino acid and the location of a specific amino acid by multiplying the voxel-wise distance values ​​for such amino acid preceding the specific amino acid in a tensor by a directionality parameter. The distances may be the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms of the corresponding amino acid. In some embodiments, the nearest atomic distances may be Euclidean distances. The nearest atomic distances may be normalized by dividing the Euclidean distances by the maximum nearest atomic distance. The amino acid may have alpha-carbon atoms, and in some embodiments, the distances may be the nearest alpha-carbon atom distances from corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding amino acid. The amino acid may have beta-carbon atoms, and in some embodiments, the distances may be the nearest beta-carbon atom distances from corresponding voxel centers to the nearest beta-carbon atoms of the corresponding amino acid. Amino acids may have backbone atoms, and in some embodiments, the distances may be the nearest backbone atom distances from corresponding voxel centers to the nearest backbone atoms of the corresponding amino acid. Amino acids may have side chain atoms, and in some embodiments, the distances may be the nearest side chain atom distances from corresponding voxel centers to the nearest side chain atoms of the corresponding amino acid.

[0220] The system may be further configured to encode a nearest atom channel in the tensor that specifies the distance from each voxel to the nearest atom. The nearest atom may be selected regardless of the amino acid and the atomic elements of the amino acid. In some embodiments, the distance is a Euclidean distance. The distance may be normalized by dividing the Euclidean distance by the maximum distance. The amino acids may include non-standard amino acids. The tensor may include an absentee atom channel that specifies atoms not found within a predefined radius of the voxel center, and the absentee atom channel may be one-hot encoded. In some embodiments, the tensor may further include a one-hot encoding of an alternative allele encoded voxel-wise in each amino acid-specific distance channel. The tensor may further include a reference allele of the protein. In some embodiments, the tensor may further include a one-hot encoding of the reference allele encoded voxel-wise in each amino acid-specific distance channel. The tensor may further include evolutionary profiles that specify the conservation levels of amino acids across multiple species.

[0221] The system may further include an evolutionary profile generator, which, for each voxel, selects the nearest atom across amino acid and atomic categories, selects a pan-amino acid conservation frequency sequence for the amino acid residue containing the nearest atom, and makes the pan-amino acid conservation frequency sequence available as one of the evolutionary profiles. The pan-amino acid conservation frequency sequence may be configured for specific positions of residues as observed in multiple species. The pan-amino acid conservation frequency sequence may determine whether there is a missing conservation frequency for a specific amino acid. In some embodiments, the evolutionary profile generator may, for each voxel, select the respective nearest atom within each of the amino acids, select the respective per-amino acid conservation frequencies for the respective residues of the amino acid containing the nearest atom, and make the per-amino acid conservation frequencies available as one of the evolutionary profiles. The per-amino acid conservation frequencies may be configured for specific positions of residues as observed in multiple species. The per-amino acid conservation frequencies may determine whether there is a missing conservation frequency for a specific amino acid.

[0222] In some embodiments of the system, the tensor may additionally include annotation channels for amino acids. The annotation channels may be one-hot encoded in the tensor. The annotation channels may be molecular processing annotations including initiator methionine, signal, modification peptide, propeptide, chain, and peptide. The annotation channels may be domain annotations including topology domains, transmembranes, intramembranes, domains, repeats, calcium binding, zinc fingers, deoxyribonucleic acid (DNA) binding, nucleotide binding, regions, coiled coils, motifs, and compositional biases. The annotation channels may be site annotations including active sites, metal binding, binding sites, and sites. The annotation channels may be amino acid modification annotations including non-standard residues, modification residues, lipidation, glycosylation, disulfide bonding, and crosslinking. The annotation channels may be secondary structure annotations including helices, spins, and beta strands. The annotation channels may be experimental information annotations including mutagenic, sequence uncertainty, sequence conflict, non-adjacent residues, and non-terminal residues.

[0223] In some embodiments of the system, the tensor additionally includes structure confidence channels for amino acids that specify the quality of each amino acid's respective structures. The structure confidence channels may be Global Model Quality Estimates (GMQEs). The structure confidence channels may include Qualitative Model Energy Analysis (QMEAN) scores.

[0224] Structural confidence channels may be temperature factors that specify the extent to which residues satisfy the physical constraints of their respective protein structures. Structural confidence channels may be template structure alignments that specify the extent to which residues of atoms closest to the voxels possess aligned template structures. Structural confidence channels may be template modeling scores of aligned template structures. Structural confidence channels may be the minimum template modeling scores, the average of the template modeling scores, and the maximum template modeling score among the template modeling scores.

[0225] In some embodiments, the system may further include a tensor generator that voxel-wise connects amino acid-wise distance channels for alpha-carbon atoms with the one-hot encoding of an alternative allele to generate a tensor. The tensor generator may voxel-wise connect amino acid-wise distance channels for beta-carbon atoms with the one-hot encoding of an alternative allele to generate a tensor. The tensor generator may voxel-wise connect amino acid-wise distance channels for alpha-carbon atoms, amino acid-wise distance channels for beta-carbon atoms, and the one-hot encoding of an alternative allele to generate a tensor. The tensor generator may voxel-wise connect amino acid-wise distance channels for alpha-carbon atoms, amino acid-wise distance channels for beta-carbon atoms, the one-hot encoding of an alternative allele, and pan-amino acid conserved frequencies to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, pan-amino acid conserved frequencies, and annotation channels voxel by voxel to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, pan-amino acid conserved frequencies, annotation channels, and structural confidence channels voxel by voxel to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, and conserved frequencies per amino acid for each amino acid voxel by voxel to generate a tensor.A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, conserved frequencies per amino acid for each amino acid, and annotation channels voxel by voxel to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, conserved frequencies per amino acid for each amino acid, annotation channels, and structural confidence channels voxel by voxel to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, and one-hot encoding of a reference allele voxel by voxel to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, and pan-amino acid conserved frequencies on a voxel basis to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, pan-amino acid conserved frequencies, and annotation channels on a voxel basis to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, pan-amino acid conserved frequencies, annotation channels, and structural confidence channels on a voxel basis to generate a tensor.A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, and conserved frequencies per amino acid for each amino acid on a voxel basis to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, conserved frequencies per amino acid for each amino acid, and annotation channels on a voxel basis to generate a tensor. A tensor generator can connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, conserved frequencies per amino acid for each amino acid, annotation channels, and structural confidence channels on a voxel basis to generate a tensor.

[0226] In some embodiments, the system may additionally include an atomic rotation engine that rotates the atoms of amino acids before amino acid-specific distance channels are generated. The pathogenicity determination engine may be a neural network. In certain embodiments, the pathogenicity determination engine may be a convolutional neural network. The convolutional neural network may use 1x1x1 convolutions, 3x3x3 convolutions, rectified linear unit activation layers, batch normalization layers, fully connected layers, dropout regularization layers, and softmax classification layers. The 1x1x1 convolutions and 3x3x3 convolutions may be three-dimensional convolutions.

[0227] In some implementations, a layer of 1x1x1 convolutions can process a tensor and generate an intermediate output, which is a convolutional representation of the tensor. A sequence of layers of 3x3x3 convolutions can process an intermediate output and generate a flattened output. A fully connected layer can process a flattened output and generate denormalized outputs. A softmax classification layer can process denormalized outputs and generate exponentially normalized outputs that identify the likelihood of a variant being pathogenic or benign. A sigmoid layer can process denormalized outputs and generate a normalized output that identifies the likelihood of a variant being pathogenic. Voxels, atoms, and distances can have three-dimensional coordinates. The tensor can have at least three dimensions, the intermediate output can have at least three dimensions, and the flattened output can have one dimension.

[0228] In some implementations, the pathogenicity decision engine is a recurrent neural network. In other implementations, the pathogenicity decision engine is an attention-based neural network. In yet another implementation, the pathogenicity decision engine is a gradient-boosted tree. In yet another implementation, the pathogenicity decision engine is a state vector machine.

[0229] In other embodiments, the system may include a memory for storing atomic category distance channels for amino acids in a protein. An amino acid may have atoms for a plurality of atomic categories, and atomic categories within the plurality of atomic categories may specify atomic elements of the amino acid. Atomic category distance channels may have voxel distance values ​​for voxels within a plurality of voxels. The voxel distance values ​​may specify distances from corresponding voxels within the plurality of voxels to atoms within corresponding atomic categories within the plurality of atomic categories. The system may further include a pathogenicity determination engine configured to process a tensor comprising atomic category distance channels and alternative alleles of a protein expressed by a variant, and to determine the pathogenicity of the variant based at least partially on the tensor.

[0230] The system may further include a distance channel generator that centers a voxel grid of voxels on the respective atoms of each atomic category within a plurality of atomic categories. The distance channel generator may center a voxel grid on the alpha-carbon atom of at least one variant amino acid residue in a protein. The distances may be the nearest atomic distances from the corresponding voxel centers within the voxel grid to the nearest atoms within the corresponding atomic categories. The nearest atomic distances may be Euclidean distances. The nearest atomic distances may be normalized by dividing the Euclidean distances by the maximum nearest atomic distances. The distances may be the nearest atomic distances from the corresponding voxel centers within the voxel grid to the nearest atoms, regardless of the amino acid and the atomic categories of the amino acid. The nearest atomic distances may be Euclidean distances. The nearest atomic distances may be normalized by dividing the Euclidean distances by the maximum nearest atomic distances.

[0231] Other embodiments of the method described in this section may include a non-transient computer-readable storage medium that stores instructions executable by a processor to perform any of the methods described above. Another embodiment of the method described in this section may include a system comprising a memory and one or more processors—operable to perform any of the methods described above by executing instructions stored in said memory.

[0232] Item 1

[0233] 1. As a computer implementation method,

[0234] A step of storing amino acid-specific distance channels for multiple amino acids within a protein,

[0235] Each amino acid-specific distance channel has voxel-specific distance values ​​for voxels within multiple voxels, and

[0236] The step of storing distance values ​​per voxel, which specify the distances from corresponding voxels within a plurality of voxels to atoms of corresponding amino acids within a plurality of amino acids;

[0237] A step of processing a tensor comprising amino acid-specific distance channels and alternative alleles of a protein expressed by a variant; and

[0238] A computer-implemented method comprising the step of determining the pathogenicity of a variant based at least partially on a tensor.

[0239] 2. A computer-implemented method according to claim 1, further comprising the step of centering a voxel grid of voxels on the alpha-carbon atoms of each residue of the amino acid.

[0240] 3. A computer-implemented method according to claim 2, further comprising the step of centering a voxel grid on the alpha-carbon atom of a specific amino acid residue corresponding to at least one variant amino acid in the protein.

