Combination learning and transfer learning of variant pathogenicity predictors using both gapped and non-gapped protein samples

Abstract

The technology disclosed relates to training a pathogenicity predictor. In particular, the technology disclosed relates to accessing a gapped training set comprising respective gapped protein samples for respective positions in a proteome, accessing a non-gapped training set comprising non-gapped benign protein samples and non-gapped pathogenic protein samples, generating respective gapped spatial representations of the gapped protein samples, and generating respective non-gapped spatial representations of the non-gapped benign protein samples and the non-gapped pathogenic protein samples, training a pathogenicity predictor and generating a trained pathogenicity predictor in one or more training cycles, wherein each of the training cycles uses a gapped spatial representation from the respective gapped spatial representations and a non-gapped spatial representation from the respective non-gapped spatial representations as training examples, and using the trained pathogenicity classifier to determine pathogenicity of a variant.

Description

[0001] Priority application

[0002] This patent application claims priority and benefit from the following U.S. application, which is incorporated by reference for all purposes.

[0003] U.S. non-provisional patent application No. 17 / 533,091 (Attorney's File No. ILLM 1050-2 / IP-2164-US), filed November 22, 2021, entitled "Protein Structure-Based Protein Language Models," claims priority to U.S. provisional patent application No. 63 / 253,122 (Attorney's File No. ILLM 1050-1 / IP-2164-PRV), filed October 6, 2021; U.S. provisional patent application No. 63 / 281,579 (Attorney's File No. ILLM 1060-1 / IP-2270-PRV), filed November 19, 2021; and U.S. provisional patent application No. 63 / 281,592 (Attorney's File No. ILLM 1061-1 / IP-2271-PRV), filed November 19, 2021; and

[0004] U.S. non-provisional patent application No. 17 / 953,286 (Attorney's File No. ILLM 1060-2 / IP-2270-US), filed on September 26, 2022, entitled "Predicting Variant Pathogenicity From Evolutionary Conservation Using Three-Dimensional (3D) Protein Structure Voxels," claims against U.S. provisional patent applications No. 63 / 253,122 (Attorney's File No. ILLM 1050-1 / IP-2164-PRV), filed on October 6, 2021; U.S. provisional patent applications No. 63 / 281,579 (Attorney's File No. ILLM 1060-1 / IP-2270-PRV), filed on November 19, 2021; and U.S. provisional patent applications No. 63 / 281,592 (Attorney's File No. ILLM), filed on November 19, 2021. Priority of 1061-1 / IP-2271-PRV; and

[0005] U.S. non-provisional patent application No. 17 / 953,293 (Attorney's File No. ILLM 1061-2 / IP-2271-US), filed on September 26, 2022, entitled "Combined And Transfer Learning of a Variant Pathogenicity Predictor Using Gapped and Non-Gapped Protein Samples," claims against U.S. provisional patent application No. 63 / 253,122 (Attorney's File No. ILLM 1050-1 / IP-2164-PRV), filed on October 6, 2021, and U.S. provisional patent application No. 63 / 281,579 (Attorney's File No. ILLM), filed on November 19, 2021. Priority claims to U.S. Provisional Patent Application No. 63 / 281,592 (Attorney’s File No. ILLM1061-1 / IP-2270-PRV), filed November 19, 2021. Technical Field

[0006] The technologies disclosed in this invention relate to artificial intelligence-type computers and digital data processing systems, as well as corresponding data processing methods and products for simulating intelligence (i.e., knowledge-based systems, inference systems, and knowledge acquisition systems); and include systems for uncertainty inference (e.g., fuzzy logic systems), adaptive systems, machine learning systems, and artificial neural networks. Specifically, the technologies disclosed in this invention relate to using deep convolutional neural networks to analyze multichannel voxelized data.

[0007] Literature merged

[0008] The following references are incorporated by way of citation, i.e., as shown in their entirety in this document, for all purposes:

[0009] Sundaram, L. et al., Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018);

[0010] Jaganathan,K. et al., Predicting splicing from primary sequence with deep learning. Cell 176,535-548(2019);

[0011] U.S. Patent Application No. 62 / 573,144 (Attorney’s File No. ILLM 1000-1 / IP-1611-PRV), filed on October 16, 2017, entitled “Training a Deep Pathogenicity Classifier Using Large-Scale Benign Training Data”;

[0012] U.S. Patent Application No. 62 / 573,149 (Attorney’s File No. ILLM 1000-2 / IP-1612-PRV), filed on October 16, 2017, entitled “Pathogenicity Classifier Based on DeepConvolutional Neural Networks (CNNs)”;

[0013] U.S. Patent Application No. 62 / 573,153 (Attorney’s File No. ILLM 1000-3 / IP-1613-PRV), filed on October 16, 2017, entitled “Deep Semi-Supervised Learning That Generates Large-Scale Pathogenic Training Data”;

[0014] U.S. Patent Application No. 62 / 582,898, filed on November 7, 2017, entitled "Pathogenicity Classification of Genomic Data Using Deep Convolutional Neural Networks (CNNs)" (Attorney's File No. ILLM 1000-4 / IP-1618-).

[0015] PRV);

[0016] U.S. Patent Application No. 16 / 160,903 (Attorney’s File No. ILLM 1000-5 / IP-1611-US), entitled “Deep Learning-Based Techniques for Training Deep Convolutional Neural Networks”, filed on October 15, 2018;

[0017] U.S. Patent Application No. 16 / 160,986 (Attorney’s File No. ILLM 1000-6 / IP-1612-US), entitled “Deep Convolutional Neural Networks for Variant Classification”, filed on October 15, 2018;

[0018] U.S. Patent Application No. 16 / 160,968 (Attorney’s File No. ILLM 1000-7 / IP-1613-US), filed on October 15, 2018, entitled “Semi-Supervised Learning for Training an Ensemble of Deep Convolutional Neural Networks”;

[0019] U.S. Patent Application No. 16 / 407,149 (Attorney’s File No. ILLM 1010-1 / IP-1734-US), filed on May 8, 2019, entitled “Deep Learning-Based Techniques for PreTraining Deep Convolutional Neural Networks”;

[0020] U.S. Patent Application No. 17 / 232,056, entitled “Deep Convolutional Neural Networks to Predict Variant Pathogenicity Using Three-Dimensional (3d) Protein Structures”, filed on April 15, 2021 (Attorney’s File No. ILLM 1037-2 / IP-2051-US);

[0021] U.S. Patent Application No. 63 / 175,495 (Attorney’s File No. ILLM 1047-1 / IP-2142-PRV), filed on April 15, 2021, entitled “Multi-Channel Protein Voxelization to Predict Variant Pathogenicity Using Deep Convolutional Neural Networks”;

[0022] U.S. Patent Application No. 63 / 175,767 (Attorney's File No. ILLM 1048-1 / IP-2143-PRV), filed on April 16, 2021, entitled "Efficient Voxelization for Deep Learning"; and

[0023] Priority is claimed in U.S. Patent Application No. 17 / 468,411 (Attorney’s File No. ILLM 1037-3 / IP-2051A-US), filed on September 7, 2021, entitled “Artificial Intelligence-Based Analysis of Protein Three-Dimensional (3d) Structures”. Background Technology

[0024] The topics discussed in this section should not be considered prior art simply because they are mentioned here. Similarly, problems mentioned in this section or related to the topics provided as background art should not be assumed to have been previously recognized in the prior art. The topics in this section merely represent different methods, which themselves may correspond to specific implementations of the technology protected by the claims.

[0025] Genomics, also broadly known as functional genomics, aims to characterize the function of every genomic element in an organism using genome-scale assays such as genome sequencing, transcriptome profiling, and proteomics. Genomics emerges as a data-driven science—it operates by discovering new properties from the exploration of genome-scale data rather than by testing preconceived models and hypotheses. Applications of genomics include discovering associations between genotypes and phenotypes, identifying biomarkers for patient stratification, predicting gene function, and mapping biochemically active genomic regions, such as transcriptional enhancers.

[0026] Genomics data is too large and complex to be mined simply by visualizing pairwise correlations. Instead, analytical tools are needed to support the discovery of unexpected relationships, leading to new hypotheses and models, and making predictions. Unlike some algorithms where hypotheses and domain expertise are hard-coded, machine learning algorithms are designed to automatically detect patterns in the data. Therefore, machine learning algorithms are well-suited for data-driven science, especially genomics. However, the performance of machine learning algorithms can be highly dependent on how the data is represented—that is, how each variable (also called a feature) is computed. For example, to classify tumors as malignant or benign from fluorescence microscopy images, preprocessing algorithms might detect cells, identify cell types, and generate a list of cell counts for each cell type.

[0027] Machine learning models can classify tumors using estimated cell counts (instances of handcrafted features) as input. A key issue is that classification performance heavily depends on the quality and relevance of these features. For example, relevant visual features (such as cell morphology, intercellular distances, or intra-organ location) may not be captured in cell counts, and this incomplete representation of the data can reduce classification accuracy.

[0028] Deep learning (a branch of machine learning) addresses this problem by embedding feature computation into the machine learning model itself to produce end-to-end models. This has been achieved through the development of deep neural networks, which are machine learning models that include sequential basic operations, where increasingly complex features are computed by taking the results of earlier operations as input. Deep neural networks can improve prediction accuracy by discovering highly complex, relevant features (such as cell morphology and spatial organization in the examples above). The construction and training of deep neural networks have become possible thanks to the explosion of data, advancements in algorithms, and a significant increase in computing power, particularly through the use of graphics processing units (GPUs).

[0029] The goal of supervised learning is to obtain a model that takes features as input and returns a prediction of a so-called target variable. An example of a supervised learning problem is predicting whether introns have been spliced ​​off on a given (target) feature of RNA, such as the presence of a typical splice site sequence, the location of the splice branch point, or the intron length. Training a machine learning model refers to learning its parameters, which typically involves minimizing a loss function on the training data with the aim of making accurate predictions on unseen data.

[0030] For many supervised learning problems in computational biology, input data can be represented as a table with multiple columns or features, each containing potentially numerical or categorical data that can be used to make predictions. Some input data are naturally represented as features in a table (such as temperature or time), while others require transformation using a process called feature extraction (such as transforming a DNA sequence into k-mer counts) to conform to a tabular representation. For intron splicing prediction problems, the presence of typical splice site sequences, the location of splice branch points, and intron lengths can be preprocessed features collected in tabular format. Tabular data is standard for a wide variety 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.

[0031] Logistic regression is a binary classifier, i.e., a supervised learning model that predicts a binary target variable. Specifically, logistic regression predicts the probability of the positive class by calculating a weighted sum of input features mapped to the [0,1] interval using a sigmoid function (a type of activation function). The parameters of logistic regression, or other linear classifiers using different activation functions, are the weights in the weighted sum. A linear classifier fails when the weighted sum of input features cannot distinguish classes well (e.g., the class of introns that have been spliced ​​or not). To improve prediction performance, new input features can be manually added by transforming or combining existing features in new ways (e.g., by exponentiation or pairwise multiplication).

[0032] Neural networks use hidden layers to automatically learn these nonlinear feature transformations. Each hidden layer can be thought of as multiple linear models whose outputs are transformed by a nonlinear activation function, such as the sigmoid function or the more popular rectified linear unit function (ReLU). Together, these layers compose the input features into relevant complex patterns, which helps in tasks that distinguish between two classes.

[0033] Deep neural networks use many hidden layers, one of which is called a fully connected layer, where each neuron receives input from all neurons in the previous layer. Neural networks are typically trained using stochastic gradient descent, an algorithm well-suited for training models on very large datasets. Implementing neural networks using modern deep learning frameworks enables rapid prototyping with different architectures and datasets. Fully connected neural networks can be used in many genomics applications, including predicting the percentage of exons spliced ​​against a given sequence from sequence features such as the presence of splicing factor binding motifs or sequence conservation; ranking potentially pathogenic genetic variants by importance; and predicting cis-regulatory elements in a given genomic region using features such as chromatin markers, gene expression, and evolutionary conservation.

[0034] For effective prediction, local dependencies in spatial and longitudinal data must be considered. For example, scrambling DNA sequences or pixels in an image can severely disrupt information patterns. These local dependencies set spatial or longitudinal data beyond tabular data, where feature ordering is arbitrary. Consider classifying genomic regions as either bound by or not by specific transcription factors, where binding regions are defined as high-confidence binding events in chromatin immunoprecipitation followed by sequencing (ChIP-seq) data. Transcription factors bind to DNA by recognizing sequence motifs. Fully connected layers based on sequence-derived features, such as the number of k-mer instances or position-weighted matrix (PWM) matching within a sequence, can be used for this task. Because k-mer or PWM instance frequencies are robust to motif shifting within a sequence, such models generalize well to sequences with the same motif located at different positions. However, they fail to recognize patterns where transcription factor binding depends on combinations of multiple motifs with well-defined intervals. Furthermore, the number of possible k-mers increases exponentially with k-mer length, posing challenges in terms of storage and overfitting.

[0035] Convolutional layers are a special form of fully connected layer where the same fully connected layer is applied locally (e.g., within a 6bp window) to all sequence locations. This approach can also be viewed as using multiple PWMs to scan the sequence, for example, for transcription factors GATA1 and TAL1. By using the same model parameters at different locations, the total number of parameters is drastically reduced, and the network is able to detect motifs at locations not seen during training. Each convolutional layer scans the sequence with several filters by producing a scalar value at each location, quantifying the match between the filter and the sequence. As in fully connected neural networks, a non-linear activation function (typically ReLU) is applied at each layer. Next, a pooling operation is applied, which gathers the activations into consecutive bins across the entire location axis, typically taking the maximum or average activation per channel. Pooling reduces the effective sequence length and coarsens the signal. Subsequent convolutional layers combine the output of the previous layer and are able to detect the presence of GATA1 and TAL1 motifs within a certain distance range. Finally, the outputs of these convolutional layers can be used as input to a fully connected neural network to perform the final prediction task. Therefore, different types of neural network layers (e.g., fully connected layers and convolutional layers) can be combined within a single neural network.