[0241] 4. A computer-implemented method according to claim 3, further comprising the step of encoding the directionality of an amino acid and the position of a specific amino acid by multiplying the voxel-wise distance values ​​for such amino acid preceding the specific amino acid in a tensor with a directionality parameter.

[0242] 5. A computer implementation method in which, in paragraph 3, the distances are the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms of corresponding amino acids.

[0243] 6. A computer implementation method in which, in paragraph 5, the nearest atomic distances are Euclidean distances.

[0244] 7. A computer implementation method in which, in paragraph 6, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distance.

[0245] 8. A computer implementation method in which, in paragraph 5, the amino acid has alpha-carbon atoms, and the distances are the nearest alpha-carbon atom distances from corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding amino acid.

[0246] 9. A computer implementation method in which, in paragraph 5, the amino acid has beta-carbon atoms, and the distances are the nearest beta-carbon atom distances from corresponding voxel centers to the nearest beta-carbon atoms of the corresponding amino acid.

[0247] 10. A computer implementation method in which, in paragraph 5, the amino acids have backbone atoms, and the distances are the nearest backbone atom distances from corresponding voxel centers to the nearest backbone atoms of the corresponding amino acids.

[0248] 11. A computer implementation method in which, in paragraph 5, the amino acid has a side chain atom, and the distances are the nearest side chain atom distances from corresponding voxel centers to the nearest side chain atoms of the corresponding amino acid.

[0249] 12. A computer implementation method according to claim 3, further comprising the step of encoding a nearest atom channel in the tensor that specifies the distance from each voxel to the nearest atom, wherein the nearest atom is selected regardless of the amino acid and the atomic elements of the amino acid.

[0250] 13. A computer implementation method in which the distance in paragraph 12 is the Euclidean distance.

[0251] 14. A computer implementation method in which the distance is normalized by dividing the Euclidean distance by the maximum distance in Clause 13.

[0252] 15. A computer-implemented method according to paragraph 12, wherein the amino acid comprises a non-standard amino acid.

[0253] 16. A computer implementation method according to claim 1, wherein the tensor further comprises absent atomic channels that specify atoms not found within a predefined radius of the voxel center, and the absent atomic channels are one-hot encoded.

[0254] 17. A computer implementation method according to claim 1, wherein the tensor further comprises one-hot encoding of alternative alleles encoded voxel by each amino acid-specific distance channel.

[0255] 18. A computer implementation method according to claim 1, wherein the tensor further comprises a reference allele of a protein.

[0256] 19. A computer implementation method according to claim 18, wherein the tensor further comprises a one-hot encoding of a reference allele encoded voxel-wise in each of the amino acid-wise distance channels.

[0257] 20. A computer-implemented method according to claim 1, wherein the tensor further comprises evolutionary profiles specifying conservation levels of amino acids across multiple species.

[0258] 21. In paragraph 20, for each of the voxels,

[0259] Step of selecting the nearest atom across amino acid and atomic categories,

[0260] A step of selecting a pan-amino acid conservation frequency sequence for the amino acid residue containing the nearest atom, and

[0261] A computer-implemented method further comprising the step of making a pan-amino acid conserved frequency sequence available as one of evolutionary profiles.

[0262] 22. A computer implementation method according to claim 21, wherein the pan-amino acid conservation frequency sequence is configured for specific positions of residues as observed in multiple species.

[0263] 23. A computer implementation method according to claim 21, wherein the pan-amino acid conservation frequency sequence determines whether there is a missing conservation frequency for a specific amino acid.

[0264] 24. In paragraph 21, for each of the voxels,

[0265] The step of selecting the nearest atoms within each of the amino acids,

[0266] A step of selecting conservation frequencies per amino acid for each residue of an amino acid containing the nearest atoms, and

[0267] A computer-implemented method further comprising the step of making amino acid sugar conservation frequencies available as one of the evolutionary profiles.

[0268] 25. A computer implementation method according to paragraph 24, wherein the conserved frequencies per amino acid are configured for specific positions of residues as observed in multiple species.

[0269] 26. A computer implementation method in which the conservation frequencies per amino acid in paragraph 24 determine whether there is a missing conservation frequency for a specific amino acid.

[0270] 27. A computer implementation method according to claim 1, wherein the tensor further comprises annotation channels for amino acids, and the annotation channels are one-hot encoded in the tensor.

[0271] 28. A computer-implemented method in which, in paragraph 27, the annotation channels are molecular processing annotations comprising an initiator methionine, a signal, a change peptide, a propeptide, a chain, and a peptide.

[0272] 29. A computer-implemented method in which, in paragraph 27, the annotation channels are domain annotations including topology domains, transmembrane, intramembrane, domains, repeats, calcium binding, zinc clamps, deoxyribonucleic acid (DNA) binding, nucleotide binding, regions, double coils, motifs, and compositional biases.

[0273] 30. A computer-implemented method in which, in paragraph 27, the annotation channels are an active site, a metal bond, a bonding site, and a site annotation including the site.

[0274] 31. A computer implementation method in which, in paragraph 27, the annotation channels are amino acid modified annotations including non-standard residues, modified residues, lipidations, glycosylations, disulfide bonds, and crosslinkings.

[0275] 32. A computer-implemented method in which, in paragraph 27, the annotation channels are secondary structure annotations including helicals, spins, and beta strands.

[0276] 33. A computer-implemented method in which, in paragraph 27, the annotation channels are experimental information annotations including mutagenic, sequence uncertainty, sequence conflict, non-adjacent residues, and non-terminal residues.

[0277] 34. A computer implementation method according to claim 1, wherein the tensor further comprises structure reliability channels for amino acids, which specify the quality of each of the structures of the amino acids.

[0278] 35. In paragraph 34, a computer-implemented method in which the structural reliability channels are Global Model Quality Estimates (GMQEs).

[0279] 36. A computer-implemented method in which, in paragraph 34, structural reliability channels include Qualitative Model Energy Analysis (QMEAN) scores.

[0280] 37. A computer implementation method in which, in paragraph 34, the structural reliability channels are temperature factors that specify the extent to which residues satisfy the physical constraints of their respective protein structures.

[0281] 38. A computer implementation method in which, in paragraph 34, the structural reliability channels are template structure alignments that specify the degree to which the residues of atoms closest to the voxels have aligned template structures.

[0282] 39. A computer implementation method in which, in paragraph 38, the structural reliability channels are mold modeling scores of aligned mold structures.

[0283] 40. In paragraph 39, the structural reliability channels are the minimum mold modeling scores, the average of the mold modeling scores, and the maximum mold modeling score among the mold modeling scores, in a computer implementation method.

[0284] 41. A computer-implemented method according to claim 1, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms to one-hot encodings of alternative alleles on a voxel-by-voxel basis to generate a tensor.

[0285] 42. A computer implementation method according to claim 41, further comprising the step of connecting amino acid-specific distance channels for beta-carbon atoms to the one-hot encoding of alternative alleles on a voxel-by-voxel basis to generate a tensor.

[0286] 43. A computer-implemented method according to claim 42, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, and one-hot encodings of alternative alleles by voxel to generate a tensor.

[0287] 44. A computer-implemented method according to claim 43, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, and pan-amino acid conserved frequency sequences by voxel to generate a tensor.

[0288] 45. A computer-implemented method according to claim 44, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, pan-amino acid conserved frequency sequences, and annotation channels by voxel to generate a tensor.

[0289] 46. ​​A computer-implemented method according to claim 45, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, pan-amino acid conserved frequency sequences, annotation channels, and structural confidence channels by voxel to generate a tensor.

[0290] 47. A computer-implemented method according to claim 46, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, and conserved frequencies per amino acid for each amino acid on a voxel-by-voxel basis to generate a tensor.

[0291] 48. A computer-implemented method according to claim 47, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, conserved frequencies per amino acid for each amino acid, and annotation channels by voxel to generate a tensor.

[0292] 49. A computer-implemented method according to claim 48, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, conserved frequencies per amino acid for each amino acid, annotation channels, and structural confidence channels by voxel to generate a tensor.

[0293] 50. A computer-implemented method according to claim 49, further comprising the step of voxel-by-voxel connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, and one-hot encoding of a reference allele to generate a tensor.

[0294] 51. A computer-implemented method according to claim 50, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, and pan-amino acid conserved frequency sequences by voxel to generate a tensor.

[0295] 52. A computer-implemented method according to claim 51, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, pan-amino acid conserved frequency sequences, and annotation channels by voxel to generate a tensor.

[0296] 53. A computer-implemented method according to claim 52, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, pan-amino acid conserved frequency sequences, annotation channels, and structural confidence channels on a voxel-by-voxel basis to generate a tensor.

[0297] 54. A computer-implemented method according to claim 53, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, one-hot encoding of reference alleles, and conserved frequencies per amino acid for each amino acid on a voxel-by-voxel basis to generate a tensor.

[0298] 55. A computer-implemented method according to claim 54, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, one-hot encoding of reference alleles, conserved frequencies per amino acid for each amino acid, and annotation channels by voxel to generate a tensor.

[0299] 56. A computer-implemented method according to claim 55, further comprising the step of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, one-hot encoding of reference alleles, conserved frequencies per amino acid for each amino acid, annotation channels, and structural confidence channels by voxel to generate a tensor.

[0300] 57. A computer-implemented method according to claim 1, further comprising the step of rotating the atoms of the amino acids before the amino acid-specific distance channels are generated.

[0301] 58. A computer-implemented method according to claim 1, further comprising the step of using 1x1x1 convolutions, 3x3x3 convolutions, rectified linear unit activation layers, batch normalization layers, fully connected layers, dropout regularization layers, and softmax classification layers in a convolutional neural network.

[0302] 59. A computer implementation method in which, in paragraph 58, the 1x1x1 convolutions and the 3x3x3 convolutions are three-dimensional convolutions.

[0303] 60. A computer implementation method according to claim 58, wherein a layer of 1x1x1 convolutions processes a tensor and generates an intermediate output which is a convolutional representation of the tensor, a sequence of layers of 3x3x3 convolutions processes the intermediate output and generates a flattened output, a fully connected layer processes the flattened output and generates denormalized outputs, and a softmax classification layer processes the denormalized outputs and generates exponentially normalized outputs that identify the likelihood that a variant is pathogenic and benign.

[0304] 61. A computer implementation method according to paragraph 60, wherein the sigmoid layer processes denormalized outputs and generates a normalized output that identifies the likelihood that a variant is pathogenic.

[0305] 62. A computer implementation method according to paragraph 60, wherein the voxels, atoms, and distances have three-dimensional coordinates, the tensor has at least three dimensions, the intermediate output has at least three dimensions, and the flattened output has one dimension.