[0036] Convolutional Neural Networks (CNNs) are capable of predicting a wide range of 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. Beyond predicting molecular phenotypes from sequences, CNNs can be applied to many more technical tasks traditionally addressed by hand-crafted bioinformatics pipelines. For example, CNNs can predict guide RNA specificity, denoise ChIP-seq, improve Hi-C data resolution, predict the laboratory of origin from DNA sequences, and detect genetic variants. CNNs have also been used to model long-range dependencies in the genome. Although interacting regulatory elements may be geographically distant on unfolded linear DNA sequences, these elements are typically close to each other in the actual 3D chromatin conformation. Therefore, while modeling molecular phenotypes from linear DNA sequences is a rough approximation of chromatin, it can be improved by allowing long-range dependencies and allowing the model to implicitly learn aspects of 3D organization, such as promoter-enhancer looping. This is achieved by using expanded convolutions, which have a receptive field of up to 32 kb. Expanded convolutions also allow for the prediction of splice sites from sequences using a receptive field of 10 kb, thus enabling the integration of genetic sequences across distances as long as those of typical human introns (see Jaganathan, K. et al., Predicting splicing from primary sequence with deep learning. Cell 176, 535-548 (2019)).

[0037] Different types of neural networks can be characterized by their parameter-sharing schemes. For example, fully connected layers do not have parameter sharing, while convolutional layers impose translation invariance by applying the same filter 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) that implement different parameter-sharing schemes. RNNs apply the same operation to each sequence element. This operation takes the previous sequence element and the new input as inputs to a memory. The operation updates the memory and optionally emits an output, which is passed to subsequent layers or used directly as a model prediction. Because the same model is applied at each sequence element, RNNs remain invariant to positional indices within the processed sequence. For example, a RNN can detect open reading frames in a DNA sequence regardless of their position within the sequence. This task requires identifying specific sequences of inputs, such as a start codon followed by an in-frame stop codon.

[0038] The main advantage of recurrent neural networks (RNNs) over convolutional neural networks (CNNs) lies in their theoretical ability to carry information through infinitely long sequences via memory. Furthermore, RNNs can naturally handle sequences of varying lengths, such as mRNA sequences. However, in sequence modeling tasks (such as audio synthesis and machine translation), CNNs combined with various techniques (such as dilated convolutions) can achieve comparable or even better performance than RNNs. RNNs can aggregate the outputs of CNNs to predict single-cell DNA methylation status, RBP binding, transcription factor binding, and DNA accessibility. Moreover, because RNNs apply sequential operations, they cannot be easily parallelized, resulting in significantly slower computation speeds compared to CNNs.

[0039] While each person has a unique genetic code, the majority of the human genetic code is common to all. In some cases, the human genetic code can include outliers, called genetic variants, which may be common to individuals within a relatively small group of people. For example, a particular human protein may contain a specific amino acid sequence, while a variant of that protein may differ by one amino acid in the same specific sequence.

[0040] Genetic variants can be pathogenic, leading to disease. Although most such variants have been depleted from the genome through natural selection, the ability to identify which variants are likely 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 genetic variants remains unclear. Some of the most common pathogenic variants are single nucleotide missense mutations that alter the amino acids in proteins. However, not all missense mutations are pathogenic.

[0041] Models that can directly predict molecular phenotypes from biological sequences can be used as computational perturbation tools to probe the association between genetic and phenotypic variations and have become novel approaches for quantitative trait locus identification and variant prioritization. These approaches are crucial because most variants identified through genome-wide association studies of complex phenotypes are non-coding, making it challenging to estimate their role and contribution to the phenotype. Furthermore, linkage disequilibrium leads to the co-heritage of variant blocks, which creates difficulties in identifying individual causal variants. Therefore, sequence-based deep learning models, which can be used as probing tools to assess the impact of such variants, offer a promising approach to discovering potential drivers of complex phenotypes. One example involves indirectly predicting the effects of non-coding single nucleotide variants and short insertions or deletions (indels) from differences between two variants in transcription factor binding, chromatin accessibility, or gene expression prediction. Another example includes predicting the generation of new splicing sites based on sequence or the quantitative effect of genetic variants on splicing.

[0042] An end-to-end deep learning approach for predicting variant effects was applied to predict the pathogenicity of missense variants from protein sequence and sequence conservation data (see Sundaram, L. et al., Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018), referred to here as “PrimateAI”). PrimateAI uses a deep neural network trained on variants known to be pathogenic, with data augmentation using cross-species information. Specifically, PrimateAI uses sequences of wild-type and mutant proteins to compare differences and uses a trained deep neural network to determine the pathogenicity of the mutation. This approach of using protein sequences for pathogenicity prediction is promising because it avoids the roundness problem and overfitting to prior knowledge. However, the amount of clinical data available in ClinVar is relatively small compared to the sufficient amount of data for effectively training deep neural networks. To overcome this data scarcity, PrimateAI uses common human and primate variants as benign data, while simulated variants based on trinucleotide backgrounds are used as unlabeled data.

[0043] When trained directly from sequence alignments, PrimateAI outperforms existing methods. PrimateAI learns important protein domains, conserved amino acid positions, and sequence dependencies directly from training data consisting of approximately 120,000 human samples. PrimateAI significantly surpasses other variant pathogenicity prediction tools in distinguishing between benign and pathogenic de novo mutations in candidate developmental disorder genes, and in replicating prior knowledge in ClinVar. These results indicate that PrimateAI represents a significant advancement in variant classification tools, reducing the reliance on prior knowledge in clinical reporting.

[0044] At the heart of protein biology is understanding how structural elements produce observed functions. The abundance of protein structural data enables the development of computational methods to systematically derive the rules governing structure-function relationships. However, the performance of these methods depends heavily on the choice of protein structural representation.

[0045] A protein site is a microenvironment within a protein structure, distinguished by its structural or functional role. A site can be defined by its three-dimensional (3D) location and the surrounding local neighborhood where its structure or function resides. At the heart of rational protein engineering is understanding how the structural arrangement of amino acids produces functional characteristics within a protein site. Determining the structural and functional roles of individual amino acids within a protein provides information that facilitates the engineering and modification of protein function. Identifying functionally or structurally important amino acids allows for focused engineering efforts, such as site-directed mutagenesis to alter the functional properties of target proteins. Alternatively, this knowledge can help avoid engineering designs that could compromise the desired function.

[0046] Because structures are now far more conserved than sequences, the increase in protein structure data provides an opportunity to systematically study the underlying patterns governing structure-function relationships using data-driven approaches. A fundamental aspect of any protein computational analysis is how protein structural 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 produce noisy distributions lacking underlying patterns.

[0047] Protein structure redundancy and the recent success of deep learning algorithms have provided an opportunity to develop tools for automatically extracting task-specific representations of protein structures. Therefore, there is an opportunity to use multichannel voxelized representations of 3D protein structures as input to deep neural networks to predict the pathogenicity of variants. Attached Figure Description

[0048] In the accompanying drawings, similar reference numerals generally refer to similar parts in all different views. Furthermore, the drawings are not necessarily drawn to scale, but rather emphasize the principles of the disclosed technology. In the following description, various embodiments of the technology disclosed in this invention are described with reference to the following drawings, wherein:

[0049] Figure 1 This is a flowchart illustrating the process of a system for determining the pathogenicity of variants according to various specific embodiments of the technology disclosed in this invention.

[0050] Figure 2 An example reference amino acid sequence of a protein and an alternative amino acid sequence of the protein are illustrated schematically according to a specific embodiment of the technology disclosed in this invention.

[0051] Figure 3 This illustrates a specific implementation of the technology disclosed in this invention. Figure 2 The amino acid formula classification of amino acid atoms in the reference amino acid sequence.

[0052] Figure 4 This illustrates a specific implementation of the technology disclosed in this invention. Figure 3 The amino acid formula is assigned based on the 3D atomic coordinates of the α-carbon atom in the classification of amino acids.

[0053] Figure 5 The process of determining voxel-based distance values ​​according to a specific embodiment of the technology disclosed in this invention is illustrated schematically.

[0054] Figure 6 An example of 21 amino acid-based distance channels is shown as a specific embodiment of the technology disclosed in this invention.

[0055] Figure 7 This is a schematic diagram of a distance channel tensor according to a specific implementation of the technology disclosed in this invention.

[0056] Figure 8 This illustrates a specific implementation of the technology disclosed in this invention, from Figure 2 The unique heat encoding of reference amino acids and alternative amino acids.

[0057] Figure 9 This is a schematic diagram of a reference amino acid encoded by voxelization and a variant / alternative amino acid encoded by voxelization, according to a specific embodiment of the technology disclosed in this invention.

[0058] Figure 10 The diagram schematically illustrates a linking process according to a specific embodiment of the technology disclosed in this invention, which links components in a voxel-like manner. Figure 7 The distance channel tensor is linked to the reference allele tensor.

[0059] Figure 11 The diagram schematically illustrates a linking process according to a specific embodiment of the technology disclosed in this invention, which links components in a voxel-like manner. Figure 7 Distance channel tensor, Figure 10 The reference allele tensor and the alternative allele tensor are linked.

[0060] Figure 12 This is a flowchart illustrating a specific embodiment of the technology disclosed in this invention, a process for determining and distributing the pan-amino acid conserved frequencies (voxolization) of the atoms closest to the voxel.

[0061] Figure 13 A voxel-to-the-nearest amino acid mapping is shown as a specific embodiment of the technology disclosed in this invention.

[0062] Figure 14 An example multiple sequence alignment of reference amino acid sequences across 99 species is shown as a specific embodiment of the technology disclosed in this invention.

[0063] Figure 15An example of determining the pan-amino acid conserved frequency sequence of a specific voxel is shown, according to a specific embodiment of the technology disclosed in this invention.

[0064] Figure 16 This illustrates a specific implementation of the technology disclosed in this invention, using... Figure 15 The position-frequency logic described herein is for the corresponding pan-amino acid conserved frequencies determined for the corresponding voxels.

[0065] Figure 17 A voxelized per-voxel evolution spectrum is shown as a specific embodiment of the technology disclosed in this invention.

[0066] Figure 18 An example of an evolutionary spectrum tensor according to a specific implementation of the technology disclosed in this invention is described.

[0067] Figure 19 This is a flowchart illustrating a specific embodiment of the technology disclosed in this invention, a process for determining and distributing the conserved frequency (voxolization) of each amino acid relative to the atom closest to the voxel.

[0068] Figure 20 Various examples of voxelized annotation channels linked to distance channel tensors are shown as a specific implementation of the technology disclosed in this invention.

[0069] Figure 21 Different combinations and arrangements of input channels according to a specific embodiment of the technology disclosed in this invention are shown, which can be provided as input to a pathogenicity classifier for determining the pathogenicity of a target variant.

[0070] Figure 22 Different methods for calculating the distance channel disclosed in this invention are shown in various specific embodiments of the technology disclosed in this invention.

[0071] Figure 23 Different examples of evolutionary pathways are shown, illustrating various specific implementations of the technology disclosed in this invention.

[0072] Figure 24 Different examples of annotation channels are shown, illustrating various specific implementations of the technology disclosed in this invention.

[0073] Figure 25 Different examples of structural confidence channels according to various specific implementations of the technology disclosed in this invention are shown.

[0074] Figure 26 An example processing architecture for a pathogenicity classifier according to a specific implementation of the technology disclosed in this invention is shown.

[0075] Figure 27An example processing architecture for a pathogenicity classifier according to a specific implementation of the technology disclosed in this invention is shown.

[0076] Figure 28 , Figure 29 , Figure 30 , Figure 31A and Figure 31B PrimateAI was used as a benchmark model to demonstrate the classification advantages of PrimateAI 3D over PrimateAI disclosed in this invention.

[0077] Figure 32A and Figure 32B The efficient voxelization process disclosed in this invention is illustrated in various specific embodiments of the technology disclosed in this invention.

[0078] Figure 33 This describes a specific embodiment of the technology disclosed in the present invention, showing how atoms are associated with voxels containing atoms.

[0079] Figure 34 A specific embodiment of the technology disclosed in this invention is shown, which generates a voxel-to-atom mapping from an atom to a voxel to identify the nearest atom on a voxel-by-voxel basis.

[0080] Figure 35A and Figure 35B This demonstrates how the efficient voxelization disclosed in this invention has a runtime complexity of O(#atoms), compared to the runtime complexity of O(#atoms*#voxels) without using the efficient voxelization disclosed in this invention.

[0081] Figure 36 It is an example computer system that can be used to implement the technology disclosed in this invention.

[0082] Figure 37 This paper demonstrates a specific implementation of determining the pathogenicity of variants of target alternative amino acids based on processing the spatial representation of nicked proteins.

[0083] Figure 38 An example of a spatial representation of a protein is shown.

[0084] Figure 39 It shows Figure 38 An example of a nicked space representation of a protein is shown.

[0085] Figure 40 It shows Figure 38 The image shows an example of the atomic space representation of a protein.

[0086] Figure 41 It shows Figure 38 Examples of nicked atomic spatial representations of proteins are shown.

[0087] Figure 42 An implementation of a pathogenicity classifier is shown, which determines the pathogenicity of variants of a target alternative amino acid based on the nicked protein spatial representation and the alternative amino acid representation processed by the target alternative amino acid.

[0088] Figure 43 A specific implementation of training data used to train a pathogenicity classifier is described.

[0089] Figure 44 A specific implementation of generating a nicked space representation of a reference protein sample by using reference amino acids as nicked amino acids is demonstrated.

[0090] Figure 45 A specific implementation of training a pathogenicity classifier on benign protein samples is shown.

[0091] Figure 46 A specific implementation of training a pathogenicity classifier on a pathogenic protein sample is shown.

[0092] Figure 47 This demonstrates how to mask certain unreachable amino acids during training.

[0093] Figure 48 A specific implementation for determining the final pathogenicity score is shown.

[0094] Figure 49A The study demonstrates the determination of variant pathogenicity of target substitute amino acids that fill vacancies created by reference gap amino acids at a given position in a protein.

[0095] Figure 49B The pathogenicity of corresponding variants of amino acids that fill vacancies created by reference gap amino acids at given positions in proteins is shown.