[0306] 63. As a computer implementation method,

[0307] As a step of storing atomic category distance channels for amino acids within a protein,

[0308] Amino acids have atoms for multiple atomic categories,

[0309] Atomic categories within multiple atomic categories specify the atomic elements of amino acids, and

[0310] Each atomic category distance channel has voxel-specific distance values ​​for voxels within multiple voxels, and

[0311] The step of storing distance values ​​per voxel, which specify distances from corresponding voxels within a plurality of voxels to atoms within corresponding atomic categories within a plurality of atomic categories;

[0312] A step of processing a tensor comprising atomic category-specific distance channels and alternative alleles of a protein expressed by a variant; and

[0313] A computer-implemented method comprising the step of determining the pathogenicity of a variant based at least partially on a tensor.

[0314] 64. A computer-implemented method according to claim 63, further comprising the step of centering a voxel grid of voxels on each atom of each atomic category within a plurality of atomic categories.

[0315] 65. A computer-implemented method according to claim 64, further comprising the step of centering a voxel grid on the alpha-carbon atom of at least one variant amino acid residue in the protein.

[0316] 66. A computer implementation method in paragraph 65, wherein the distances are the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms within corresponding atomic categories.

[0317] 67. A computer implementation method in which, in paragraph 66, the nearest atomic distances are Euclidean distances.

[0318] 68. A computer implementation method in which, in paragraph 67, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distances.

[0319] 69. A computer implementation method in which, in paragraph 68, the distances are the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms, regardless of amino acids and atomic categories of amino acids.

[0320] 70. A computer implementation method in paragraph 69, wherein the nearest atomic distances are Euclidean distances.

[0321] 71. A computer implementation method in which, in paragraph 70, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distances.

[0322] 1. One or more computer-readable media storing computer-executable instructions, wherein the computer-executable instructions, when executed on one or more processors,

[0323] As an operation of storing amino acid-specific distance channels for multiple amino acids within a protein,

[0324] Each amino acid-specific distance channel has voxel-specific distance values ​​for voxels within multiple voxels, and

[0325] The operation of storing voxel-specific distance values, which specify distances from corresponding voxels within a plurality of voxels to atoms of corresponding amino acids within a plurality of amino acids; the operation of processing a tensor comprising amino acid-specific distance channels and alternative alleles of a protein expressed by a variant; and

[0326] Computer-readable media configured to perform operations including operations determining the pathogenicity of a variant based at least partially on a tensor.

[0327] 2. Computer-readable media according to claim 1, wherein the operations further comprise centering a voxel grid of voxels on the alpha-carbon atoms of each residue of the amino acid.

[0328] 3. Computer-readable media according to paragraph 2, wherein the operations further comprise centering a voxel grid on the alpha-carbon atom of a specific amino acid residue corresponding to at least one variant amino acid in the protein.

[0329] 4. Computer-readable media according to paragraph 3, wherein the operations further include an operation of encoding the directionality of an amino acid and the position of a specific amino acid by multiplying the voxel-wise distance values ​​for such amino acid preceding the specific amino acid in a tensor with a directionality parameter.

[0330] 5. Computer-readable media, wherein, in paragraph 3, the distances are the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms of corresponding amino acids.

[0331] 6. Computer-readable media in which, in paragraph 5, the nearest atomic distances are Euclidean distances.

[0332] 7. Computer-readable media in which, in paragraph 6, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distance.

[0333] 8. Computer-readable media, wherein, in paragraph 5, the amino acid has alpha-carbon atoms, and the distances are the nearest alpha-carbon atom distances from the corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding amino acid.

[0334] 9. Computer-readable media, wherein, in paragraph 5, the amino acid has beta-carbon atoms, and the distances are the nearest beta-carbon atom distances from the corresponding voxel centers to the nearest beta-carbon atoms of the corresponding amino acid.

[0335] 10. Computer-readable media, wherein, in paragraph 5, the amino acids have backbone atoms, and the distances are the nearest backbone atom distances from corresponding voxel centers to the nearest backbone atoms of the corresponding amino acids.

[0336] 11. Computer-readable media, wherein, in paragraph 5, the amino acid has side chain atoms, and the distances are the nearest side chain atom distances from corresponding voxel centers to the nearest side chain atoms of the corresponding amino acid.

[0337] 12. Computer-readable media, wherein, in paragraph 3, the operations further include an operation to encode a nearest atom channel in the tensor that specifies the distance from each voxel to the nearest atom, and the nearest atom is selected regardless of the amino acid and the atomic elements of the amino acid.

[0338] 13. Computer-readable media in paragraph 12, wherein the distance is Euclidean distance.

[0339] 14. Computer-readable media in which, in paragraph 13, the distance is normalized by dividing the Euclidean distance by the maximum distance.

[0340] 15. In paragraph 12, the amino acids are computer-readable media comprising non-standard amino acids.

[0341] 16. Computer-readable media, wherein the tensor further comprises absent atomic channels specifying atoms not found within a predefined radius of the voxel center, and the absent atomic channels are one-hot encoded.

[0342] 17. Computer-readable media according to claim 1, wherein the tensor further comprises one-hot encoding of alternative alleles encoded voxel-wise in each amino acid-wise distance channel.

[0343] 18. Computer-readable media according to paragraph 1, wherein the tensor further comprises a reference allele of a protein.

[0344] 19. Computer-readable media according to paragraph 18, wherein the tensor further comprises a one-hot encoding of a reference allele encoded voxel-wise in each of the amino acid-wise distance channels.

[0345] 20. Computer-readable media according to claim 1, wherein the tensor further comprises evolutionary profiles specifying conservation levels of amino acids across multiple species.

[0346] 21. In paragraph 20, the operations, for each voxel,

[0347] The operation of selecting the nearest atom across amino acid and atomic categories,

[0348] Operation of selecting a pan-amino acid conservation frequency sequence for the amino acid residue containing the nearest atom, and

[0349] Computer-readable media further comprising an operation that makes a pan-amino acid conserved frequency sequence available as one of evolutionary profiles.

[0350] 22. Computer-readable media in which, in paragraph 21, the pan-amino acid conservation frequency sequence is configured for specific positions of residues as observed in multiple species.

[0351] 23. Computer-readable media in which, in paragraph 21, the pan-amino acid conservation frequency sequence specifies whether there is a missing conservation frequency for a specific amino acid.

[0352] 24. In paragraph 21, the operations, for each voxel,

[0353] The action of selecting the nearest atom within each of the amino acids,

[0354] The operation of selecting conservation frequencies per amino acid for each residue of an amino acid containing the nearest atoms, and

[0355] Computer-readable media further comprising an operation that makes amino acid sugar conservation frequencies available as one of evolutionary profiles.

[0356] 25. Computer-readable media in which, in paragraph 24, the conserved frequencies per amino acid are configured for specific positions of residues as observed in multiple species.

[0357] 26. Computer-readable media in paragraph 24, wherein the retention frequencies per amino acid specify whether there is a missing retention frequency for a particular amino acid.

[0358] 27. Computer-readable media, wherein the tensor further comprises annotation channels for amino acids, and the annotation channels are one-hot encoded in the tensor.

[0359] 28. In paragraph 27, computer-readable media in which the annotation channels are molecular processing annotations comprising initiator methionine, signal, change peptide, propeptide, chain, and peptide.

[0360] 29. Computer-readable media, wherein the annotation channels are domain annotations including topology domains, transmembranes, intramembranes, domains, repeats, calcium bindings, zinc clips, deoxyribonucleic acid (DNA) bindings, nucleotide bindings, regions, double coils, motifs, and compositional biases.

[0361] 30. Computer-readable media, wherein the annotation channels are active sites, metal bonds, bonding sites, and site annotations including sites.

[0362] 31. Computer-readable media, wherein the annotation channels are amino acid modified annotations including non-standard residues, modified residues, lipidations, glycosylations, disulfide bonds, and crosslinks.

[0363] 32. Computer-readable media, wherein the annotation channels are secondary structure annotations including helicals, spins, and beta strands in paragraph 27.

[0364] 33. Computer-readable media, wherein the annotation channels are experimental information annotations including mutagenic, sequence uncertainty, sequence conflict, non-adjacent residues, and non-terminal residues.

[0365] 34. Computer-readable media according to claim 1, wherein the tensor further comprises structure confidence channels for amino acids that specify the quality of each of the structures of the amino acids.

[0366] 35. In paragraph 34, the structural reliability channels are computer-readable media that are Global Model Quality Estimates (GMQEs).

[0367] 36. In paragraph 34, the structural reliability channels are computer-readable media containing Qualitative Model Energy Analysis (QMEAN) scores.

[0368] 37. In paragraph 34, the structure reliability channels are computer-readable media, which are temperature factors that specify the extent to which residues satisfy the physical constraints of their respective protein structures.

[0369] 38. Computer-readable media, wherein, in paragraph 34, the structural reliability channels are template structure alignments that specify the degree to which the residues of atoms closest to the voxels have aligned template structures.

[0370] 39. Computer-readable media, wherein, in paragraph 38, the structural reliability channels are mold modeling scores of aligned mold structures.

[0371] 40. In paragraph 39, the structural reliability channels are computer-readable media that are the minimum mold modeling scores, the average of the mold modeling scores, and the maximum mold modeling score among the mold modeling scores.

[0372] 41. Computer-readable media according to claim 1, wherein the operations further comprise connecting amino acid-specific distance channels for alpha-carbon atoms to one-hot encodings of alternative alleles by voxel to generate a tensor.

[0373] 42. Computer-readable media, wherein, in paragraph 41, the operations further comprise connecting amino acid-specific distance channels for beta-carbon atoms to one-hot encodings of alternative alleles on a voxel-by-voxel basis to generate a tensor.

[0374] 43. Computer-readable media, wherein, in paragraph 42, the operations further comprise connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, and one-hot encodings of alternative alleles by voxel to generate a tensor.

[0375] 44. Computer-readable media, wherein, in paragraph 43, the operations further comprise amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, and operations of linking pan-amino acid conserved frequency sequences by voxel to generate a tensor.

[0376] 45. Computer-readable media, wherein, in paragraph 44, the operations further comprise connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, pan-amino acid conserved frequency sequences, and annotation channels by voxel to generate a tensor.

[0377] 46. ​​Computer-readable media, wherein, in paragraph 45, the operations further comprise operations of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, pan-amino acid conserved frequency sequences, annotation channels, and structural confidence channels by voxel to generate a tensor.

[0378] 47. Computer-readable media, wherein, in paragraph 46, the operations further comprise amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, and voxel-specific conserved frequencies per amino acid for each amino acid to generate a tensor.

[0379] 48. Computer-readable media, wherein, in paragraph 47, the operations further comprise amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, conserved frequencies per amino acid for each amino acid, and voxel-specific annotation channels to generate a tensor.

[0380] 49. Computer-readable media, wherein, in paragraph 48, the operations further comprise operations to connect amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, conserved frequencies per amino acid for each amino acid, annotation channels, and structural confidence channels by voxel to generate a tensor.