[0096] Figure 50 This paper demonstrates a specific implementation of determining the pathogenicity of variants of multiple alternative amino acids based on processing the spatial representation of nicked proteins.

[0097] Figure 51 An implementation of a pathogenicity classifier is shown, which determines the pathogenicity of variants of multiple alternative amino acids based on processing the spatial representation of a nicked protein.

[0098] Figure 52 A specific implementation of training a pathogenicity classifier simultaneously on both benign and pathogenic protein samples is demonstrated.

[0099] Figure 53This paper presents a specific implementation for determining the pathogenicity of variants of multiple alternative amino acids based on processing the spatial representation of a nicked protein and generating an evolutionary conservation score of multiple alternative amino acids in response.

[0100] Figure 54 An evolutionary conservation determiner is shown based on an operational implementation.

[0101] Figure 55 This demonstrates a specific implementation of determining pathogenicity based on predicted evolutionary scores.

[0102] Figure 56 A specific implementation of training data used to train an evolutionary conservation determiner is shown.

[0103] Figure 57 A specific implementation of training an evolutionary conservation determinant on both benign and pathogenic protein samples is demonstrated.

[0104] Figure 58 Different specific implementations of the benchmark truth tag encoding used for training evolutionary conservatism determiners are described.

[0105] Figure 59 An example location-specific frequency matrix (PSFM) is shown.

[0106] Figure 60 An example location-specific scoring matrix (PSSM) is depicted.

[0107] Figure 61 A specific implementation for generating PSFM and PSSM is shown.

[0108] Figure 62 Example PSFM encoding is shown.

[0109] Figure 63 Example PSSM encoding is depicted.

[0110] Figure 64 Two datasets are shown on which the models disclosed in this paper can be trained.

[0111] Figures 65A to 65B This paper demonstrates a specific implementation of the combinatorial learning of the model disclosed in this paper.

[0112] Figures 66A to 66B Demonstrates the use of transfer learning and Figure 64 The two datasets shown are used to train a specific implementation of the model disclosed in this paper.

[0113] Figure 67 A specific implementation for generating training data and labels to train the model disclosed herein is shown.

[0114] Figure 68 A specific implementation of a method for determining the pathogenicity of nucleotide variants is demonstrated.

[0115] Figure 69 A specific implementation of a system for predicting the structural tolerance of amino acid substitutes is demonstrated.

[0116] Figure 70A , Figure 70B and Figure 70C Performance results are described using objective metrics that demonstrate non-obviousness and creativity. Detailed Implementation

[0117] The following discussion is presented to enable any person skilled in the art to implement and use the disclosed technology, and is provided in the context of a particular application and its requirements. Various modifications to the specific embodiments disclosed will be apparent to those skilled in the art, and the general principles defined herein can be applied to other specific embodiments and applications without departing from the spirit and scope of the disclosed technology. Therefore, the disclosed technology is not intended to be limited to the specific embodiments shown, but is to be accorded the broadest scope consistent with the principles and features disclosed herein.

[0118] The detailed descriptions of the various embodiments will be better understood when read in conjunction with the accompanying drawings. With regard to the diagrams illustrating the functional blocks of the various embodiments, the functional blocks do not necessarily indicate the division between hardware circuits. Therefore, for example, one or more functional blocks (e.g., modules, processors, or memories) may be implemented in a single piece of hardware (e.g., a general-purpose signal processor or a block of random access memory, a hard disk, etc.) or in multiple pieces of hardware. Similarly, a program may be a standalone program, may be incorporated as a subroutine into an operating system, may be a function within an installed software package, etc. It should be understood that the various embodiments are not limited to the arrangements and tools shown in the accompanying drawings.

[0119] The processing engine and database designated as modules in the accompanying figures can be implemented in hardware or software and do not need to be precisely divided into identical blocks as shown in the figures. Some of these modules may also be implemented on different processors, computers, or servers, or distributed among multiple different processors, computers, or servers. Furthermore, it should be understood that some modules may be combined, operated synchronously, or operated in a different sequence than shown in the figures, without affecting the implemented functionality. The modules in the accompanying figures can also be considered as flowchart steps in a method. A module does not necessarily need to place all its code contiguously in memory; some parts of the code may be separated from other parts of the code, with code from other modules or other functions positioned between them.

[0120] Pathogenicity assays based on protein structure

[0121] Figure 1 This is a flowchart illustrating process 100 of a system for determining the pathogenicity of variants. At step 102, the system's sequence accessor 104 accesses a reference amino acid sequence and a substituted amino acid sequence. At step 112, the system's 3D structure generator 114 generates a 3D protein structure of the reference amino acid sequence. In some embodiments, the 3D protein structure is a homology model of a human protein. In one embodiment, the so-called SwissModel homology modeling pipeline provides a public knowledge base of predicted human protein structures. In another embodiment, the so-called HHpred homology modeling uses a tool called Modeller to predict the structure of a target protein from a template structure.

[0122] Proteins are represented by sets of atoms and their coordinates in 3D space. Amino acids can have various types of atoms, such as carbon, oxygen (O), nitrogen (N), and hydrogen (H). These atoms can be further classified into side-chain atoms and main-chain atoms. The main-chain carbon atom can include α-carbon (C0). α ) atoms and β-carbon (C β )atom.

[0123] At step 122, the system's coordinate classifier 124 classifies the 3D atomic coordinates of the 3D protein structure based on amino acids. In one specific implementation, amino acid classification involves assigning the 3D atomic coordinates to 21 amino acid categories (including terminating amino acid categories or gap amino acid categories). In one example, the amino acid classification of α-carbon atoms could list the α-carbon atoms under each of the 21 amino acid categories. In another example, the amino acid classification of β-carbon atoms could list the β-carbon atoms under each of the 21 amino acid categories.

[0124] In yet another example, the amino acid classification for oxygen atoms can be listed separately for each of the 21 amino acid categories. In yet another example, the amino acid classification for nitrogen atoms can be listed separately for each of the 21 amino acid categories. In yet another example, the amino acid classification for hydrogen atoms can be listed separately for each of the 21 amino acid categories.

[0125] Those skilled in the art will recognize that, in various specific implementations, the amino acid classification can include subsets of 21 amino acid categories and subsets of different atomic elements.

[0126] At step 132, the system's voxel mesh generator 134 instantiates a voxel mesh. The voxel mesh can have any resolution, such as 3×3×3, 5×5×5, 7×7×7, etc. The voxels in the voxel mesh can have any size, for example, 1 angstrom per side. Each side Each side And so on. Those skilled in the art will understand that these example dimensions refer to cubic dimensions, since a voxel is a cube. Furthermore, those skilled in the art will understand that these example dimensions are not limiting, and a voxel can have any cubic dimension.

[0127] At step 142, the system's voxel grid center locator 144 centers the voxel grid at a reference amino acid of the target variant at the amino acid level. In one embodiment, the voxel grid center is located at the atomic coordinates of a specific atom of the reference amino acid of the target variant, for example, at the 3D atomic coordinates of the α-carbon atom of the reference amino acid of the target variant.

[0128] Distance Channel

[0129] Voxels in a voxel grid can have multiple channels (or features). In one specific implementation, voxels in a voxel grid have multiple distance channels (e.g., 21 distance channels for each of the 21 amino acid categories, including termination or nick amino acid categories). At step 152, the system's distance channel generator 154 generates amino acid-based distance channels for the voxels in the voxel grid. Distance channels are generated independently for each of these 21 amino acid categories.

[0130] For example, considering the alanine (A) amino acid category, and further considering, for example, a voxel grid of size 3×3×3 with 27 voxels. Then, in one specific embodiment, the alanine distance channel comprises 27 distance values ​​for each of the 27 voxels in the voxel grid. The 27 distance values ​​in the alanine distance channel are measured from the corresponding center of each of the 27 voxels in the voxel grid to the corresponding nearest atom in the alanine amino acid category.

[0131] In one example, the alanine amino acid class includes only α-carbon atoms, so the closest atoms are those alanine α-carbon atoms that are closest to the 27 voxels in the voxel grid, respectively. In another example, the alanine amino acid class includes only β-carbon atoms, so the closest atoms are those alanine β-carbon atoms that are closest to the 27 voxels in the voxel grid, respectively.

[0132] In yet another example, the alanine amino acid class includes only oxygen atoms, so the closest atoms are the alanine oxygen atoms that are closest to each of the 27 voxels in the voxel grid. In yet another example, the alanine amino acid class includes only nitrogen atoms, so the closest atoms are the alanine nitrogen atoms that are closest to each of the 27 voxels in the voxel grid. In yet another example, the alanine amino acid class includes only hydrogen atoms, so the closest atoms are the alanine hydrogen atoms that are closest to each of the 27 voxels in the voxel grid.

[0133] Similar to the alanine distance channel, the distance channel generator 154 generates distance channels (i.e., a set of voxel-based distance values) for each of the remaining amino acid categories. In other embodiments, the distance channel generator 154 generates distance channels only for a subset of the 21 amino acid categories.

[0134] In other specific implementations, the selection of the nearest atom is not limited to a specific atom type. That is, within the subject amino acid category, the atom closest to a specific voxel is selected, regardless of the atomic element of the closest atom, and the distance value of the specific voxel is calculated to include it in the distance channel of the subject amino acid category.

[0135] In other specific implementations, distance channels are generated based on atomic elements. As an alternative or supplement to the distance channels used for amino acid categories, distance values ​​can be generated for atomic element categories, regardless of the amino acid to which the atom belongs.

[0136] For example, consider an amino acid sequence in which the atoms span seven atomic elements: carbon, oxygen, nitrogen, hydrogen, calcium, iodine, and sulfur. The voxels in the voxel grid are then configured with seven distance channels, each of which has twenty-seven voxel-based distance values ​​specifying the distance to the nearest atom only within the corresponding atomic element class. In other embodiments, distance channels can be generated for a subset of the seven atomic elements. In still other embodiments, the atomic element classes and distance channels can be further hierarchically stratified into variants of the same atomic element, for example, α-carbon (C1). α ) atoms and β-carbon (C β )atom.

[0137] In other specific implementations, distance channels can be generated based on atom type, for example, distance channels for side-chain atoms only and distance channels for main-chain atoms only.

[0138] A predefined maximum scan radius (e.g., six angstroms) can be used at the voxel center. The search function finds the nearest atom within the voxel grid. Furthermore, multiple atoms can be found to be closest to the same voxel within the voxel grid.

[0139] The distance between the 3D coordinates of the voxel center and the 3D atomic coordinates of the atom is calculated. Additionally, a distance channel is generated using a voxel grid centered at the same location (e.g., at the 3D atomic coordinates of the α-carbon atom of the reference amino acid undergoing the target variant).

[0140] These distances can be Euclidean distances. Alternatively, they can be parameterized by atomic size (or atomic influence) (e.g., by using the Leonard-Jones potential and / or van der Waals atomic radius of the atoms under consideration). Furthermore, these distance values ​​can be normalized by the maximum scan radius, or by the maximum observed distance value of the nearest atom within a subject amino acid category, subject atom element category, or subject atom type category. In some embodiments, the distance between a voxel and an atom is calculated based on the polar coordinates of the voxel and the atom. The polar coordinates are parameterized by the angle between the voxel and the atom. In one embodiment, this angle information is used to generate the voxel's angle channel (i.e., independent of the distance channel). In some embodiments, the angle between the nearest atom and its neighboring atom (e.g., main chain atoms) can be used as a feature utilizing voxel encoding.

[0141] Reference allele pathway and alternative allele pathway

[0142] Voxels in the voxel grid can also have reference allele channels and alternative allele channels. At step 162, the system's one-hot encoder 164 generates a reference one-hot code for a reference amino acid in the reference amino acid sequence and an alternative one-hot code for an alternative amino acid in the alternative amino acid sequence. The reference amino acid undergoes a target variant. The alternative amino acid is the target variant. The reference amino acid and the alternative amino acid are located at the same positions 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. This exception is the position where the reference amino acid is present in the reference amino acid sequence and the alternative amino acid is present in the alternative amino acid sequence.

[0143] At step 172, the system's connector 174 connects the amino acid-type distance channel to the reference one-hot encoding and the alternative one-hot encoding. In another embodiment, connector 174 connects the atomic element-type distance channel to the reference one-hot encoding and the alternative one-hot encoding. In yet another embodiment, connector 174 connects the atomic type-type distance channel to the reference one-hot encoding and the alternative one-hot encoding.

[0144] At step 182, the system's runtime logic 184 processes the linked amino acid-based distance channels / atomic element-based distance channels / atomic type-based distance channels with reference one-hot encoding and alternative one-hot encoding via a pathogenicity classifier (pathogenicity determination engine) to determine the pathogenicity of the target variant. This pathogenicity is then inferred as the pathogenicity determination result of the underlying nucleotide variant that produces the target variant at the amino acid level. The pathogenicity classifier is trained using labeled datasets of benign and pathogenic variants (e.g., using a backpropagation algorithm). Additional details regarding the labeled datasets of benign and pathogenic variants, as well as example architectures and training of the pathogenicity classifier, can be found in commonly owned U.S. Patent Applications Nos. 16 / 160,903, 16 / 160,986, 16 / 160,968, and 16 / 407,149.

[0145] Figure 2 The reference amino acid sequence 202 and the alternative amino acid sequence 212 of protein 200 are schematically shown. Protein 200 contains N amino acids. The positions of the amino acids in protein 200 are labeled 1, 2, 3…N. In the example shown, position 16 is the position that undergoes amino acid variant 214 (mutation) caused by a variant of the base nucleotide. For example, for 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 sequence 202) has the reference amino acid leucine (L). Although not shown for clarity, the remaining positions in 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 variant 214 at position 16, which contains the alternative amino acid alanine (A) 214 instead of the reference amino acid glycine (G) 204.

[0146] Figure 3 The amino acid classification of the amino acid atoms in reference amino acid sequence 202 is shown, also referred to herein as "Atomic Classification 300". Of the 20 naturally occurring amino acids listed in column 302, certain types of amino acids can appear repeatedly in proteins. That is, a certain type of amino acid can appear more than once in a protein. Proteins may also have some unidentified amino acids, which are classified according to 21 terminating or nicked amino acid categories. Figure 3 The right column contains α-carbons (C60) from different amino acids. α ) The number of atoms.

[0147] Specifically, Figure 3 The α-carbon (C) of the amino acid in reference amino acid sequence 202 is shown. α The amino acid classification of atoms. Figure 3 Column 308 lists the total number of α-carbon atoms observed for reference amino acid sequence 202 under each of the 21 amino acid categories. For example, column 308 lists 11 α-carbon atoms observed for the alanine (A) amino acid category. Since each amino acid has only one α-carbon atom, this means that alanine appears 11 times in reference amino acid sequence 202. In another example, arginine (R) appears 35 times in reference amino acid sequence 202. The total number of α-carbon atoms across the 21 amino acid categories is 828.

[0148] Figure 4 Showing based on Figure 3 The amino acid classification 300 refers to the 3D atomic coordinates of the α-carbon atom in the amino acid sequence 202. This is referred to herein as "atomic coordinate grouping aggregation 400". Figure 4 In the table, lists 404 to 440 list the 3D atomic coordinates of the α-carbon atoms in each of the 21 amino acid categories.

[0149] In the specific implementation shown, Figure 4 Grouping aggregation 400 follows Figure 3 Category 300. For example, in Figure 3 In this context, alanine, an amino acid, has 11 α-carbon atoms; therefore, in... Figure 4 In this context, the alanine amino acid class has [derived from] Figure 3 This corresponds to the 11 3D atomic coordinates of the 11 α-carbon atoms. For other amino acid categories, this classification to grouping aggregation logic also changes from... Figure 3 Flow direction Figure 4 However, this classification-to-grouping aggregation logic is for representative purposes only. In other embodiments, the technology disclosed in this invention does not require performing classification 300 and grouping aggregation 400 to locate the nearest voxel-form atom, and fewer, additional, or different steps can be performed. For example, in some embodiments, the technology disclosed in this invention can locate the nearest voxel-form atom by using a sorting and search algorithm that returns the nearest voxel-form atom from one or more databases in response to a search query configured to accept query parameters such as sorting criteria (e.g., amino acid formula, element formula, atom type formula), a predefined maximum scan radius, and distance type (e.g., Euclidean, Markovnikov, normalized, non-normalized). In various embodiments of the technology disclosed in this invention, those skilled in the art can similarly use a variety of sorting and search algorithms from current or future technologies to locate the nearest voxel-form atom.

[0150] exist Figure 4In this context, 3D atomic coordinates are represented by Cartesian coordinates x, y, and z, but any type of coordinate system can be used, such as spherical or cylindrical coordinates, and the subject matter for which protection is sought is not limited in this respect. In some specific implementations, one or more databases may include information on the 3D atomic coordinates of α-carbon atoms and other amino acid atoms in proteins. Such databases can be searched by specific proteins.

[0151] As discussed above, voxels and voxel meshes are 3D entities. However, for clarity, the accompanying drawings depict and the description discusses voxels and voxel meshes in a two-dimensional (2D) format. For example, a 3×3×3 voxel mesh of 27 voxels is depicted and described herein as a 3×3 2D pixel mesh with 9 2D pixels. Those skilled in the art will appreciate that the 2D format is used for representative purposes only and is intended to cover the 3D counterpart (i.e., 2D pixels represent 3D voxels, and 2D pixel meshes represent 3D voxel meshes). Furthermore, the drawings are not drawn to scale. For example, a single pixel is used to depict a size of 2 angstroms. The voxels.

[0152] Voxel distance calculation

[0153] Figure 5 The process for determining voxel-based distance values, also referred to herein as "Voxel-based Distance Calculation 500," is illustrated schematically. In the example shown, voxel-based distance values ​​are calculated only for the alanine (A) distance channel. However, the same distance calculation logic is performed for each of the 21 amino acid categories to generate 21 amino acid-based distance channels, and this can be further extended to other atom types, such as β-carbon atoms and other atomic elements, such as oxygen, nitrogen, and hydrogen, as discussed above. Figure 1 As discussed, in some specific implementations, atoms are randomly rotated before distance calculation so that the training of the pathogenicity classifier remains unchanged in terms of atomic orientation.

[0154] exist Figure 5 In the voxel grid 522, there are 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 center of voxel grid 522 is located at, for example, position 16 in reference amino acid sequence 202, at the 3D atomic coordinate 532 of the α-carbon atom of the glycine (G) amino acid, because in the alternative amino acid sequence 212, position 16 undergoes a variant that mutates the glycine (G) amino acid to the alanine (A) amino acid, as described above regarding... Figure 2 As discussed above. Additionally, the center of voxel mesh 522 coincides with the center of voxel (2,2).

[0155] A centered voxel grid 522 is used for voxel distance calculations for each of the 21 amino acid distance channels. For example, starting with the alanine (A) distance channel, the distances between the 3D coordinates of the corresponding centers of the nine voxels 14 and the 3D atomic coordinates 402 of the 11 alanine α-carbon atoms are measured to locate the nearest alanine α-carbon atom for each of the nine voxels 514. Then, the alanine distance channel is constructed using the nine distance values ​​between the nine voxels 514 and their respective nearest alanine α-carbon atoms. The resulting alanine distance channel arranges the nine alanine distance values ​​in the same order as the nine voxels 514 in the voxel grid 522.

[0156] The above process is performed for each of the 21 amino acid categories. For example, the centered voxel grid 522 is similarly used to calculate the arginine (R) distance channels, such that the distance between the 3D coordinates of the corresponding centers of the 9 voxels 514 and the 3D atomic coordinates 404 of the 35 arginine α-carbon atoms is measured to locate the nearest arginine α-carbon atom in each of the 9 voxels 514. Then, the arginine distance channels are constructed using the 9 distance values ​​between the 9 voxels 514 and their respective nearest arginine α-carbon atoms. The resulting arginine distance channels arrange the 9 arginine distance values ​​in the same order as the 9 voxels 514 in the voxel grid 522. The 21 amino acid-based distance channels are encoded in a voxel fashion to form a distance channel tensor.

[0157] Specifically, in the example shown, a distance of 512 is at the center of voxel (1,1) in voxel grid 522, and is closest to the α-carbon (C). α Between the atoms, the α-carbon atom is C in List 402. α A5 Atoms. Therefore, the value assigned to voxel (1,1) is a distance of 512. In another example, C α A4 The atom is the C closest to the center of voxel (1,2). α Atoms. Therefore, the value assigned to voxel (1,2) is the center of voxel (1,2) and C. α A4 The distance between atoms. In yet another example, C α A6 The atom is the C closest to the center of the voxel (2,1). α Atoms. Therefore, the value assigned to voxel (2,1) is the center of voxel (2,1) and C. α A6 The distance between atoms. In yet another example, C α A6 The atom is also the C atom closest to the center of voxels (3,2) and (3,3). αAtoms. Therefore, the value assigned to voxel (3,2) is the center of voxel (3,2) and C. α A6 The distance between atoms, and the value assigned to voxel (3,3) is the distance between the center of voxel (3,3) and C. α A6 Distances between atoms. In some implementations, the distance value assigned to voxel 514 can be a normalized distance. For example, the distance value assigned to voxel (1,1) can be distance 512 divided by a maximum distance 502 (a predefined maximum scan radius). In some implementations, the nearest atom distance can be a Euclidean distance, and the nearest atom distance can be normalized by dividing the Euclidean distance by the maximum nearest atom distance (e.g., such as maximum distance 502).

[0158] As described above, for amino acids having an α-carbon atom, the distance can be the distance from the center of the corresponding voxel to the nearest α-carbon atom of the corresponding amino acid. Similarly, for amino acids having a β-carbon atom, the distance can be the distance from the center of the corresponding voxel to the nearest β-carbon atom of the corresponding amino acid. Likewise, for amino acids having a main chain atom, the distance can be the distance from the center of the corresponding voxel to the nearest main chain atom of the corresponding amino acid. Similarly, for amino acids having a side chain atom, the distance can be the distance from the center of the corresponding voxel to the nearest side chain atom of the corresponding amino acid. In some specific embodiments, the distance may, in addition to / alternatively, include distances to a second, third, fourth, or other nearby atoms.

[0159] Amino acid distance channels

[0160] Figure 6 An example of 21 amino acid-based distance channels 600 is shown. Figure 6 Each column in the table corresponds to one of the 21 amino acid-based distance channels 602 to 642. Each amino acid-based distance channel includes the distance value of each voxel in voxel 514 of voxel grid 522. For example, the amino acid-based distance channel 602 for alanine (A) includes the distance value of the corresponding voxel in voxel 514 of voxel grid 522. As mentioned above, voxel grid 522 is a 3D grid with a volume of 3×3×3, comprising 27 voxels. Similarly, although... Figure 6 Voxel 514 is displayed in a two-dimensional manner (e.g., 9 voxels in a 3×3 grid), but each amino acid-based distance channel can include 27 voxel-based distance values ​​in a 3×3×3 voxel grid.

[0161] Directional coding

[0162] In some embodiments, the technique disclosed in this invention uses directional parameters to specify the directionality of reference amino acids in reference amino acid sequence 202. In some embodiments, the technique disclosed in this invention uses directional parameters to specify the directionality of alternative amino acids in alternative amino acid sequence 212. In some embodiments, the technique disclosed in this invention uses directional parameters to specify the location in protein 200 that undergoes a target variant at the amino acid level.

[0163] As described above, all distance values ​​in the 21 amino acid distance channels 602 to 642 are measured from the corresponding nearest atom in voxel grid 522 to voxel 514. These nearest atoms originate from a reference amino acid in reference amino acid sequence 202. These originating reference amino acids containing the nearest atom can be divided into two categories: (1) those originating reference amino acids preceding reference amino acid 204 that has undergone a variant in reference amino acid sequence 202, and (2) those originating reference amino acids following reference amino acid 204 that has undergone a variant in reference amino acid sequence 202. The originating reference amino acids in the first category can be referred to as the prior reference amino acids. The originating reference amino acids in the second category can be referred to as the subsequent reference amino acids.

[0164] The directional parameter is applied to the distance values ​​in the 21 amino acid-based distance channels 602 to 642, which are measured from the closest atoms originating from the prior reference amino acid. In one embodiment, the directional parameter is multiplied by these distance values. The directional parameter can be any number, such as -1.

[0165] As a result of the application of directional parameters, these 21 amino acid-based distance channels 600 include distance values ​​that indicate to the pathogenicity classifier which end of protein 200 is the start and which end is the end. This also allows the pathogenicity classifier to reconstruct the protein sequence from the 3D protein structure information provided by the distance channels, as well as the reference and allele channels.

[0166] Distance Channel Tensor

[0167] Figure 7 This is a schematic diagram of the distance-channel tensor 700. The distance-channel tensor 700 comes from... Figure 6 The amino acid-based distance channels 600 are represented in voxel format. In the distance channel tensor 700, like the RGB channels of a color image, the 21 amino acid-based distance channels 602 to 642 are connected in a voxel fashion. The voxelization dimension of the distance channel tensor 700 is 21×3×3×3 (where 21 represents the 21 amino acid classes, and 3×3×3 represents a 3D voxel grid with 27 voxels); however... Figure 7 It is a 2D depiction with dimensions of 21×3×3.

[0168] One-hot encoding

[0169] Figure 8 The unique heat codes 800 for reference amino acid 204 and alternative amino acid 214 are shown. Figure 8 In the table, the left column is the unique heat code 802 for the reference amino acid glycine (G)204, where 1 indicates the glycine amino acid class and 0 indicates all other amino acid classes. Figure 8 In the table, the right column is the unique heat code 804 for the variant / alternative amino acid alanine (A)214, where 1 indicates the alanine amino acid class and 0 indicates all other amino acid classes.

[0170] Figure 9 This is a schematic diagram of the voxelized one-heat encoded reference amino acid 902 and the voxelized one-heat encoded variant / substitute amino acid 912. The voxelized one-heat encoded reference amino acid 902 is derived from... Figure 8 The voxelized representation of the uniquely encoded amino acid glycine (G)204, 802. The voxelized uniquely encoded alternative amino acid 912 is derived from... Figure 8 The voxelized representation of the one-hot encoded amino acid alanine (A)214, 804. The voxelized one-hot encoded reference amino acid 902 has a voxelized dimension of 21×1×1×1 (where 21 represents 21 amino acid classes); but Figure 9 It is a 2D depiction with dimensions 21×1×1. Similarly, the voxelization dimension of the uniquely encoded alternative amino acid 912 is 21×1×1×1 (where 21 represents 21 amino acid classes); but Figure 9 It is a 2D depiction with dimensions of 21×1×1.

[0171] Reference allele tensor

[0172] Figure 10 This schematically demonstrates how to arrange data in a voxel-based manner. Figure 7 The connection process 1000 connects the distance channel tensor 700 to the reference allele tensor 1004. The reference allele tensor 1004 is derived from... Figure 9 The voxel-like aggregation (repetition / cloning / replication) of the voxelized one-hot encoded reference amino acid 902. That is, multiple copies of the voxelized one-hot encoded reference amino acid 902 are linked together in a voxel manner according to the spatial arrangement of voxels 514 in the voxel grid 522, such that the reference allele tensor 1004 has a corresponding copy of the voxelized one-hot encoded reference amino acid 910 for each voxel in the voxel grid 522.

[0173] The linking process 1000 generates a linking tensor 1010. The voxelization dimension of the reference allele tensor 1004 is 21×3×3×3 (where 21 represents 21 amino acid classes and 3×3×3 represents a 3D voxel grid with 27 voxels); but Figure 10 It is a 2D representation of the reference allele tensor 1004 with dimensions 21×3×3. The voxelized dimension of the link tensor 1010 is 42×3×3×3; but Figure 10 It is a 2D representation of a connection tensor of dimension 42×3×3 and a capacity of 1010.