[0381] 50. Computer-readable media, wherein, in paragraph 49, the operations further comprise connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, and one-hot encoding of a reference allele by voxel to generate a tensor.

[0382] 51. Computer-readable media, wherein the operations further comprise amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, one-hot encoding of reference alleles, and voxel-specific linking of pan-amino acid conserved frequency sequences to generate a tensor.

[0383] 52. Computer-readable media according to claim 51, wherein the operations further comprise connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, pan-amino acid conserved frequency sequences, and annotation channels by voxel to generate a tensor.

[0384] 53. Computer-readable media, wherein, in paragraph 52, the operations further comprise operations of connecting amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, pan-amino acid conserved frequency sequences, annotation channels, and structural confidence channels by voxel to generate a tensor.

[0385] 54. Computer-readable media, wherein, in paragraph 53, the operations further comprise amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of alternative alleles, one-hot encoding of reference alleles, and voxel-specific conserved frequencies per amino acid for each amino acid to generate a tensor.

[0386] 55. Computer-readable media, wherein the operations further comprise, for generating a tensor, amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, conserved frequencies per amino acid for each amino acid, and voxel-specific annotation channels.

[0387] 56. Computer-readable media, wherein the operations further comprise, for generating a tensor, amino acid-specific distance channels for alpha-carbon atoms, amino acid-specific distance channels for beta-carbon atoms, one-hot encoding of an alternative allele, one-hot encoding of a reference allele, conservation frequencies per amino acid for each amino acid, annotation channels, and structure confidence channels by voxel.

[0388] 57. Computer-readable media according to claim 1, wherein the operations further include the operation of rotating the atoms of the amino acids before the amino acid-specific distance channels are created.

[0389] 58. Computer-readable media, wherein, in paragraph 1, the operations further comprise operations using 1x1x1 convolutions, 3x3x3 convolutions, rectified linear unit activation layers, batch normalization layers, full connection layers, dropout regularization layers, and softmax classification layers in a convolutional neural network.

[0390] 59. Computer-readable media in which, in paragraph 58, 1x1x1 convolutions and 3x3x3 convolutions are three-dimensional convolutions.

[0391] 60. Computer-readable media, wherein, in paragraph 58, a layer of 1x1x1 convolutions processes a tensor and generates an intermediate output which is a convolutional representation of the tensor, a sequence of layers of 3x3x3 convolutions processes the intermediate output and generates a flattened output, a fully connected layer processes the flattened output and generates denormalized outputs, and a softmax classification layer processes the denormalized outputs and generates exponentially normalized outputs which identify the likelihood that a variant is pathogenic and benign.

[0392] 61. Computer-readable media in which, in paragraph 60, the sigmoid layer processes denormalized outputs and generates a normalized output that identifies the likelihood that a variant is pathogenic.

[0393] 62. Computer-readable media in which, in paragraph 60, the voxels, atoms, and distances have three-dimensional coordinates, the tensor has at least three dimensions, the intermediate output has at least three dimensions, and the flattened output has one dimension.

[0394] 63. One or more computer-readable media storing computer-executable instructions, wherein the computer-executable instructions, when executed on one or more processors,

[0395] As an operation of storing atomic category-specific distance channels for amino acids within a protein,

[0396] Amino acids have atoms for multiple atomic categories,

[0397] Atomic categories within multiple atomic categories specify the atomic elements of amino acids, and

[0398] Each atomic category distance channel has voxel-specific distance values ​​for voxels within multiple voxels, and

[0399] The operation of storing distance values ​​per voxel, which specify distances from corresponding voxels within a plurality of voxels to atoms within corresponding atomic categories within a plurality of atomic categories;

[0400] An operation to process a tensor comprising atomic category-specific distance channels and alternative alleles of a protein expressed by a variant; and

[0401] Computer-readable media configured to perform operations including operations determining the pathogenicity of a variant based at least partially on a tensor.

[0402] 64. Computer-readable media according to paragraph 63, wherein the operations further include the operation of centering a voxel grid of voxels on the respective atoms of each atomic category within a plurality of atomic categories.

[0403] 65. Computer-readable media according to claim 64, wherein the operations further comprise centering a voxel grid on the alpha-carbon atom of at least one variant amino acid residue in the protein.

[0404] 66. Computer-readable media, wherein, in paragraph 65, the distances are the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms within corresponding atomic categories.

[0405] 67. Computer-readable media in which, in paragraph 66, the nearest atomic distances are Euclidean distances.

[0406] 68. Computer-readable media in which, in paragraph 67, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distances.

[0407] 69. Computer-readable media, wherein, in paragraph 68, the distances are the nearest atomic distances from corresponding voxel centers within a voxel grid to the nearest atoms, regardless of amino acids and atomic categories of amino acids.

[0408] 70. Computer-readable media in which, in paragraph 69, the nearest atomic distances are Euclidean distances.

[0409] 71. Computer-readable media in which, in paragraph 70, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distances.

[0410] Specific implementation examples 2

[0411] In some embodiments, the system includes a voxelizer that accesses the three-dimensional structure of a reference amino acid sequence of a protein and creates amino acid-specific distance channels by fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure on an amino acid-by-amino acid basis. Each amino acid-specific distance channel has a three-dimensional distance value for each voxel within the three-dimensional grid of voxels. The three-dimensional distance value specifies the distance from a corresponding voxel within the three-dimensional grid of voxels to atoms of the corresponding reference amino acid within the reference amino acid sequence. The system further includes an alternative allele encoder that encodes an alternative allele amino acid for each voxel within the three-dimensional grid of voxels. The alternative allele amino acid is a three-dimensional representation of the one-hot encoding of the variant amino acid expressed by the variant nucleotide. The system further includes an evolutionary conservation encoder that encodes an evolutionary conservation sequence for each voxel within the three-dimensional grid of voxels. The evolutionary conservation sequence may be a three-dimensional representation of amino acid-specific conservation frequencies across multiple species. Amino acid-specific conservation frequencies can be selected based on amino acid proximity to corresponding voxels. The system further includes a convolutional neural network configured to apply three-dimensional convolutions to a tensor containing amino acid-specific distance channels encoded with alternative allele amino acids and their respective evolutionarily conserved sequences. The convolutional neural network can also be configured to determine the pathogenicity of variant nucleotides based at least partially on the tensor.

[0412] The voxelizer can center a three-dimensional grid of voxels on the alpha-carbon atoms of each residue of the reference amino acid within the reference amino acid sequence. The voxelizer can center a three-dimensional grid of voxels on the alpha-carbon atoms of a specific residue of the reference amino acid located in the variant amino acid.

[0413] In some embodiments, the system may be further configured to encode the directionality of a reference amino acid and the position of a specific reference amino acid within a reference amino acid sequence by multiplying three-dimensional distance values ​​for such reference amino acid preceding the specific reference amino acid in a tensor by a directionality parameter. The distances may be the nearest atomic distances from corresponding voxel centers within a three-dimensional grid of voxels to the nearest atoms of the corresponding reference amino acid. The nearest atomic distances may be Euclidean distances and may be normalized by dividing the Euclidean distances by the maximum nearest atomic distance.

[0414] In some embodiments, the reference amino acid may have alpha-carbon atoms, and the distances may be the nearest alpha-carbon atom distances from the corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding reference amino acid. In some embodiments, the reference amino acid may have beta-carbon atoms, and the distances may be the nearest beta-carbon atom distances from the corresponding voxel centers to the nearest beta-carbon atoms of the corresponding reference amino acid. In some embodiments, the reference amino acid may have backbone atoms, and the distances may be the nearest backbone atom distances from the corresponding voxel centers to the nearest backbone atoms of the corresponding reference amino acid. In some embodiments, the amino acid may have side chain atoms, and the distances may be the nearest side chain atom distances from the corresponding voxel centers to the nearest side chain atoms of the corresponding reference amino acid.

[0415] In some embodiments, the system may be further configured to encode a nearest atom channel in the tensor that specifies the distance from each voxel to the nearest atom. The nearest atom may be selected regardless of the amino acid and the atomic elements of the amino acid. The distance may be a Euclidean distance and may be normalized by dividing the Euclidean distance by the maximum distance. The amino acids may include non-standard amino acids. The tensor may further include an absent atom channel that specifies atoms not found within a predefined radius of the voxel center. The absent atom channel may be one-hot encoded.

[0416] In some embodiments, the system may additionally include a reference allele encoder that encodes a reference allele amino acid on a voxel-by-voxel basis for each 3D distance value at the amino acid position level. The reference allele amino acid may be a 3D representation of the one-hot encoding of a reference amino acid sequence. Amino acid-specific conservation frequencies may specify the conservation levels of each amino acid across multiple species.

[0417] In some embodiments, the evolutionary conservation encoder may select the atom closest to the corresponding voxel across reference amino acid and atomic categories, select pan-amino acid conservation frequencies for the residue of the reference amino acid containing the closest atom, and use a three-dimensional representation of the pan-amino acid conservation frequencies as an evolutionary conservation sequence. The pan-amino acid conservation frequencies may be configured for specific positions of the residue, as observed in multiple species. The pan-amino acid conservation frequencies may determine whether there are missing conservation frequencies for a specific reference amino acid.

[0418] In some embodiments, the evolutionary conservation encoder may select the nearest atoms for each corresponding voxel within each of the reference amino acids, select the conservation frequencies per amino acid for each residue of the reference amino acid containing the nearest atoms, and use a three-dimensional representation of the conservation frequencies per amino acid as the evolutionary conservation sequence. The conservation frequencies per amino acid may be configured for specific positions of residues as observed in multiple species. The conservation frequencies per amino acid may determine whether there are missing conservation frequencies for a specific reference amino acid.

[0419] In some embodiments, the system may further include an annotation encoder that encodes one or more annotation channels per voxel for each three-dimensional distance value. The annotation channels may be three-dimensional representations of one-hot encoding of residue annotations and may be molecular processing annotations including initiator methionine, signal, change peptide, propeptide, chain, and peptide. In some embodiments, the annotation channels may be region annotations including topology domains, transmembranes, intramembranes, domains, repeats, calcium binding, zinc clamps, deoxyribonucleic acid (DNA) binding, nucleotide binding, regions, double coils, motifs, and compositional biases, or may be site annotations including active sites, metal binding, binding sites, and sites. In some embodiments, the annotation channels may be amino acid modification annotations including non-standard residues, modification residues, lipidation, glycosylation, disulfide binding, and crosslinking, or may be secondary structure annotations including helices, spins, and beta strands. The annotation channels may be experimental information annotations including mutagenic, sequence uncertainty, sequence conflict, non-adjacent residues, and non-terminal residues.