[0174] Alternative allele tensor

[0175] Figure 11 This schematically demonstrates how to arrange data in a voxel-based manner. Figure 7 Distance channel tensor 700, Figure 10 The linking process 1100 involves connecting the reference allele tensor 1004 to the alternative allele tensor 1104. The alternative allele tensor 1104 is derived from... Figure 9 The voxel-like aggregation (repetition / cloning / replication) of the voxelized one-hot encoded alternative amino acid 912. That is, multiple copies of the voxelized one-hot encoded alternative amino acid 912 are linked together in a voxel manner according to the spatial arrangement of voxels 514 in voxel grid 522, such that the alternative allele tensor 1104 has a corresponding copy of the voxelized one-hot encoded alternative amino acid 910 for each voxel in voxel 514 in voxel grid 522.

[0176] The linking process 1100 produces the linking tensor 1110. The voxelization dimension of the alternative allele tensor 1104 is 21×3×3×3 (where 21 represents 21 amino acid classes and 3×3×3 represents a 3D voxel grid with 27 voxels); but Figure 11 It is a 2D representation of the alternative allele tensor 1104 with dimensions 21×3×3. The voxelized dimension of the linking tensor 1110 is 63×3×3×3; but Figure 11 It is a 2D representation of a connection tensor of dimension 63×3×3 and a capacity of 1110.

[0177] In some specific implementations, runtime logic 184 processes the link tensor 1110 through a pathogenicity classifier to determine the pathogenicity of the variant / alternative amino acid alanine (A) 214, which is then inferred as the pathogenicity determination result of the underlying nucleotide variant that produces the variant / alternative amino acid alanine (A) 214.

[0178] Evolutionary Conservative Channel

[0179] Predicting the functional consequences of variants relies at least in part on the assumption that key amino acids in a protein family are conserved during evolution due to negative selection (i.e., amino acid changes at these sites have been harmful in the past), and that mutations at these sites increase the likelihood of pathogenicity (causing disease) in humans. Generally, homologous sequences of the target protein are collected and aligned, and a measure of conservation is calculated based on the weighted frequencies of different amino acids observed at the target position in the alignment.

[0180] Therefore, the technique disclosed in this invention links the distance channel tensor 700, the reference allele tensor 1004, and the alternative allele tensor 1004 to evolutionary channels. One example of an evolutionary channel is the pan-amino acid conserved frequency. Another example of an evolutionary channel is the per-amino acid conserved frequency.

[0181] In some implementations, position-weighted matrices (PWM) are used to construct evolutionary channels. In others, position-specific frequency matrices (PSFM) are used. Still others employ computational tools like SIFT, PolyPhen, and PANTHER-PSEC. And still others are conservation-based channels built upon evolutionary conservation. Conservation is associated with conservation because it also reflects the role of negative selection, which has been used to prevent evolutionary changes at a given site in a protein.

[0182] Pan-amino acid evolutionary spectrum

[0183] Figure 12 This is a flowchart illustrating a specific embodiment of the technology disclosed in this invention, detailing a process 1200 for determining and allocating pan-amino acid conserved frequencies (voxolization) relative to the nearest atom to a voxel. The following will be discussed sequentially. Figure 12 , Figure 13 , Figure 14 , Figure 15 , Figure 16 , Figure 17 and Figure 18 .

[0184] At step 1202, the system's similarity sequence finder 1204 searches for amino acid sequences that are similar (homologous) to the reference amino acid sequence 202. Similar amino acid sequences can be selected from multiple species, such as primates, mammals, and vertebrates.

[0185] In step 1212, the system's aligner 1214 compares the reference amino acid sequence 202 with similar amino acid sequences in a positional manner; that is, the aligner 1214 performs multiple sequence alignment. Figure 14An example multiple sequence alignment 1400 is shown, representing a reference amino acid sequence 202 across 99 species. In some embodiments, the multiple sequence alignment 1400 can be divided, for example, to generate 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, single position frequency matrices are generated across the 99 species.

[0186] At step 1222, the system’s panamino acid conserved frequency calculator 1224 uses multiple sequence alignment to determine the panamino acid conserved frequency of the reference amino acid in the reference amino acid sequence 202.

[0187] At step 1232, the system's nearest atom finder 1234 searches for the atom in the voxel grid 522 that is closest to voxel 514. In some embodiments, the search for the voxel-based nearest atom may not be limited to any specific amino acid class or atom type. That is, voxel-based nearest atoms can be selected across amino acid classes and types, as long as they are the atoms closest to the corresponding voxel center. In other embodiments, the search for the voxel-based nearest atom may be limited to specific atom classes, such as limited to specific atomic elements (e.g., oxygen, nitrogen, and hydrogen), or limited to α-carbon atoms, or limited to β-carbon atoms, or limited to side-chain atoms, or limited to main-chain atoms.

[0188] At step 1242, the system's amino acid selector 1244 selects those reference amino acids in the reference amino acid sequence 202 that contain the closest atom identified at step 1232. These reference amino acids can be referred to as the closest reference amino acids. Figure 13 An example is shown where the nearest atom 1302 is located to voxel 514 in voxel grid 522, and where the nearest reference amino acid 1312 containing the nearest atom 1302 is mapped to voxel 514 in voxel grid 522. This is in Figure 13 The label is "Voxel to closest amino acid mapping 1300".

[0189] At step 1252, the system's voxelizer 1254 voxels the panamino acid conserved frequency that is closest to the reference amino acid. Figure 15 An example of determining the pan-amino acid conserved frequency sequence of the first voxel (1,1) in voxel grid 522 is shown, also referred to herein as “determining 1500 per voxel evolutionary spectrum”.

[0190] Go to Figure 13The closest reference amino acid mapped to the first voxel (1,1) is the aspartic (D) amino acid at position 15 in reference amino acid sequence 202. Then, multiple sequence alignments of reference amino acid sequence 202 with, for example, 99 homologous amino acid sequences from 99 species were analyzed at position 15. This position-specific and cross-species analysis revealed how many instances of amino acids from each of the 21 amino acid classes were found at position 15 across 100 aligned amino acid sequences (i.e., reference amino acid sequence 202 plus 99 homologous amino acid sequences).

[0191] exist Figure 15 In the example shown, the aspartic (D) amino acid is found at position 15 in 96 of the 100 aligned amino acid sequences. Therefore, aspartic amino acid category 1504 is assigned a pan-amino acid conserved frequency of 0.96. Similarly, in the example shown, the valine (V) amino acid is found at position 15 in 4 of the 100 aligned amino acid sequences. Therefore, valine amino acid category 1514 is assigned a pan-amino acid conserved frequency of 0.04. Since no amino acid instances from other amino acid categories were detected at position 15, the remaining amino acid categories are assigned a pan-amino acid conserved frequency of 0. Thus, a corresponding pan-amino acid conserved frequency is assigned to each of the 21 amino acid categories, which can be encoded in the pan-amino acid conserved frequency sequence 1502 of the first voxel (1,1).

[0192] Figure 16 It shows the use of Figure 15 The position-frequency logic described herein (also referred to herein as “voxel-to-evolutionary spectrum mapping 1600”) determines the corresponding pan-amino acid conserved frequencies 1612 to 1692 for the corresponding voxels in voxel 514 of voxel grid 522.

[0193] Then, the voxelizer 1254 uses the per-voxel evolution spectrum 1602 to generate the voxelized per-voxel evolution spectrum 1700, such as Figure 17 As shown. Typically, each voxel in voxel 514 within voxel grid 522 has a different pan-amino acid conserved frequency sequence, and therefore a different voxelization per-voxel evolutionary spectrum, because voxels are regularly mapped to different nearest-neighbor atoms, and thus to different nearest-neighbor reference amino acids. Of course, when two or more voxels have the same nearest-neighbor atom and therefore the same nearest-neighbor reference amino acid, the same pan-amino acid conserved frequency sequence and the same voxelization per-voxel evolutionary spectrum are assigned to each voxel in the two or more voxels.

[0194] Figure 18An instance of the evolutionary spectrum tensor 1800 is depicted, where each voxel of the evolutionary spectrum 1700 is voxelized and interconnected in a voxel-like manner according to the spatial arrangement of voxels 514 in a voxel grid 522. The voxelization dimension of the evolutionary spectrum tensor 1800 is 21×3×3×3 (where 21 represents 21 amino acid classes and 3×3×3 represents a 3D voxel grid with 27 voxels); but Figure 18 It is a 2D depiction with an evolutionary spectrum tensor of 1800 and dimensions of 21×3×3.

[0195] At step 1262, connector 174 connects the evolutionary spectrum tensor 1800 to the distance channel tensor 700 in a voxel manner. In some implementations, the evolutionary spectrum tensor 1800 is connected to connector tensor 1110 in a voxel manner to generate another connection tensor (not shown) with dimensions 84×3×3×3.

[0196] At step 1272, runtime logic 184 processes the 84×3×3×3 dimension of the further link tensor through the pathogenicity classifier to determine the pathogenicity of the target variant, which is then inferred as the pathogenicity determination result of the underlying nucleotide variant that produces the target variant at the amino acid level.

[0197] Evolutionary profile of each amino acid

[0198] Figure 19 This is a flowchart illustrating the process of a system for determining and assigning the conserved frequency (voxelization) of each amino acid to the atom closest to the voxel. (1900) Figure 19 In the middle, steps 1202 and 1212 are related to Figure 12 The same as in [the previous sentence].

[0199] At step 1922, the system’s per-amino acid conservation frequency calculator 1924 uses multiple sequence alignment to determine the per-amino acid conservation frequency of the reference amino acid in the reference amino acid sequence 202.

[0200] At step 1932, the system's nearest atom finder 1934 searches for 21 nearest atoms across 21 amino acid categories for each voxel in voxel 514 of voxel grid 522. Each of the 21 nearest atoms is distinct from the others because they are selected from different amino acid categories. This allows for the selection of 21 unique nearest reference amino acids for a particular voxel, which in turn generates 21 unique position-frequency matrices for that particular voxel, which in turn determines 21 unique per-amino acid conserved frequencies for that particular voxel.

[0201] At step 1942, the system's amino acid selector 1944 selects 21 reference amino acids from the reference amino acid sequence 202 for each voxel in voxel 514 of voxel grid 522, which contain the 21 closest atoms identified in step 1932. These reference amino acids can be referred to as the closest reference amino acids.

[0202] At step 1952, the system's voxelizer 1954 voxels the conserved frequencies per amino acid of the 21 nearest-neighbor reference amino acids identified at step 1942 for a specific voxel. These 21 nearest-neighbor reference amino acids must be located at 21 different positions in the reference amino acid sequence 202 because they correspond to different fundamental nearest-neighbor atoms. Therefore, for a specific voxel, 21 position-frequency matrices can be generated for these 21 nearest-neighbor reference amino acids. As mentioned above regarding... Figures 12 to 15 The 21 position frequency matrices discussed can be generated across multiple species by aligning their homologous amino acid sequences with the reference amino acid sequence 202 in a positional manner.

[0203] Then, using this 21 position-frequency matrix, 21 position-specific conservation scores can be calculated for the 21 closest reference amino acids identified for a specific voxel. These 21 position-specific conservation scores form the per-amino acid conservation frequency for a specific voxel, similar to... Figure 12 The pan-amino acid conserved frequency sequence 1502 is different in that sequence 1502 has many 0 entries, while each element (feature) in the per-amino acid conserved frequency sequence has a certain value (e.g., a floating-point number) because the 21 closest reference amino acids across 21 amino acid classes must have different positions that produce different positional frequency matrices, and thus different per-amino acid conserved frequencies.

[0204] The above processing is performed on each voxel in voxel 514 of voxel grid 522, and the resulting voxel formula is voxelized, tensorized, linked, and processed for the conserved frequency per amino acid, so as to be compatible with the information about... Figures 12 to 18 The conserved frequencies of the pan-amino acids discussed were similarly used for pathogenicity determination.

[0205] Comment channel

[0206] Figure 20Various examples of voxelized annotation channels 2000 linked to a distance channel tensor 700 are shown. In some embodiments, voxelized annotation channels are unique indicators of different protein annotations, such as whether an amino acid (residue) is part of a transmembrane region, signal peptide, active site, or any other binding site, or whether the residue undergoes post-translational modification, path ratio (see Pei P, Zhang A: A Topological Measurement for Weighted Protein Interaction Network. CSB 2005, 268-278), etc. Additional examples of annotation channels can be found in the specific embodiments section below and in the claims.

[0207] These voxelized annotation channels are arranged in a voxel-like manner, so that voxels can have the same annotation sequence, such as voxelized reference allele sequence and alternative allele sequence (e.g., annotation channels 2002, 2004, 2006), or voxels can have their own annotation sequence, such as voxelized evolutionary spectrum 1700 per voxel (e.g., annotation channels 2012, 2014, 2016 (as indicated by different colors)).

[0208] These annotation channels are voxelized, tensorized, linked, and processed to allow for their integration with information about... Figures 12 to 18 The conserved frequencies of the pan-amino acids discussed were similarly used for pathogenicity determination.

[0209] Structural confidence channel

[0210] The technology disclosed in this invention can also link various voxelized structure confidence channels with a distance channel tensor of 700. Some examples of structure confidence channels include GMQE scores (provided by SwissModel); B-factor; temperature factor column of the homology model (indicating the degree to which residues satisfy (physical) constraints in the protein structure); normalized number of alignments of the template protein for the residues closest to the voxel center (alignments provided by HHpred, e.g., the residues closest to the voxel of the three template structures aligned out of six, which means the value of this feature is 3 / 6 = 0.5); minimum, maximum, and average TM scores; and predicted TM scores of the template protein structure aligned to the residues closest to the voxel (continuing with the above examples, assuming the TM scores of these three template structures are 0.5, 0.5, and 1.5, then the minimum is 0.5, the average is 2 / 3, and the maximum is 1.5). The TM score for each protein template can be provided by HHpred. Additional examples of structure confidence channels can be found in the “Specific Embodiments” section below and in the claims.

[0211] These voxelized structural confidence channels are arranged in a voxel-like manner, allowing voxels to have the same structural confidence sequence, such as a voxelized reference allele sequence and an alternative allele sequence, or voxels to have their own structural confidence sequence, such as voxelized evolutionary spectrum 1700 per voxel.

[0212] These structural confidence channels are voxelized, tensorized, linked, and processed to enable their association with... Figures 12 to 18 The conserved frequencies of the pan-amino acids discussed were similarly used for pathogenicity determination.