[0420] In some embodiments, the system may additionally include a structure confidence encoder that encodes one or more structure confidence channels per voxel for each three-dimensional distance value. The structure confidence channels may be three-dimensional representations of confidence scores that specify the quality of their respective residual structures. The structure confidence channels may be Global Model Quality Estimates (GMQEs), Qualitative Model Energy Analysis (QMEAN) scores, temperature factors that specify the extent to which residues satisfy the physical constraints of their respective protein structures, template structure alignments that specify the extent to which residues of atoms closest to the voxels have aligned template structures, template modeling scores of aligned template structures, or minimum template modeling scores, the mean of template modeling scores, and maximum template modeling scores among the template modeling scores.

[0421] In some embodiments, the system may additionally include an atomic rotation engine that rotates atoms before amino acid-specific distance channels are generated.

[0422] Convolutional neural networks may use 1x1x1 convolutions, 3x3x3 convolutions, rectified linear unit activation layers, batch normalization layers, fully connected layers, dropout regularization layers, and softmax classification layers. The 1x1x1 convolutions and 3x3x3 convolutions may be three-dimensional convolutions. In some embodiments, a layer of 1x1x1 convolutions may process a tensor and generate an intermediate output which is a convolutional representation of the tensor. A sequence of layers of 3x3x3 convolutions may process the intermediate output and generate a flattened output. A fully connected layer may process the flattened output and generate denormalized outputs. A softmax classification layer may process the denormalized outputs and generate exponentially normalized outputs that identify the likelihood that variant nucleotides are pathogenic or benign.

[0423] In some embodiments, the sigmoid layer can process denormalized outputs and generate a normalized output that identifies the likelihood that a variant nucleotide is pathogenic. The convolutional neural network may be an attention-based neural network. The tensor may include amino acid-specific distance channels further encoded with reference allele amino acids, amino acid-specific distance channels further encoded with annotation channels, or amino acid-specific distance channels further encoded with structural confidence channels.

[0424] In some embodiments, the system may include a voxelizer that accesses the three-dimensional structure of a reference amino acid sequence of a protein and creates distance channels by atomic category by fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level. The atoms span multiple atomic categories that specify the atomic elements of the amino acids. Each distance channel by atomic category has a three-dimensional distance value for each voxel within the three-dimensional grid of voxels. The three-dimensional distance value specifies the distance from a corresponding voxel within the three-dimensional grid of voxels to atoms of corresponding atomic categories within the multiple atomic categories. The system further includes an alternative allele encoder that encodes an alternative allele amino acid for each voxel within the three-dimensional grid of voxels. The alternative allele amino acid is a three-dimensional representation of the one-hot encoding of a variant amino acid expressed by a variant nucleotide. The system further includes an evolutionary conservation encoder that encodes an evolutionary conservation sequence for each voxel within the three-dimensional grid of voxels. The evolutionarily conserved sequence may be a three-dimensional representation of amino acid-specific conserved frequencies across multiple species. The amino acid-specific conserved frequencies may be selected based on amino acid proximity to the corresponding voxel. The system further includes a convolutional neural network configured to apply three-dimensional convolutions to a tensor containing atomic category-specific distance channels encoded with alternative allele amino acids and their respective evolutionarily conserved sequences, and to determine the pathogenicity of variant nucleotides based at least partially on the tensor.

[0425] In some embodiments, the system includes a voxelizer that accesses the three-dimensional structure of a reference amino acid sequence of a protein and creates amino acid-specific distance channels by fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure on an amino acid-by-amino acid basis. Each amino acid-specific distance channel may have a three-dimensional distance value for each voxel within the three-dimensional grid of voxels. The three-dimensional distance value may specify the distance from a corresponding voxel within the three-dimensional grid of voxels to atoms of the corresponding reference amino acid within the reference amino acid sequence. The system further includes an alternative allele encoder that encodes an alternative allele amino acid for each voxel within the three-dimensional grid of voxels. The alternative allele amino acid is a three-dimensional representation of the one-hot encoding of the variant amino acid expressed by the variant nucleotide. The system further includes an evolutionary conservation encoder that encodes an evolutionary conservation sequence for each voxel within the three-dimensional grid of voxels. The evolutionary conservation sequence may be a three-dimensional representation of amino acid-specific conservation frequencies across multiple species. Amino acid-specific conservation frequencies can be selected based on amino acid proximity to the corresponding voxel. The system further includes a tensor generator configured to generate a tensor containing alternative allele amino acids and amino acid-specific distance channels encoded with their respective evolutionarily conserved sequences.

[0426] In some embodiments, the system includes a voxelizer that accesses the three-dimensional structure of a reference amino acid sequence of a protein and creates atomic category distance channels by fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level. Atoms may span multiple atomic categories that specify the atomic elements of the amino acid. Each atomic category distance channel may have a three-dimensional distance value for each voxel within the three-dimensional grid of voxels. The three-dimensional distance value may specify the distance from a corresponding voxel within the three-dimensional grid of voxels to atoms of corresponding atomic categories within the multiple atomic categories. The system further includes an alternative allele encoder that encodes an alternative allele amino acid for each voxel within the three-dimensional grid of voxels. The alternative allele amino acid is a three-dimensional representation of the one-hot encoding of a variant amino acid expressed by a variant nucleotide. The system further includes an evolutionary conservation encoder that encodes an evolutionary conservation sequence for each voxel within the three-dimensional grid of voxels. An evolutionarily conserved sequence may be a three-dimensional representation of amino acid-specific conserved frequencies across multiple species. Amino acid-specific conserved frequencies may be selected based on amino acid proximity to corresponding voxels. The system further includes a tensor generator configured to generate a tensor containing atomic category-specific distance channels encoded with alternative allele amino acids and their respective evolutionarily conserved sequences.

[0427] Item 2

[0428] 1. As a computer implementation method,

[0429] A step of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate amino acid-specific distance channels,

[0430] Each amino acid-specific distance channel has a 3D distance value for each voxel within the 3D grid of voxels, and

[0431] A step of generating a three-dimensional distance value that specifies the distance from a corresponding voxel within a three-dimensional grid of voxels to atoms of a corresponding reference amino acid within a reference amino acid sequence;

[0432] A step of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0433] A step of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across amino acid-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel;

[0434] A step of applying three-dimensional convolutions to a tensor comprising amino acid-specific distance channels encoded by alternative allele channels and their respective evolutionarily conserved channels; and

[0435] A computer-implemented method comprising the step of determining the pathogenicity of a variant nucleotide based at least partially on a tensor.

[0436] 2. A computer-implemented method according to Item 1, further comprising the step of centering a three-dimensional grid of voxels on the alpha-carbon atoms of each residue of a reference amino acid within a reference amino acid sequence.

[0437] 3. A computer implementation method according to Item 2, further comprising the step of centering a three-dimensional grid of voxels on the alpha-carbon atom of a residue of a specific reference amino acid corresponding to a variant amino acid.

[0438] 4. A computer-implemented method according to Item 3, further comprising the step of encoding the directionality of a reference amino acid and the position of a specific reference amino acid within a reference amino acid sequence by multiplying three-dimensional distance values ​​for such reference amino acid preceding a specific reference amino acid in a tensor by a directionality parameter.

[0439] 5. A computer implementation method in which, in Item 4, the distances are the nearest atomic distances from corresponding voxel centers within a three-dimensional grid of voxels to the nearest atoms of a corresponding reference amino acid.

[0440] 6. A computer implementation method of Item 5 in which the nearest atomic distances are Euclidean distances.

[0441] 7. A computer implementation method of Item 6 in which the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distance.

[0442] 8. A computer implementation method in which, in Item 5, the reference amino acid has alpha-carbon atoms, and the distances are the nearest alpha-carbon atom distances from the corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding reference amino acid.

[0443] 9. A computer implementation method in which, in Item 5, the reference amino acid has beta-carbon atoms, and the distances are the nearest beta-carbon atom distances from the corresponding voxel centers to the nearest beta-carbon atoms of the corresponding reference amino acid.

[0444] 10. A computer implementation method in which, in Item 5, the reference amino acid has backbone atoms, and the distances are the nearest backbone atom distances from the corresponding voxel centers to the nearest backbone atoms of the corresponding reference amino acid.

[0445] 11. A computer implementation method in which, in Item 5, the amino acids have side chain atoms, and the distances are the nearest side chain atom distances from corresponding voxel centers to the nearest side chain atoms of the corresponding reference amino acid.

[0446] 12. A computer implementation method according to Item 3, further comprising the step of encoding a nearest atom channel in the tensor that specifies the distance from each voxel to the nearest atom, wherein the nearest atom is selected regardless of the amino acid and the atomic elements of the amino acid.

[0447] 13. A computer implementation method of Item 12, wherein the distance is the Euclidean distance.

[0448] 14. A computer implementation method of Item 13 in which the distance is normalized by dividing the Euclidean distance by the maximum distance.

[0449] 15. A computer implementation method of Item 12, wherein the amino acid includes a non-standard amino acid.

[0450] 16. A computer implementation method according to Item 1, wherein the tensor further comprises an absent atomic channel that specifies atoms not found within a predefined radius of the voxel center.

[0451] 17. A computer implementation method in which, in Item 16, the absent atomic channel is one-hot encoded.

[0452] 18. A computer-implemented method according to Item 1, further comprising the step of encoding a reference allele channel on a voxel-by-voxel basis in each voxel within a three-dimensional grid of voxels.

[0453] 19. A computer implementation method of Item 18, wherein the reference allele amino acid is a three-dimensional representation of the one-hot encoding of the reference amino acid experiencing the variant amino acid.

[0454] 20. A computer implementation method of Item 1, wherein amino acid-specific conservation frequencies specify the conservation levels of each amino acid across multiple species.

[0455] 21. Regarding Item 20,

[0456] A step of selecting the nearest atom to the corresponding voxel across reference amino acid and atomic categories,

[0457] A step of selecting pan-amino acid conservation frequencies for a residue of a reference amino acid containing the nearest atom, and

[0458] A computer-implemented method further comprising the step of using a three-dimensional representation of pan-amino acid conservation frequencies as an evolutionary conservation channel.

[0459] 22. A computer implementation method of Item 21 in which pan-amino acid conservation frequencies are configured for specific positions of residues as observed in multiple species.

[0460] 23. A computer implementation method for item 21, wherein pan-amino acid conservation frequencies determine whether there are missing conservation frequencies for a specific reference amino acid.

[0461] 24. Regarding Item 21,

[0462] A step of selecting the nearest atoms to the corresponding voxels within each of the reference amino acids,

[0463] A step of selecting conservation frequencies per amino acid for each residue of a reference amino acid containing the nearest atoms, and

[0464] A computer-implemented method further comprising the step of using a three-dimensional representation of amino acid conserved frequencies as an evolutionary conservation channel.

[0465] 25. A computer implementation method of Item 24, wherein the conservation frequencies per amino acid are configured for specific positions of residues as observed in multiple species.

[0466] 26. A computer implementation method for item 24, wherein the conservation frequencies per amino acid determine whether there are missing conservation frequencies for a specific reference amino acid.

[0467] 27. A computer implementation method according to Item 1, further comprising the step of encoding one or more annotation channels on a voxel-by-voxel basis for each voxel in a three-dimensional grid of voxels, wherein the annotation channels are three-dimensional representations of one-hot encoding of residual annotations.

[0468] 28. A computer-implemented method in which, in item 27, the annotation channels are molecular processing annotations comprising an initiator methionine, a signal, a change peptide, a propeptide, a chain, and a peptide.

[0469] 29. Computer implementation method of Item 27, wherein the annotation channels are domain annotations including topology domains, transmembrane, intramembrane, domains, repeats, calcium binding, zinc clamps, deoxyribonucleic acid (DNA) binding, nucleotide binding, regions, double coils, motifs, and compositional biases.

[0470] 30. A computer-implemented method of Item 27, wherein the annotation channels are site annotations including an active site, a metal bond, a bonding site, and a site.

[0471] 31. A computer implementation method in which, in item 27, the annotation channels are amino acid modified annotations including non-standard residues, modified residues, lipidations, glycosylations, disulfide bonds, and crosslinkings.

[0472] 32. A computer implementation method of Item 27, wherein the annotation channels are secondary structure annotations including helicals, spins, and beta strands.

[0473] 33. A computer-implemented method of Item 27, wherein the annotation channels are experimental information annotations including mutagenic, sequence uncertainty, sequence conflict, non-adjacent residues, and non-terminal residues.

[0474] 34. A computer implementation method according to Item 1, further comprising the step of encoding one or more structural reliability channels on a voxel-by-voxel basis for each voxel within a three-dimensional grid of voxels, wherein the structural reliability channels are three-dimensional representations of reliability scores specifying the quality of each residual structure.

[0475] 35. In Item 34, the structural reliability channels are computer-implemented methods that are Global Model Quality Estimates (GMQEs).

[0476] 36. In Item 34, the structural reliability channels are computer implementation methods in which the qualitative model energy analysis (QMEAN) scores.

[0477] 37. In item 34, a computer implementation method in which structural reliability channels are temperature factors that specify the extent to which residues satisfy the physical constraints of their respective protein structures.

[0478] 38. A computer implementation method of Item 34, wherein the structural reliability channels are template structure alignments that specify the degree to which the residues of atoms closest to the voxels have aligned template structures.

[0479] 39. A computer implementation method in which, in Item 38, the structural reliability channels are the mold modeling scores of aligned mold structures.

[0480] 40. In Item 39, the structural reliability channels are the minimum mold modeling scores, the average of the mold modeling scores, and the maximum mold modeling score among the mold modeling scores, in a computer implementation method.

[0481] 41. A computer-implemented method according to Item 1, further comprising the step of rotating atoms before amino acid-specific distance channels are generated.

[0482] 42. A computer-implemented method according to Item 1, further comprising the step of using 1x1x1 convolutions, 3x3x3 convolutions, rectified linear unit activation layers, batch normalization layers, fully connected layers, dropout regularization layers, and softmax classification layers in a convolutional neural network.

[0483] 43. A computer implementation method in which, in Item 42, 1x1x1 convolutions and 3x3x3 convolutions are three-dimensional convolutions.

[0484] 44. A computer implementation method according to Item 42, wherein a layer of 1x1x1 convolutions processes a tensor and generates an intermediate output which is a convolutional representation of the tensor, a sequence of layers of 3x3x3 convolutions processes the intermediate output and generates a flattened output, a fully connected layer processes the flattened output and generates denormalized outputs, and a softmax classification layer processes the denormalized outputs and generates exponentially normalized outputs that identify the likelihood that variant nucleotides are pathogenic and benign.

[0485] 45. A computer implementation method of Item 44, wherein the sigmoid layer processes denormalized outputs and generates a normalized output that identifies the likelihood that a variant nucleotide is pathogenic.

[0486] 46. ​​In Item 1, the convolutional neural network is a computer implementation method that is an attention-based neural network.

[0487] 47. A computer implementation method of Item 1, wherein the tensor comprises amino acid-specific distance channels additionally encoded as a reference allele channel.

[0488] 48. A computer implementation method of Item 1, wherein the tensor comprises amino acid-specific distance channels additionally encoded as annotation channels.

[0489] 49. A computer implementation method of Item 1, wherein the tensor comprises amino acid-specific distance channels additionally encoded as structural confidence channels.

[0490] 50. As a computer implementation method,

[0491] A step of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate distance channels by atomic category,

[0492] Atoms span multiple atomic categories, and

[0493] Atomic categories within multiple atomic categories specify the atomic elements of amino acids, and

[0494] Each of the atomic category distance channels has a 3D distance value for each voxel within the 3D grid of voxels, and

[0495] A step of generating a 3D distance value that specifies the distance from a corresponding voxel within a 3D grid of voxels to atoms of corresponding atomic categories within a plurality of atomic categories;

[0496] A step of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0497] A step of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across atomic category-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel;

[0498] A step of applying three-dimensional convolutions to a tensor comprising atomic category-specific distance channels encoded with alternative allele channels and their respective evolutionarily conserved channels; and

[0499] A computer-implemented method comprising the step of determining the pathogenicity of a variant nucleotide based at least partially on a tensor.

[0500] 51. As a computer implementation method,

[0501] A step of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate amino acid-specific distance channels,

[0502] Each amino acid-specific distance channel has a 3D distance value for each voxel within the 3D grid of voxels, and

[0503] A step of generating a three-dimensional distance value that specifies the distance from a corresponding voxel within a three-dimensional grid of voxels to atoms of a corresponding reference amino acid within a reference amino acid sequence;

[0504] A step of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0505] A step of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across amino acid-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel; and

[0506] A computer-implemented method comprising the step of generating a tensor containing alternative allele channels and amino acid-specific distance channels encoded by their respective evolutionarily conserved channels.

[0507] 52. As a computer implementation method,

[0508] A step of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate distance channels by atomic category,

[0509] Atoms span multiple atomic categories, and

[0510] Atomic categories within multiple atomic categories specify the atomic elements of amino acids, and

[0511] Each of the atomic category distance channels has a 3D distance value for each voxel within the 3D grid of voxels, and

[0512] A step of generating a 3D distance value that specifies the distance from a corresponding voxel within a 3D grid of voxels to atoms of corresponding atomic categories within a plurality of atomic categories;

[0513] A step of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0514] A step of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across atomic category-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel; and

[0515] A computer-implemented method comprising the step of generating a tensor containing atomic category-specific distance channels encoded with alternative allele channels and their respective evolutionarily conserved channels.

[0516] 1. One or more computer-readable media storing computer-executable instructions, wherein the computer-executable instructions, when executed on one or more processors,

[0517] An operation of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate amino acid-specific distance channels,

[0518] Each amino acid-specific distance channel has a 3D distance value for each voxel within the 3D grid of voxels, and

[0519] An operation of generating a three-dimensional distance value that specifies the distance from a corresponding voxel within a three-dimensional grid of voxels to atoms of a corresponding reference amino acid within a reference amino acid sequence;

[0520] An operation of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0521] An operation of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across amino acid-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel;

[0522] An operation of applying three-dimensional convolutions to a tensor comprising amino acid-specific distance channels encoded by alternative allele channels and their respective evolutionarily conserved channels; and

[0523] Computer-readable media configured to perform operations including operations determining the pathogenicity of variant nucleotides based at least partially on a tensor.

[0524] 2. Computer-readable media according to Item 1, wherein the operations further include the operation of centering a three-dimensional grid of voxels on the alpha-carbon atoms of each residue of a reference amino acid within a reference amino acid sequence.

[0525] 3. Computer-readable media according to Item 2, wherein the operations further include the operation of centering a three-dimensional grid of voxels on the alpha-carbon atom of a specific reference amino acid residue corresponding to a variant amino acid.

[0526] 4. Computer-readable media, wherein, in Item 3, the operations further comprise an operation of encoding the directionality of a reference amino acid and the position of a specific reference amino acid within a reference amino acid sequence by multiplying three-dimensional distance values ​​for such reference amino acid preceding a specific reference amino acid in a tensor by a directionality parameter.

[0527] 5. Computer-readable media, wherein, in Item 4, the distances are the nearest atomic distances from corresponding voxel centers within a three-dimensional grid of voxels to the nearest atoms of the corresponding reference amino acid.

[0528] 6. Computer-readable media in which, in Item 5, the nearest atomic distances are Euclidean distances.

[0529] 7. Computer-readable media in which, in Item 6, the nearest atomic distances are normalized by dividing the Euclidean distances by the maximum nearest atomic distance.

[0530] 8. Computer-readable media, wherein, in Item 5, the reference amino acid has alpha-carbon atoms, and the distances are the nearest alpha-carbon atom distances from the corresponding voxel centers to the nearest alpha-carbon atoms of the corresponding reference amino acid.

[0531] 9. Computer-readable media, wherein, in Item 5, the reference amino acid has beta-carbon atoms, and the distances are the nearest beta-carbon atom distances from the corresponding voxel centers to the nearest beta-carbon atoms of the corresponding reference amino acid.

[0532] 10. Computer-readable media, wherein, in Item 5, the reference amino acid has backbone atoms, and the distances are the nearest backbone atom distances from the corresponding voxel centers to the nearest backbone atoms of the corresponding reference amino acid.

[0533] 11. Computer-readable media, wherein, in Item 5, the amino acids have side chain atoms, and the distances are the nearest side chain atom distances from corresponding voxel centers to the nearest side chain atoms of the corresponding reference amino acid.

[0534] 12. In Item 3, the operations further include an operation to encode a nearest atom channel in the tensor that specifies the distance from each voxel to the nearest atom, wherein the nearest atom is selected regardless of the amino acid and the atomic elements of the amino acid, computer-readable media.

[0535] 13. Computer-readable media in which, in Item 12, the distance is a Euclidean distance.

[0536] 14. Computer-readable media in which, in Item 13, the distance is normalized by dividing the Euclidean distance by the maximum distance.

[0537] 15. In Item 12, the amino acids are computer-readable media containing non-standard amino acids.

[0538] 16. Computer-readable media, wherein the tensor of Item 1 further comprises absent atomic channels that specify atoms not found within a predefined radius of the voxel center.

[0539] 17. In Item 16, computer-readable media in which the absent atomic channel is one-hot encoded.

[0540] 18. Computer-readable media according to Item 1, wherein the operations further include the operation of encoding a reference allele channel on a voxel-by-voxel basis in each voxel within a three-dimensional grid of voxels.