[0213] Pathogenicity classifier

[0214] Figure 21 Different combinations and arrangements of input channels are shown, which can be provided as input 2102 to the pathogenicity classifier 2108 for pathogenicity determination 2106 of the target variant. One of the inputs 2102 can be a distance channel 2104 generated by the distance channel generator 2272. Figure 22Different methods for calculating distance channel 2104 are illustrated. In one embodiment, distance channel 2104 is generated based on the distance 2202 between the voxel center and atoms spanning multiple atomic elements, independent of amino acids. In some embodiments, distance 2202 is normalized by the maximum scan radius to generate a normalized distance 2202a. In another embodiment, distance channel 2104 is generated on an amino acid basis based on the distance 2212 between the voxel center and the α-carbon atom. In some embodiments, distance 2212 is normalized by the maximum scan radius to generate a normalized distance 2212a. In yet another embodiment, distance channel 2104 is generated on an amino acid basis based on the distance 2222 between the voxel center and the β-carbon atom. In some embodiments, distance 2222 is normalized by the maximum scan radius to generate a normalized distance 2222a. In yet another embodiment, distance channel 2104 is generated on an amino acid basis based on the distance 2232 between the voxel center and the side chain atom. In some embodiments, distance 2232 is normalized by the maximum scan radius to generate a normalized distance 2232a. In yet another embodiment, distance channel 2104 is generated on an amino acid basis based on the distance 2242 between the voxel center and the main chain atom. In some embodiments, distance 2242 is normalized by the maximum scan radius to generate a normalized distance 2242a. In yet another embodiment, distance channel 2104 is generated based on the distance 2252 (a feature) between the voxel center and the corresponding nearest atom, regardless of the atom type and amino acid type. In yet another embodiment, distance channel 2104 is generated based on the distance 2262 (a feature) between the voxel center and the atom from a non-standard amino acid. In some embodiments, the distance between the voxel and the atom is calculated based on the polar coordinates of the voxel and the atom. The polar coordinates are parameterized by the angle between the voxel and the atom. In one embodiment, this angle information is used to generate the voxel's angle channel (i.e., independent of the distance channel). In some specific implementations, the angle between the nearest atom and its neighboring atoms (e.g., main chain atoms) can be used as a feature that utilizes voxel encoding.

[0215] Another input in 2102 could be feature 2114, which indicates the absence of atoms within a specified radius.

[0216] Another option in input 2102 could be the one-hot code 2124 for the reference amino acid. Another option in input 2102 could be the one-hot code 2134 for a variant / substitute amino acid.

[0217] The other one in input 2102 could be from Figure 23The evolutionary spectrum generator 2372 shown generates evolutionary channel 2144. In one embodiment, evolutionary channel 2144 can be generated based on pan-amino acid conserved frequencies 2302. In another embodiment, evolutionary channel 2144 can be generated based on pan-amino acid conserved frequencies 2312.

[0218] Another input in 2102 can be a feature 2154 indicating a missing residue or a missing evolutionary spectrum.

[0219] The other input in 2102 could be annotation channel 2164 generated by annotation generator 2472, such as... Figure 24 As shown. In one embodiment, annotation channel 2154 can be generated based on molecular processing annotation 2402. In another embodiment, annotation channel 2154 can be generated based on region annotation 2412. In yet another embodiment, annotation channel 2154 can be generated based on site annotation 2422. In yet another embodiment, annotation channel 2154 can be generated based on amino acid modification annotation 2432. In yet another embodiment, annotation channel 2154 can be generated based on secondary structure annotation 2442. In yet another embodiment, annotation channel 2154 can be generated based on experimental information annotation 2452.

[0220] Another input 2102 can be the structure confidence channel 2174 generated by the structure confidence generator 2572, such as Figure 25 As shown. In one embodiment, the structural confidence score 2174 can be generated based on the Global Model Quality Estimate (GMQE) 2502. In another embodiment, the structural confidence score 2174 can be generated based on the Qualitative Model Energy Analysis (QMEAN) score 2512. In yet another embodiment, the structural confidence score 2174 can be generated based on the temperature factor 2522. In yet another embodiment, the structural confidence score 2174 can be generated based on the template modeling score 2542. Examples of template modeling scores 2542 include the minimum template modeling score 2542a, the average template modeling score 2542b, and the maximum template modeling score 2542c.

[0221] Those skilled in the art will recognize that any permutation and combination of input channels can be concatenated as input for processing by the pathogenicity classifier 2108 to determine the pathogenicity of the target variant 2106. In some embodiments, only a subset of the input channels may be concatenated. These input channels can be concatenated in any order. In one embodiment, these input channels may be concatenated into a single tensor by a tensor generator (input encoder) 2110. This single tensor can then be provided as input to the pathogenicity classifier 2108 to determine the pathogenicity of the target variant 2106.

[0222] In one implementation, the pathogenicity classifier 2108 uses a convolutional neural network (CNN) with multiple convolutional layers. In another implementation, the pathogenicity classifier 2108 uses a recurrent neural network (RNN), such as a Long Short-Term Memory (LSTM), a Bi-LSTM (Bi-LSTM), and a Gated Recurrent Unit (GRU). In yet another implementation, the pathogenicity classifier 2108 uses both CNN and RNN. In yet another implementation, the pathogenicity classifier 2108 uses a graph convolutional neural network that models dependencies in graph-structured data. In yet another implementation, the pathogenicity classifier 2108 uses a variational autoencoder (VAE). In yet another implementation, the pathogenicity classifier 2108 uses a generative adversarial network (GAN). In yet another implementation, the pathogenicity classifier 2108 can also be based on, for example, a self-attention language model, such as a language model implemented by a transformer and BERT.

[0223] In other specific implementations, the pathogenicity classifier 2108 may use 1D convolution, 2D convolution, 3D convolution, 4D convolution, 5D convolution, dilated or dilated convolution, transposed convolution, depthwise separable convolution, pointwise convolution, 1×1 convolution, grouped convolution, flattened convolution, spatial and cross-channel convolution, shuffled grouped convolution, spatially separable convolution, and deconvolution. The pathogenicity classifier may 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, smoothed L1 loss, and Huber loss. The pathogenicity classifier may use any parallelism, efficiency, and compression scheme, such as TFRecord, compressed encoding (e.g., PNG), sharpening, parallel detection of map transformations, batch processing, prefetching, model parallelism, data parallelism, and synchronous / asynchronous stochastic gradient descent (SGD). The pathogenicity classifier may include upsampling layers, downsampling layers, recursive connections, gates and gate memory cells (such as LSTM or GRU), residual blocks, residual connections, high-speed connections, skip connections, peephole connections, activation functions (e.g., nonlinear transformation functions such as modified linear unit (ReLU), leaky ReLU, exponential linear unit (ELU), sigmoid hyperbolic tangent (tanh)), batch normalization layers, regularization layers, dropout layers, pooling layers (e.g., max or average pooling), global average pooling layers, attention mechanisms, and Gaussian error linear units.

[0224] The pathogenicity classifier 2108 is trained using a gradient update technique based on backpropagation. Example gradient descent techniques that can be used to train the pathogenicity classifier 2108 include stochastic gradient descent, batch gradient descent, and micro-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 specific implementations, the pathogenicity classifier 2108 can be trained using unsupervised learning, semi-supervised learning, self-learning, reinforcement learning, multi-task learning, multimodal learning, transfer learning, knowledge distillation, etc.

[0225] Figure 26 An example processing architecture 2600 for a pathogenicity classifier 2108 according to a specific embodiment of the technology disclosed in this invention is shown. The processing architecture 2600 includes a cascade of processing modules 2606, 2610, 2614, 2618, 2622, 2626, 2630, 2634, 2638, and 2642, wherein each processing module may include 1D convolution (1×1×1CONV), 3D convolution (3×3×3CONV), ReLU nonlinearity, and batch normalization (BN). Other instances of processing modules include fully connected (FC) layers, dropout layers, flattening layers, and a final Softmax layer that generates exponentially normalized scores for target variants belonging to the benign and pathogenic categories. Figure 26 In this context, "64" indicates the number of convolutional filters applied by a specific processing module. Figure 26 In the input voxel 2602, the dimensions are 15×15×15×8. Figure 26 The corresponding volumetric dimensions of the intermediate inputs 2604, 2608, 2612, 2616, 2620, 2624, 2628, 2632, 2636 and 2640 generated by the processing architecture 2600 are also shown.

[0226] Figure 27An example processing architecture 2700 of a pathogenicity classifier 2108 according to a specific embodiment of the technology disclosed in this invention is shown. The processing architecture 2700 includes a cascade of processing modules 2708, 2714, 2720, 2726, 2732, 2738, 2744, 2750, 2756, 2762, 2768, 2774, and 2780, such as 1D convolution (CONV 1D), 3D convolution (CONV 3D), ReLU nonlinearity, and batch normalization (BN). Other instances of processing modules include fully connected (dense) layers, dropout layers, flattening layers, and a final Softmax layer that generates exponentially normalized scores for target variants belonging to the benign and pathogenic categories. Figure 27 In this context, "64" and "32" represent the number of convolutional filters applied by a specific processing module. Figure 27 In this context, the size of the input voxel 2704 provided by the input layer 2702 is 7×7×7×108. Figure 27 The corresponding volumetric dimensions of the intermediate inputs 2710, 2716, 2722, 2728, 2734, 2740, 2746, 2752, 2758, 2764, 2770, 2776 and 2782 generated by the processing architecture 2700 and the resulting intermediate outputs 2706, 2712, 2718, 2724, 2730, 2736, 2742, 2748, 2754, 2760, 2766, 2772, 2778 and 2784 are also shown.

[0227] Those skilled in the art will recognize that other current and future artificial intelligence, machine learning and deep learning models, datasets and training techniques can be incorporated into the variant pathogenicity classifiers disclosed in this invention without departing from the spirit of the techniques disclosed herein.

[0228] Performance results as objective indicators of creativity and non-obviousness

[0229] The variant pathogenicity classifier disclosed in this paper predicts pathogenicity based on 3D protein structure and is referred to as "PrimateAI 3D". "Primate AI" is a co-owned and previously disclosed variant pathogenicity classifier that predicts pathogenicity based on protein sequence. Additional details about PrimateAI can be found in co-owned U.S. patent applications 16 / 160,903, 16 / 160,986, 16 / 160,968 and 16 / 407,149, and Sundaram, L. et al., Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018).

[0230] Figure 28 , Figure 29 , Figure 30 and Figure 31A We used PrimateAI as a benchmark model to demonstrate the classification advantages of PrimateAI 3D over PrimateAI. Figure 28 , Figure 29 , Figure 30 , Figure 31A and Figure 31B The performance results were generated on a classification task that accurately distinguishes benign variants from pathogenic variants across multiple validation sets. PrimateAI 3D was trained on training sets different from the multiple validation sets. PrimateAI 3D was trained on common human and primate variants used as benign datasets, while simulated variants based on trinucleotide backgrounds were used as unlabeled or spurious pathogenic datasets.

[0231] A novel developmental retardation disorder (novel DDD) is an example of a validation set used to compare the classification accuracy of Primate AI 3D with Primate AI. The novel DDD validation set labels variants from DDD individuals as pathogenic, while labeling the same variants from healthy relatives of DDD individuals as benign. A similar labeling scheme is used... Figure 31A and Figure 31B The validation set for autism spectrum disorder (ASD) is shown in the figure.

[0232] BRCA1 is another example of a validation set used to compare the classification accuracy of Primate AI 3D with Primate AI. The BRCA1 validation set labels the reference amino acid sequence of synthetically generated BRCA1 gene-mimicking proteins as benign variants and the allelic amino acid sequences of synthetically generated altered BRCA1 gene-mimicking proteins as pathogenic variants. A similar labeling scheme is used for the TP53 gene, TP53S3 gene and its variants, and... Figure 31A and Figure 31B The other genes and their variants shown are different validation sets.

[0233] Figure 28The performance of the baseline PrimateAI model is indicated by blue horizontal bars, and the performance of the PrimateAI 3D model disclosed in this invention is indicated by orange horizontal bars. Green horizontal bars depict the pathogenicity predictions derived by combining the corresponding pathogenicity predictions of the PrimateAI 3D model disclosed in this invention and the baseline PrimateAI model. In the legend, “ens10” represents an ensemble of 10 PrimateAI 3D models, each trained with a different seed training dataset and randomly initialized with different weights and biases. Additionally, “7×7×7×2” depicts the size of the voxel grid used to encode the input channels during the training of the ensemble of 10 PrimateAI 3D models. For a given variant, the ensemble of 10 PrimateAI 3D models generates 10 pathogenicity predictions, which are then combined (e.g., by averaging) to generate a final pathogenicity prediction for the given variant. This logic similarly applies to ensembles with different group sizes.

[0234] In addition, Figure 28 In the diagram, the y-axis has different validation sets, and the x-axis has p-values. A larger p-value, i.e., a longer horizontal bar, indicates higher accuracy in distinguishing between benign and pathogenic variants. For example... Figure 28 As confirmed by the p-values, PrimateAI 3D outperforms PrimateAI on most validation sets (except for the tp53s3_A549 validation set). That is, the orange horizontal bar of PrimateAI 3D is always longer than the blue horizontal bar of PrimateAI.

[0235] In addition, Figure 28 In the y-axis category, the "mean value" category is calculated as the average of the p-values ​​determined for each validation set in the validation set. Also within this mean value category, PrimateAI 3D outperforms PrimateAI.

[0236] exist Figure 29 In the diagram, PrimateAI is represented by blue horizontal bars, the ensemble of 20 PrimateAI 3D models trained with a 3×3×3 voxel mesh is represented by red horizontal bars, the ensemble of 10 PrimateAI 3D models trained with a 7×7×7×2 voxel mesh is represented by purple horizontal bars, the ensemble of 20 PrimateAI 3D models trained with a 7×7×7×2 voxel mesh is represented by brown horizontal bars, and the ensemble of 20 PrimateAI 3D models trained with a 17×17×17×2 voxel mesh is represented by purple horizontal bars.

[0237] In addition, Figure 29In the diagram, the y-axis has different validation sets, and the x-axis has p-values. As before, a larger p-value, i.e., a longer horizontal bar, indicates higher accuracy in distinguishing between benign and pathogenic variants. Figure 20 As confirmed by the p-values, different configurations of PrimateAI 3D outperform PrimateAI on most validation sets. Specifically, the red, purple, brown, and pink horizontal bars of PrimateAI 3D are mostly longer than the blue horizontal bars of PrimateAI.

[0238] In addition, Figure 29 In this context, the "Average" category along the y-axis is calculated as the average of the p-values ​​determined for each validation set in the validation set. Also within this Average category, different configurations of PrimateAI 3D outperform PrimateAI.

[0239] exist Figure 30 In the image, the red vertical bar represents PrimateAI, and the cyan vertical bar represents PrimateAI 3D. Figure 30 In this context, the y-axis has p-values, and the x-axis has different validation sets. Figure 30 Without exception, PrimateAI 3D consistently outperforms PrimateAI on all validation sets. That is, the cyan vertical bars in PrimateAI 3D are always longer than the red vertical bars in PrimateAI.

[0240] Figure 31A and Figure 31B The performance of the baseline PrimateAI model is indicated by blue vertical bars, and the performance of the PrimateAI 3D model disclosed in this invention is indicated by orange vertical bars. Green vertical bars depict the pathogenicity prediction derived by combining the corresponding pathogenicity predictions of the PrimateAI 3D model disclosed in this invention and the baseline PrimateAI model. Figure 31A and Figure 31B In the diagram, the y-axis has a p-value, and the x-axis has different validation sets.

[0241] like Figure 31A and Figure 31B As confirmed by the p-values, PrimateAI 3D outperforms PrimateAI on most validation sets (except for the tp53s3_A549_p53NULL_Nutlin-3 validation set). That is, the orange bars of PrimateAI 3D are always longer than the blue bars of PrimateAI.

[0242] In addition, Figure 31A and Figure 31BIn the graph, the individual "average" chart is calculated as the average of the p-values ​​determined for each validation set in the validation set. Also in this average chart, PrimateAI 3D outperforms PrimateAI.

[0243] The average statistic may be biased due to outliers. To address this issue, Figure 31A and Figure 31B The document also includes a separate "Method Rank" chart. Higher ranks indicate lower classification accuracy. Similarly, in this method rank chart, PrimateAI 3D outperforms PrimateAI because PrimateAI 3D has more counts of lower ranks 1 and 2, compared to PrimateAI's rank of 3 for all of them.

[0244] exist Figure 28 , Figure 29 , Figure 30 , Figure 31A and Figure 31B Similarly, it is evident that combining PrimateAI 3D with PrimateAI produces superior classification accuracy. Specifically, a protein can be fed to PrimateAI as its amino acid sequence to generate the first output, and the same protein can be fed to PrimateAI3D as a 3D voxelized protein structure to generate the second output. The first and second outputs can then be combined or aggregated for analysis to produce a final prediction of the pathogenicity of the variants the protein has undergone.

[0245] High-efficiency voxelization

[0246] Figure 32 is a flowchart illustrating the efficient voxelization process 3200 that efficiently identifies the closest atom on a voxel-by-voxel basis.

[0247] Now let's revisit the distance channel. As discussed above, the reference amino acid sequence 202 can contain different types of atoms, such as α-carbon atoms, β-carbon atoms, oxygen atoms, nitrogen atoms, hydrogen atoms, etc. Therefore, as discussed above, the distance channel can be arranged by the nearest α-carbon atom, the nearest β-carbon atom, the nearest oxygen atom, the nearest nitrogen atom, the nearest hydrogen atom, etc. For example, in... Figure 6 In this study, each of the nine voxels 514 has 21 amino acid-based distance channels for the nearest α-carbon atom. Figure 6 This can be further extended to each of the 9 voxels 514 having 21 amino acid-based distance channels for the nearest β-carbon atom, and each of the 9 voxels 514 also having a universal atomic distance channel for the nearest atom, regardless of the atom type or amino acid type. Thus, each of the 9 voxels 514 could have 43 distance channels.

[0248] The discussion now turns to the number of distances required to include the nearest atom in the distance channel, calculated on a voxel-by-voxel basis. Consider... Figure 3 The example in the text depicts a total of 828 α-carbon atoms distributed across all 21 amino acid classes. To calculate... Figure 6 The amino acid-based distance channels 602 to 642, i.e., to determine 189 distance values, measure the distance from each voxel in 9 voxels 514 to each α-carbon atom in 828 α-carbon atoms, resulting in 9 * 828 = 7,452 calculated distance values. With 27 voxels being 3D voxels, this yields 828 * 27 = 22,356 calculated distance values. With the inclusion of 828 β-carbon atoms, this number increases to 27 * 1656 = 44,712 calculated distance values.

[0249] like Figure 35A As shown, this means that the runtime complexity of identifying the nearest atom on a voxel-by-voxel basis for a single protein voxelization is O(#atom * #voxel). Furthermore, when calculating distance channels across multiple attributes (e.g., different features or channels for each voxel, such as annotation channels and structure confidence channels), the runtime complexity of a single protein voxelization increases to O(#atom * #voxel * #attribute).

[0250] Therefore, distance computation can become the most computationally intensive part of the voxelization process, thus diverting 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 dataset of 7,000 proteins. Generating distance channels for multiple voxels across multiple amino acids, atoms, and properties can involve more than 100 voxels per protein, resulting in approximately 800,000 voxels in a single training iteration (epoch). A training run of 20 to 40 epochs (with rotations of atomic coordinates in each epoch) can generate up to 32 million voxels.

[0251] In addition to the high computational cost, 32 million voxelizations also resulted in a massive amount of data that could not be stored in main memory (e.g., for a 15×15×15 voxel grid, the data volume exceeds 20TB). Considering the repetitive training runs used for parameter optimization and ensemble learning, the memory footprint of this voxelization process became too large to be stored on disk, making it a part of model training rather than a pre-computation step.

[0252] The technology disclosed in this invention provides an efficient voxelization process that achieves up to approximately 100 times speedup with a runtime complexity of O(#atoms * #voxels). This efficient voxelization process reduces the runtime complexity of a single protein voxelization to O(#atoms). When each voxel has different features or channels, this efficient voxelization process reduces the runtime complexity of a single protein voxelization to O(#atoms * #attributes). Therefore, this voxelization process becomes as fast as model training, thereby shifting the computational bottleneck from voxelization back to computing neural network weights on processors such as GPUs, ASICs, TPUs, FPGAs, and CGRAs.

[0253] In some specific embodiments of the efficient voxelization process involving large voxel grids disclosed in this invention, the runtime complexity of a single protein voxelization is O(#atom + voxel) and O(#atom * #attribute + voxel) for cases where each voxel has different characteristics or channels. When the number of atoms is negligible compared to the number of voxels, for example, when there is only one atom in a 100×100×100 voxel grid (i.e., 1 million voxels per atom), a "+voxel" complexity is observed. In this case, runtime is dominated by the overhead of a massive number of voxels, such as initializing 1 million voxels to 0 in order to allocate memory to them, etc.

[0254] The discussion now turns to the details of the efficient voxelization process disclosed in this invention. This will be discussed in turn. Figure 32A , Figure 32B , Figure 33 , Figure 34 and Figure 35B .

[0255] from Figure 32A Beginning at step 3202, each atom (e.g., each α-carbon atom out of 828 α-carbon atoms and each β-carbon atom out of 828 β-carbon atoms) is associated with a voxel containing that atom (e.g., one voxel out of 9 voxels 514). The term "containing" means that the 3D atomic coordinates of the atom are located within the voxel. The voxel containing the atom is also referred to herein as an "atom-containing voxel".

[0256] Figure 32B and Figure 33 It describes how to select voxels containing specific atoms. Figure 33 2D atomic coordinates are used as a representation of 3D atomic coordinates. Note that the voxel grid 522 is regularly spaced, with each voxel in voxel 514 having the same step size (e.g., 1 angstrom). or ).

[0257] In addition, Figure 33 In the voxel grid 522, magenta indices [0,1,2] are present along the first dimension (e.g., the x-axis), and cyan indices [0,1,2] are present along the second dimension (e.g., the y-axis). Additionally, in... Figure 33 In the text, the corresponding voxel 514 in voxel 512 is identified by the green voxel index [voxel 0, voxel 1, ..., voxel 8] and the black voxel center index [(1,1),(1,2),...,(3,3)].

[0258] In addition, Figure 33 In the diagram, the center coordinates of the voxel centers along the first dimension (i.e., the first-dimensional voxel coordinates) are marked in orange. Additionally, in... Figure 33 In the diagram, the center coordinates of the voxel centers along the second dimension (i.e., the second-dimensional voxel coordinates) are marked in red.

[0259] First, in step 3202a ( Figure 33 In step 1), the 3D atomic coordinates (1.7456, 2.14323) of a specific atom are quantized to generate quantized 3D atomic coordinates (1.7, 2.1). This quantization can be achieved by rounding or truncating the bits.

[0260] Then, in step 3202b ( Figure 33 At step 2), the voxel coordinates (or voxel center or voxel center coordinates) of voxel 514 are assigned to the quantized 3D atomic coordinates based on the dimension. For the first dimension, the quantized atomic coordinate 1.7 is assigned to voxel 1 because this atomic coordinate covers the first dimension voxel coordinates from 1 to 2 and is centered at 1.5 in the first dimension. Note that voxel 1 has an index of 1 along the first dimension, compared to having an index of 0 along the second dimension.

[0261] For the second dimension, starting with 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, since this atomic coordinate covers the 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 an index of 2 along the second dimension, compared to having an index of 1 along the first dimension.

[0262] Then, in step 3202c ( Figure 33 At step 3), the dimension index corresponding to the assigned voxel coordinates is 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 know that the above steps can be performed similarly for the third dimension to select the dimension index along the third dimension.

[0263] Then, in step 3202d ( Figure 33In step 4), a cumulative sum is generated based on positional weighting of the selected dimension indices using powers of the base. The general idea behind positional numbering systems is to represent values ​​by powers of an increasing base (or radix), for example, base 2 for binary, base 3 for ternary, base 8 for octal, and base 16 for hexadecimal. This is often called a weighted numbering system because each position is weighted by a power of the base. The size of the set of significant values ​​in a positional numbering system is equal to the base of the system. For example, a decimal system has 10 digits, 0 through 9, and a ternary system has 3 digits, 0, 1, and 2. The largest significant digit in a base system is one less than the base (therefore, 8 is not a significant digit in any base system less than 9). Any decimal integer can be precisely represented using any other integer base system, and vice versa.

[0264] Return to Figure 33 In the example, the selected dimension indices 1 and 2 are converted into single integers by multiplying each dimension index by its respective cardinality power in a positional manner, and then summing the results of these positional multiplications. Cardinality 3 is chosen here because 3D atomic coordinates have three dimensions (but for simplicity...). Figure 33 Only 2D atomic coordinates along two dimensions are shown.

[0265] Since index 2 is in the rightmost position (i.e., the least significant bit), multiplying it by 3 to the power of 0 gives 2. Since index 1 is in the second rightmost position (i.e., the second least significant bit), multiplying it by 3 to the power of 1 gives 3. This makes the sum 5.

[0266] Then, in step 3202e( Figure 33 In step 5), the voxel index containing a specific atom is selected based on the summation. That is, the summation is interpreted as the voxel index containing the specific atom.

[0267] At step 3212, after each atom is associated with the 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 “neighborhood voxels”). Neighborhood voxels can be based on a predefined radius of the atom-containing voxel (e.g., 5 angstroms). The selection is made within a range. In other implementations, neighboring voxels can be selected based on their consecutive adjacency to the voxel containing the atom (e.g., top neighbor voxel, bottom neighbor voxel, right neighbor voxel, left neighbor voxel). The resulting association of each atom with the voxel containing the atom and neighboring voxels is encoded in an atom-to-voxel mapping 3402, also referred to herein as an element-to-cell mapping. In one example, a first α-carbon atom is associated with a first subset 3404 of voxels, which includes the voxel containing the atom and neighboring voxels of the first α-carbon atom. In another example, a second α-carbon atom is associated with a second subset 3406 of voxels, which includes the voxel containing the atom and neighboring voxels of the second α-carbon atom.

[0268] It is important to note that distance calculations are not performed to determine the atomic voxels and neighboring voxels. The atomic voxels are selected based on the spatial arrangement of the voxels, which allows quantized 3D atomic coordinates to be assigned to corresponding regularly spaced voxel centers within the voxel mesh (without using any distance calculations). Additionally, neighboring voxels are selected based on their spatial adjacency to the atomic voxels in the voxel mesh (again, without using any distance calculations).

[0269] At step 3222, each voxel is mapped to the atom associated with it at steps 3202 and 3212. In one specific implementation, this mapping is encoded in a voxel-to-atom mapping 3412 generated based on the atom-to-voxel mapping 3402 (e.g., by applying a voxel-based sorting key to the atom-to-voxel mapping 3402). The voxel-to-atom mapping 3412 is also referred to herein as a “cell-to-element mapping”. In one example, a first voxel is mapped to a first subset 3414 of α-carbon atoms, which includes the α-carbon atoms associated with the first voxel at steps 3202 and 3212. In another example, a second voxel is mapped to a second subset 3416 of α-carbon atoms, which includes the α-carbon atoms associated with the second voxel at steps 3202 and 3212.

[0270] At step 3232, for each voxel, the distance between that voxel and the atom mapped to that voxel at step 3222 is calculated. The runtime complexity of step 3232 is O(#atom) because the distance to a particular atom is measured only once in the voxel-to-atom mapping 3412, which uniquely maps the corresponding voxel to that particular atom. This is true without considering adjacent voxels. In the absence of adjacent voxels, the constant factor implied in the Big O notation is 1. In the presence of adjacent voxels, the Big O notation is equal to the number of adjacent voxels + 1, because the number of adjacent voxels is constant for each voxel, so the runtime complexity O(#atom) remains true. In contrast, in Figure 35AIn this process, the distance to a particular atom is redundantly measured as many times as the number of voxels (e.g., since there are 27 voxels, the distance to a particular atom is measured 27 times).