[0541] 19. Computer-readable media in which, in item 18, the reference allele amino acid is a three-dimensional representation of the one-hot encoding of the reference amino acid experiencing the variant amino acid.

[0542] 20. Computer-readable media in which, in Item 1, amino acid-specific conservation frequencies specify the conservation levels of each amino acid across multiple species.

[0543] 21. In Item 20, the operations are the operation of selecting the nearest atom to a corresponding voxel across reference amino acid and atomic categories,

[0544] An operation to select pan-amino acid conservation frequencies for a residue of a reference amino acid containing the nearest atom, and

[0545] Computer-readable media further comprising an operation using a three-dimensional representation of pan-amino acid conservation frequencies as an evolutionary conservation channel.

[0546] 22. Computer-readable media in which, in Item 21, pan-amino acid conservation frequencies are configured for specific positions of residues as observed in multiple species.

[0547] 23. Computer-readable media in which, in Item 21, pan-amino acid preservation frequencies specify whether there are missing preservation frequencies for a specific reference amino acid.

[0548] 24. In Item 21, the operations are the operations of selecting the respective nearest atoms to the corresponding voxels within each of the reference amino acids,

[0549] The operation of selecting conservation frequencies per amino acid for each residue of a reference amino acid containing the nearest atoms, and

[0550] Computer-readable media further comprising an operation using a three-dimensional representation of amino acid conserved frequencies as an evolutionary conservation channel.

[0551] 25. Computer-readable media in which, in Item 24, the conservation frequencies per amino acid are configured for specific positions of residues as observed in multiple species.

[0552] 26. Computer-readable media in which, in Item 24, the retention frequencies per amino acid specify whether there is a missing retention frequency for a specific reference amino acid.

[0553] 27. In Item 1, the operations further include the operation of encoding one or more annotation channels on a voxel-by-voxel basis for each voxel within a three-dimensional grid of voxels, wherein the annotation channels are three-dimensional representations of one-hot encoding of residual annotations, computer-readable media.

[0554] 28. In item 27, the annotation channels are computer-readable media that are molecular processing annotations including initiator methionine, signal, change peptide, propeptide, chain, and peptide.

[0555] 29. Computer-readable media, wherein the annotation channels in Item 27 are domain annotations including topology domains, transmembrane, intramembrane, domains, repeats, calcium binding, zinc clamps, deoxyribonucleic acid (DNA) binding, nucleotide binding, regions, double coils, motifs, and compositional biases.

[0556] 30. Computer-readable media, wherein the annotation channels in Item 27 are site annotations including an active site, a metal bond, a bonding site, and a site.

[0557] 31. Computer-readable media in which, in item 27, the annotation channels are amino acid modified annotations including non-standard residues, modified residues, lipidations, glycosylations, disulfide bonds, and crosslinks.

[0558] 32. Computer-readable media in which, in item 27, the annotation channels are secondary structure annotations including helicals, spins, and beta strands.

[0559] 33. Computer-readable media in which, in item 27, the annotation channels are experimental information annotations including mutagenic, sequence uncertainty, sequence conflict, non-adjacent residues, and non-terminal residues.

[0560] 34. In Item 1, the operations further include the operation of encoding one or more structural confidence channels on a voxel-by-voxel basis for each voxel within a three-dimensional grid of voxels, wherein the structural confidence channels are three-dimensional representations of confidence scores specifying the quality of their respective residual structures, computer-readable media.

[0561] 35. In Item 34, the structural reliability channels are computer-readable media that are Global Model Quality Estimates (GMQEs).

[0562] 36. In Item 34, the structural reliability channels are computer-readable media that are Qualitative Model Energy Analysis (QMEAN) scores.

[0563] 37. In Item 34, the structure reliability channels are computer-readable media that are temperature factors specifying the extent to which residues satisfy the physical constraints of their respective protein structures.

[0564] 38. Computer-readable media, wherein, in Item 34, the structural reliability channels are template structure alignments that specify the degree to which the residues of atoms closest to the voxels have aligned template structures.

[0565] 39. In Item 38, computer-readable media in which structural reliability channels are mold modeling scores of aligned mold structures.

[0566] 40. In Item 39, the structural reliability channels are computer-readable media that are the minimum mold modeling scores, the average of the mold modeling scores, and the maximum mold modeling score among the mold modeling scores.

[0567] 41. Computer-readable media, wherein, in Item 1, the operations further include the operation of rotating atoms before amino acid-specific distance channels are formed.

[0568] 42. Computer-readable media, wherein, in Item 1, the operations further include operations using 1x1x1 convolutions, 3x3x3 convolutions, rectified linear unit activation layers, batch normalization layers, full connection layers, dropout regularization layers, and softmax classification layers in a convolutional neural network.

[0569] 43. Computer-readable media in which, in Item 42, 1x1x1 convolutions and 3x3x3 convolutions are three-dimensional convolutions.

[0570] 44. Computer-readable media according to Item 42, wherein a layer of 1x1x1 convolutions processes a tensor and generates an intermediate output which is a convolutional representation of the tensor, a sequence of layers of 3x3x3 convolutions processes the intermediate output and generates a flattened output, a fully connected layer processes the flattened output and generates denormalized outputs, and a softmax classification layer processes the denormalized outputs and generates exponentially normalized outputs which identify the likelihood that variant nucleotides are pathogenic and benign.

[0571] 45. Computer-readable media in which, in item 44, the sigmoid layer processes denormalized outputs and generates normalized outputs that identify the likelihood that variant nucleotides are pathogenic.

[0572] 46. ​​Computer-readable media in which, in Item 1, the convolutional neural network is an attention-based neural network.

[0573] 47. Computer-readable media of Item 1, wherein the tensor comprises amino acid-specific distance channels additionally encoded as a reference allele channel.

[0574] 48. Computer-readable media of Item 1, wherein the tensor comprises amino acid-specific distance channels additionally encoded as annotation channels.

[0575] 49. Computer-readable media of Item 1, wherein the tensor comprises amino acid-specific distance channels additionally encoded as structural confidence channels.

[0576] 50. One or more computer-readable media storing computer-executable instructions, wherein the computer-executable instructions, when executed on one or more processors,

[0577] An operation of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate distance channels by atomic category,

[0578] Atoms span multiple atomic categories, and

[0579] Atomic categories within multiple atomic categories specify the atomic elements of amino acids, and

[0580] Each of the atomic category distance channels has a 3D distance value for each voxel within the 3D grid of voxels, and

[0581] An operation to generate a 3D distance value that specifies the distance from a corresponding voxel within a 3D grid of voxels to atoms of corresponding atomic categories within a plurality of atomic categories;

[0582] An operation of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0583] An operation of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across atomic category-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel;

[0584] An operation of applying three-dimensional convolutions to a tensor comprising atomic category-specific distance channels encoded with alternative allele channels and their respective evolutionarily conserved channels; and

[0585] One or more computer-readable media configured to perform operations including operations to determine the pathogenicity of variant nucleotides based at least partially on a tensor.

[0586] 51. One or more computer-readable media storing computer-executable instructions, wherein the computer-executable instructions, when executed on one or more processors,

[0587] An operation of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate amino acid-specific distance channels,

[0588] Each amino acid-specific distance channel has a 3D distance value for each voxel within the 3D grid of voxels, and

[0589] An operation of generating a three-dimensional distance value that specifies the distance from a corresponding voxel within a three-dimensional grid of voxels to atoms of a corresponding reference amino acid within a reference amino acid sequence;

[0590] As an operation of encoding an alternative allele channel for each 3D distance value in each amino acid-specific distance channel at the amino acid position level,

[0591] An alternative allele channel is an encoding operation that is a three-dimensional representation of the one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0592] An operation of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across amino acid-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel; and

[0593] One or more computer-readable media configured to perform operations including generating a tensor comprising alternative allele channels and amino acid-specific distance channels encoded with their respective evolutionarily conserved channels.

[0594] 52. One or more computer-readable media storing computer-executable instructions, wherein the computer-executable instructions, when executed on one or more processors,

[0595] An operation of accessing the three-dimensional structure of a reference amino acid sequence of a protein and fitting a three-dimensional grid of voxels onto atoms within the three-dimensional structure at the amino acid level to generate distance channels by atomic category,

[0596] Atoms span multiple atomic categories, and

[0597] Atomic categories within multiple atomic categories specify the atomic elements of amino acids, and

[0598] Each of the atomic category distance channels has a 3D distance value for each voxel within the 3D grid of voxels, and

[0599] An operation to generate a 3D distance value that specifies the distance from a corresponding voxel within a 3D grid of voxels to atoms of corresponding atomic categories within a plurality of atomic categories;

[0600] An operation of encoding an alternative allele channel in each voxel within a three-dimensional grid of voxels, wherein the alternative allele channel is a three-dimensional representation of a one-hot encoding of a variant amino acid expressed by a variant nucleotide;

[0601] An operation of encoding an evolutionary conservation channel in each sequence of three-dimensional distance values ​​across atomic category-specific distance channels at the voxel position level, wherein the evolutionary conservation channel is a three-dimensional representation of amino acid-specific conservation frequencies across multiple species, and the amino acid-specific conservation frequencies are selected according to amino acid proximity to the corresponding voxel; and

[0602] One or more computer-readable media configured to perform operations including generating a tensor comprising alternative allele channels and atomic category distance channels encoded with their respective evolutionary conservation channels.

[0603] Other embodiments of the method described in this section may include a non-transient computer-readable storage medium that stores instructions executable by a processor to perform any of the methods described above. Another embodiment of the method described in this section may include a system comprising a memory and one or more processors—operable to perform any of the methods described above by executing instructions stored in said memory.

[0604] Specific implementation examples 3

[0605] Item 3

[0606] 1. A computer implementation method for efficiently determining sequence elements closest to non-uniformly spaced cells within a grid, wherein elements have element coordinates and cells have dimensional cell indices and cell coordinates, and a step of generating an element-cell mapping that maps a subset of cells to each element, wherein the subset of cells mapped to a specific element in the sequence includes the nearest cell within the grid and one or more neighboring cells within the grid, and

[0607] The nearest cell is selected based on matching the element coordinates of a specific element to the cell coordinates, and

[0608] A step of generating an element-cell mapping in which neighboring cells are selected based on being closely adjacent to the nearest cell and within a distance proximity range from a specific element; a step of generating a cell-element mapping in which a subset of elements is mapped to each of the cells, wherein the subset of elements mapped to a specific cell in the grid includes such elements in the sequence mapped to the specific cell by the element-cell mapping; and

[0609] The method includes the step of determining the nearest element within the sequence for each cell using cell-element mapping,

[0610] A computer-implemented method in which the closest element to a specific cell is determined based on the distances between the specific cell and elements within a subset of elements.

[0611] 2. A computer implementation method according to Item 1, wherein matching the element coordinates of a specific element to cell coordinates further comprises truncating the decimal part of the element coordinates to generate truncated element coordinates.

[0612] 3. In Item 2, matching the element coordinates of a specific element to cell coordinates is,

[0613] For the first dimension, matching the first truncated element coordinates within the truncated element coordinates to the first cell coordinates of the first cell within the grid, and selecting the first dimension index of the first cell;

[0614] For the second dimension, matching the second clipped element coordinates within the clipped element coordinates to the second cell coordinates of the second cell within the grid, and selecting the second dimension index of the second cell;

[0615] For the third dimension, matching the third truncated element coordinates within the truncated element coordinates to the third cell coordinates of the third cell within the grid, and selecting the third dimension index of the third cell;

[0616] Generating a cumulative sum based on positionally weighting the first, second, and third dimensional indices selected by squares of the radix using the selected first, second, and third dimensional indices; and

[0617] A computer implementation method further comprising using the accumulated sum as a cell index for selecting the nearest cell.

[0618] 4. A computer implementation method of Item 1, wherein distances are calculated between the cell coordinates of a specific cell and the element coordinates of elements within a subset of elements.

[0619] 5. A computer implementation method in which, in Item 1, the sequence is a protein sequence of amino acids.

[0620] 6. A computer implementation method in which, in Item 5, the elements are atoms of amino acids.

[0621] 7. In Item 6, the steps of generating an element-cell mapping, generating a cell-element mapping, and determining the closest element for each of the cells using the cell-element mapping have a runtime complexity of O(a * f + v), and

[0622] a is the number of atoms, and

[0623] f is the number of amino acids, and

[0624] v is the number of cells, and

[0625] * is a computer implementation method that is a multiplication operation.

[0626] 8. A computer implementation method in which, in Item 7, the atoms include alpha-carbon atoms.

[0627] 9. A computer implementation method in which, in Item 7, the atoms include beta-carbon atoms.

[0628] 10. A computer implementation method in which, in Item 7, the atoms include non-carbon atoms.

[0629] 11. A computer implementation method of Item 1, wherein the cells are three-dimensional voxels.

[0630] 12. A computer implementation method in which, in Item 11, the cell coordinates are three-dimensional coordinates.

[0631] 13. A computer implementation method of Item 12 in which the element coordinates are three-dimensional coordinates.

[0632] 14. A computer implementation method of Item 1, wherein neighboring cells are selected based on being within an index proximity range from the nearest cell.

[0633] 15. A computer implementation method of Item 1, wherein neighboring cells are selected based on being within cell neighbors within a grid that includes the nearest cell.

[0634] 16. A computer implementation method of Item 1, wherein the sequence comprises M elements, a subset of the elements comprises N elements, and M>> N.

[0635] 17. A computer implementation method for efficiently determining which atoms within a protein are closest to voxels within a grid, wherein atoms have three-dimensional (3D) atomic coordinates, voxels have 3D voxel coordinates, and the computer implementation method comprises:

[0636] A step of generating an atomic-voxel mapping that maps selected containing voxels to each of the atoms based on matching the 3D atomic coordinates of a specific atom of the protein to the 3D voxel coordinates within a grid;

[0637] A computer-implemented method comprising: generating a voxel-atom mapping that maps a subset of atoms to each of the voxels, wherein the subset of atoms mapped to a specific voxel in the grid includes such atoms in the protein that are mapped to the specific voxel by the atom-voxel mapping; and using the voxel-atom mapping to determine the nearest atom in the protein for each of the voxels.

[0638] 18. In Item 17, the steps of Item 17 are a computer implementation method having a runtime complexity of O(number of atoms).

[0639] Other embodiments of the method described in this section may include a non-transient computer-readable storage medium that stores instructions executable by a processor to perform any of the methods described above. Another embodiment of the method described in this section may include a system comprising a memory and one or more processors—operable to perform any of the methods described above by executing instructions stored in said memory.

[0640] Although the present invention is disclosed with reference to the preferred embodiments and examples described above, it will be understood that these examples are intended to be illustrative rather than restrictive. Modifications and combinations will readily come to mind of those skilled in the art, and it is considered that such modifications and combinations will be within the spirit of the invention and the scope of the following claims.

Claims

Claim 1 A system comprising: a memory for storing amino acid-wise distance channels for a plurality of amino acids in an amino acid sequence of a protein, wherein each of the amino acid-wise distance channels has a voxel-wise distance value for a voxel within a plurality of voxels, and the voxel-wise distance value (voxel-wise distance value) specifies the distance from a corresponding voxel within the plurality of voxels to an atom of a corresponding amino acid within the plurality of amino acids; and a neural network-based variant pathogenicity classifier executed on at least one processor coupled to the memory, wherein the neural network-based variant pathogenicity classifier is trained to process as input a tensor comprising the amino acid-wise distance channels and alternative allele amino acids of the protein expressed by the variant, and to classify the variant as positive or pathogenic based at least partially on said tensor. Claim 2 A system according to claim 1, further comprising a distance channel generator executed on at least one processor coupled to the memory, wherein the generator centers a voxel grid of the voxels on the alpha carbon atom of each corresponding amino acid residue, calculates a distance value per voxel, and the distance value per voxel specifies the distance between the center of the voxels within the voxel grid and the atoms of the corresponding amino acids. Claim 3 In paragraph 2, the distance channel generator centers the voxel grid on the alpha-carbon atom of a specific amino acid residue corresponding to at least one variant amino acid in the protein. Claim 4 A system according to paragraph 3, further configured to encode the directionality of the corresponding amino acid and the position of the specific amino acid by multiplying the voxel-wise distance value for the preceding amino acid preceding the specific amino acid in the tensor by a directionality parameter. Claim 5 In paragraph 3, the voxel-specific distance value specifies the nearest atomic distance from the center of the corresponding voxel within the voxel grid to the nearest atom of the corresponding amino acid, in a system. Claim 6 A system according to claim 5, wherein the corresponding amino acid has alpha-carbon atoms, and the voxel-specific distance value specifies the nearest alpha-carbon atom distance from the center of the corresponding voxel to the nearest alpha-carbon atoms of the corresponding amino acid. Claim 7 A system according to claim 5, wherein the corresponding amino acid has beta-carbon atoms, and the voxel-specific distance value specifies the distance of the nearest beta-carbon atoms from the center of the corresponding voxel to the nearest beta-carbon atoms of the corresponding amino acid. Claim 8 A system according to claim 5, wherein the corresponding amino acid has backbone atoms, and the voxel-specific distance value specifies the distance of the nearest backbone atoms from the center of the corresponding voxel to the nearest backbone atoms of the corresponding amino acid. Claim 9 A system according to paragraph 3, further configured to encode a nearest atom channel in the tensor specifying the distance from each voxel to the nearest atom, wherein the nearest atom is selected regardless of the plurality of amino acids and the atomic elements of the plurality of amino acids. Claim 10 A system according to claim 1, wherein the tensor further comprises an absentee atomic channel that specifies atoms not found within a predefined radius of the voxel center, and the absentee atomic channel is one-hot encoded. Claim 11 A system according to claim 1, wherein the tensor further comprises one-hot encoding of the alternative allele amino acid encoded voxel by each amino acid-specific distance channel. Claim 12 A system according to claim 11, wherein the tensor further comprises (i) a one-hot encoding of a reference allele amino acid corresponding to the alternative allele amino acid and (ii) voxel-wise encoded for each of the amino acid-wise distance channels. Claim 13 A system according to claim 1, wherein the tensor further comprises an evolutionary profile that specifies the conservation level of the corresponding amino acid across a plurality of species having a sequence homologous to the amino acid sequence of the protein. Claim 14 A system according to claim 13, further comprising an evolutionary profile generator executed on at least one processor coupled to the memory, wherein the evolutionary profile generator determines a pan-amino acid conservation frequency using multiple sequence alignment for each of the voxels, selects the nearest atom across the plurality of amino acid and atomic categories, selects a pan-amino acid conservation frequency sequence for a residue of the amino acid containing the nearest atom, voxels the pan-amino acid conservation frequency for the residue of the amino acid, and makes the pan-amino acid conservation frequency sequence available as one of the evolutionary profiles. Claim 15 In claim 14, the above-mentioned pan-amino acid conservation frequency sequence is configured for a specific position of the residue as observed in the above-mentioned plurality of species, in a system. Claim 16 A system according to claim 14, wherein the evolutionary profile generator determines the conservation frequency per amino acid using the multiple sequence alignment for each of the voxels, selects the respective nearest atom within each of the plurality of amino acids, selects the respective conservation frequency per amino acid for each residue of the plurality of amino acids containing the respective nearest atom, voxels the conservation frequency per amino acid for each residue of the plurality of amino acids, and makes the conservation frequency per amino acid available as one of the evolutionary profiles. Claim 17 In paragraph 16, the above amino acid retention frequency is configured for specific positions of each residue observed in multiple species, in a system. Claim 18 A system according to claim 1, wherein the tensor comprises: a comment channel for the corresponding amino acid that annotates the characteristics of the corresponding amino acid—the comment channel being one-hot encoded in the tensor—; or one or more of a structure confidence channel for the corresponding amino acid that specifies the quality of the respective structure of the corresponding amino acid. Claim 19 A system comprising: a memory storing atomic category distance channels for amino acids within an amino acid sequence of a protein, wherein the amino acid has atoms for a plurality of atomic categories, the atomic categories within the plurality of atomic categories specify atomic elements of the amino acid, each atomic category distance channel has a voxel-wise distance value for a voxel within a plurality of voxels, and the voxel-wise distance value specifies a distance from a corresponding voxel within the plurality of voxels to an atom of a corresponding atomic category within the plurality of atomic categories; and a neural network-based variant pathogenicity classifier executed on at least one processor coupled to the memory, wherein the variant pathogenicity classifier is trained to process as input a tensor comprising the atomic category distance channels and alternative allele amino acids of the protein expressed by the variant, and to classify the variant as positive or pathogenic based at least partially on said tensor. Claim 20 A computer implementation method comprising: a step of storing amino acid-specific distance channels for a plurality of amino acids within an amino acid sequence of a protein, wherein each of the amino acid-specific distance channels has a voxel-specific distance value for a voxel within a plurality of voxels, and the voxel-specific distance value specifies the distance from a corresponding voxel within the plurality of voxels to an atom of a corresponding amino acid within the plurality of amino acids; a step of processing a tensor comprising the amino acid-specific distance channels and alternative allele amino acids of the protein expressed by the variant as input to a neural network-based variant pathogenicity classifier; and a step of classifying the variant as positive or pathogenic based at least partially on the tensor. Claim 21 delete Claim 22 delete Claim 23 delete Claim 24 delete Claim 25 delete Claim 26 delete Claim 27 delete Claim 28 delete Claim 29 delete Claim 30 delete

Images