[0271] exist Figure 35B In the diagram, based on the voxel-to-atom mapping 3412, each voxel is mapped to a corresponding subset of 828 atoms (excluding the calculated distances to neighboring voxels), as shown by the corresponding ellipse for each voxel. The corresponding subsets are largely non-overlapping, with some exceptions. When multiple atoms are mapped to the same voxel, there is subtle overlap due to several conditions, such as... Figure 35B The yellow overlap between the apostrophe "'" and the ellipse is indicated by this minimal overlap. This minimal overlap has an additive, rather than multiplicative, effect on runtime complexity for O (# atoms). This overlap is the result of considering adjacent voxels after determining the voxel containing the atom. Without adjacent voxels, there might be no overlap because an atom is associated with only one voxel. However, when considering adjacent voxels, each adjacent voxel can potentially be associated with the same atom (provided that no other atom of the same amino acid is closer to an adjacent voxel).

[0272] At step 3242, for each voxel, the nearest atom to that voxel is identified based on the distance calculated at step 3232. In one embodiment, this identification is encoded in a voxel-to-nearest-atom mapping 3422, also referred to herein as a “cell-to-nearest-element mapping.” In one example, the first voxel is mapped to the second α-carbon atom 3424, which is its nearest α-carbon atom. In another example, the second voxel is mapped to the 31st α-carbon atom 3426, which is its nearest α-carbon atom.

[0273] Furthermore, when using the techniques discussed above to calculate voxel-based distances, the atom type and amino acid type classification of the stored atoms, along with their corresponding distance values, are used to generate categorized distance channels.

[0274] Once the distances to the nearest atom are identified using the techniques described above, these distances can be encoded in the distance channel for voxelization and subsequently processed by the pathogenicity classifier 2108.

[0275] Computer System

[0276] Figure 36An example computer system 3600 that can be used to implement the techniques disclosed in this invention is shown. The computer system 3600 includes at least one central processing unit (CPU) 3672 that communicates with a plurality of peripheral devices via a bus subsystem 3655. These peripheral devices may include a storage subsystem 3610 (including, for example, memory devices and a file storage subsystem 3636), a user interface input device 3638, a user interface output device 3676, and a network interface subsystem 3674. The input and output devices allow users to interact with the computer system 3600. The network interface subsystem 3674 provides an interface to an external network, including interfaces to corresponding interface devices in other computer systems.

[0277] In one implementation, the pathogenicity classifier 2108 is communicatively linked to the storage subsystem 3610 and the user interface input device 3638.

[0278] User interface input device 3638 may include: a keyboard; pointing devices such as a mouse, trackball, touchpad, or graphics tablet; a scanner; a touchscreen integrated into a display; audio input devices such as a voice recognition system and a microphone; and other types of input devices. Generally, the term "input device" is intended to encompass all possible types of devices and methods for inputting information into computer system 3600.

[0279] User interface output device 3676 may include a display subsystem, printer, fax machine, or non-visual display (such as an audio output device). The display subsystem may include an LED display, cathode ray tube (CRT), flat panel device such as a liquid crystal display (LCD), projection device, or other mechanisms for producing visible images. The display subsystem may also provide non-visual displays, such as audio output devices. Generally, the term "output device" is intended to encompass all possible types of devices and methods for outputting information from computer system 3600 to a user or to another machine or computer system.

[0280] The storage subsystem 3610 stores the programming and data structures that provide the functionality of some or all of the modules and methods described herein. These software modules are typically executed by the processor 3678.

[0281] Processor 3678 can be a graphics processing unit (GPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), and / or a coarse-grained reconfigurable architecture (CGRA). Processor 3678 can be powered by deep learning cloud platforms such as Google Cloud Platform. TM Xilinx TM and Cirrascale TMManaged. Examples of processors 3678 include Google's Tensor Processing Unit (TPU). TM Rackmount solutions (such as the GX4 Rackmount Series) TM GX36 Rackmount Series TM NVIDIA DGX-1 TM Microsoft's Stratix V FPGA TM Graphcore's Intelligent Processing Unit (IPU) TM Qualcomm's Snapdragon processors TM Zeroth Platform TM NVIDIA's Volta TM NVIDIA's Drive PX TM , NVIDIA's JETSON TX1 / TX2 MODULE TM Intel's Nirvana TM Movidius VPU TM ,Fujitsu DPI TM ARM's DynamicIQ TM IBM TrueNorth TM With Testa VI 00s TM Lambda GPU servers, etc.

[0282] The memory subsystem 3622 used in the storage subsystem 3610 may include multiple 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 can provide persistent storage for program files and data files and may include hard disk drives, floppy disk drives, and associated removable media, CD-ROM drives, optical disk drives, or removable media enclosures. Modules implementing certain specific functionalities may be stored by the file storage subsystem 3636 in the storage subsystem 3610 or in other machines accessible to the processor.

[0283] The bus subsystem 3655 provides a mechanism for enabling various components and subsystems of the computer system 3600 to communicate with each other as intended. Although the bus subsystem 3655 is schematically shown as a single bus, alternative implementations of this bus subsystem may use multiple buses.

[0284] The computer system 3600 itself can be of different types, including personal computers, portable computers, workstations, computer terminals, network computers, televisions, mainframes, server clusters, a loosely networked group of widely distributed computers, or any other data processing system or user equipment. Due to the constantly evolving nature of computers and networks, [the following text appears to be incomplete and requires further context: "for..."] Figure 36 The description of the computer system 3600 depicted herein is intended only as a specific example to illustrate a preferred embodiment of the invention. The computer system 3600 may also have many other configurations, in which the components differ... Figure 36 The computer systems depicted in the book may be more or less.

[0285] Amino acid prediction

[0286] Supervised training of a protein language model using masked language modeling targets outputs the probability of an amino acid appearing at a given location within a protein. Proteins are linear polymers that fold into a variety of specific conformations to perform their functions. The incredible variety of three-dimensional (3D) structures determined by the combination and order of 20 amino acids that run through the protein polymer chain (the protein's sequence) enables proteins responsible for most biological activities to perform complex functions. Therefore, obtaining the structure of proteins is crucial for understanding the fundamental biology of health and disease, as well as for developing therapeutic molecules. While protein structures are primarily determined through sophisticated experimental techniques such as X-ray crystallography, NMR spectroscopy, and increasingly, cryo-electron microscopy, computational structure prediction from the genetically encoded amino acid sequence of a protein has been used as an alternative approach when experimental methods are limited.

[0287] Computational methods have been used to predict protein structures to elucidate the mechanisms of biological processes and determine protein properties. Furthermore, all naturally occurring proteins are the result of an evolutionary process of random variants produced under various selective pressures. Through this process, nature has explored only a small subset of the theoretically possible protein sequence space. Advances in machine learning, particularly deep learning, are catalyzing a revolution in scientific research paradigms. Some deep learning-based methods, especially in structure prediction, are now often combined with higher-resolution physical modeling, outperforming traditional methods. Challenges remain in experimental validation, benchmark labeling, utilizing known physics and interpreting models, and extending to other biomolecules and contexts.

[0288] Protein sites are microenvironments within a protein structure, distinguished by their structural or functional roles. Sites can be defined by their three-dimensional location and the surrounding local neighborhood where the structure or function resides. At the heart of rational protein engineering is understanding how the structural arrangement of amino acids produces functional characteristics within protein sites. Determining the structural and functional roles of individual amino acids within a protein provides information that facilitates the engineering and alteration of protein function. Identifying functionally or structurally important amino acids allows for focused engineering efforts, such as site-directed mutagenesis to alter the functional properties of target proteins. In one specific embodiment, the disclosed technique involves predicting the spatial tolerance of amino acid substitutes. In such embodiments, the disclosed technique includes gapping logic and substitution logic. Gapping logic is configured to remove a specific amino acid at a specific location from a protein, creating an amino acid vacancy at that specific location. Substitution logic is configured to process proteins with amino acid vacancies and score the tolerance of substituted amino acids as candidates for filling / fitting the vacancies. The substitution logic is further configured to score the tolerance of substituted amino acids based at least in part on the structural (or spatial) compatibility between the substituted amino acid and its neighboring amino acids (e.g., right and left amino acids) near the vacancy. The substitution logic assesses how well an amino acid “fits” its surrounding protein environment and indicates that mutations that disrupt strong amino acid preferences are more likely to be harmful. When the substitution logic is a convolutional neural network, during training, the weights of the convolutional filters are optimized to detect local spatial patterns that best capture local biochemical features, thereby separating the microenvironments of 20 amino acids. During training, filters in the convolutional layers of the convolutional neural network are activated when a desired feature appears at a spatial location in the input. Structural (or spatial) compatibility can be defined by altering or influencing protein function. A substituted amino acid is considered structurally (or spatially) incompatible when its substitution at a specific location within the protein structure causes a change in protein function. A substituted amino acid is considered structurally (or spatially) compatible when its substitution at a specific location within the protein structure does not cause a change in protein function. Structural (or spatial) compatibility can be defined by spatial deviation measured by a distance metric. First, the pre-insertion spatial measurement of the protein structure can be determined, for example, by measuring the distance between amino acids in the protein structure before amino acid substitution at a specific position. This distance can be an atomic distance based on the atomic coordinates of the amino acid atoms. The distance between pairs of amino acids can be measured. Then, the post-insertion spatial measurement of the protein structure can be determined, for example, by re-measuring the distance between amino acids in the protein structure after amino acid substitution at a specific position. When the spatial deviation between the pre-insertion and post-insertion spatial measurements exceeds a threshold, the substituted amino acids are considered structurally (or spatially) incompatible.If the spatial deviation between the pre-insertion spatial measurement and the post-insertion spatial measurement does not exceed the threshold, the substituted amino acid is considered to be structurally (or spatially) compatible.

[0289] In another embodiment, the disclosed technique relates to predicting the evolutionary conservation of amino acid substitutes. In such an embodiment, the disclosed technique includes cleavage logic and substitution logic. The cleavage logic is configured to remove a specific amino acid at a specific position from a protein, creating an amino acid vacancy at that specific position in the protein. The substitution logic is configured to process the protein having the amino acid vacancy and score the evolutionary conservation of the substituted amino acid as a candidate to fill the vacancy. The substitution logic is further configured to score the evolutionary conservation of the substituted amino acid based at least in part on the structural (or spatial) compatibility between the substituted amino acid and its neighboring amino acids (e.g., right and left amino acids) near the vacancy. In some embodiments, evolutionary conservation frequencies are used to score evolutionary conservation. In one embodiment, the evolutionary conservation frequencies are based on a position-specific frequency matrix (PSFM). In another embodiment, the evolutionary conservation frequencies are based on a position-specific score matrix (PSSM). In one embodiment, the evolutionary conservation scores of the substituted amino acids are sorted by size.

[0290] In yet another embodiment, the disclosed technique relates to predicting the evolutionary conservation of amino acid substitutions. In this embodiment, the disclosed technique includes vacancy logic and evolutionary conservation prediction logic. The vacancy logic is configured to remove a specific amino acid at a specific position from a protein, creating an amino acid vacancy at that specific position in the protein. The evolutionary conservation prediction logic is configured to process proteins with amino acid vacancies and to rank the evolutionary conservation of substitution amino acids as candidates to fill the vacancies.

[0291] Determination of pathogenicity based on nicked protein spatial representation using target alternative amino acids

[0292] Figure 37 This paper demonstrates a specific implementation of determining the pathogenicity of variants of 3700 target alternative amino acids based on processing the spatial representation of nicked proteins. Proteins are amino acid sequences. A specific amino acid removed or masked from a protein is called a "nicked amino acid." The resulting protein lacking the nicked amino acid is called a "nicked protein" or "vacancy-containing protein."

[0293] The "spatial representation" of a protein characterizes structural information about the amino acids in the protein. A protein's spatial representation can be based on the shape, location, position, pattern, and / or arrangement of the amino acids. Protein spatial representations can be one-dimensional (1D), two-dimensional (2D), three-dimensional (3D), or n-dimensional (nD) information.

[0294] In one specific implementation, the spatial representation of a protein includes the amino acid-based distance channels discussed above, for example, regarding... Figure 6 The amino acid-based distance channel 600 is described. In another specific embodiment, the spatial representation of the protein includes the distance channel tensor discussed above, for example, the one described above regarding... Figure 7 The distance channel tensor described is 700. In yet another specific embodiment, the spatial representation of the protein includes the evolutionary spectrum tensor discussed above, for example, the one described above regarding... Figure 18 The evolutionary spectrum tensor described is 1800. In yet another specific implementation, the spatial representation of the protein includes the voxelized annotation channels discussed above, for example, regarding... Figure 20 The described voxelization annotation channel 2000. In yet another embodiment, the spatial representation of the protein includes the structure confidence channel discussed above. In other embodiments, the spatial representation may also include other channels.

[0295] A "gap-filled spatial representation" of a protein is a spatial representation of a protein excluding at least one gap-filled amino acid. In one embodiment, when generating a gap-filled spatial representation, the gap-filled amino acid is excluded by excluding (or ignoring or disregarding) one or more atoms or types of atoms that contain the gap-filled amino acid. For example, atoms containing the gap-filled amino acid can be excluded from calculations (or selections or computations) that generate distance channels, evolutionary profiles, annotation channels, and / or structural confidence channels. In other embodiments, a gap-filled spatial representation can also be generated by excluding the gap-filled amino acid from other feature channels.

[0296] Consider the following example of generating a nicked spatial representation of a protein by excluding atoms of nicked amino acids from the calculation of amino acid-based distance channels. Figure 5 In the middle, C α A5 The atom belongs to the fifth alanine amino acid in the protein. Now, assume that this fifth alanine amino acid is chosen as the missing amino acid. Then, by considering the center of voxel (1,1) in voxel grid 522 and the nearest α-carbon (C... α The distance between the atoms is 512, and the distance channel is calculated to generate a gapped space representation. The α-carbon atom is the C of the gapped amino acid (i.e., the alanine amino acid at the fifth position). α A5 atom.

[0297] It should also be noted that the terms "spatial representation of a protein" and "protein structure" are used interchangeably in this application. It should also be noted that the terms "spatial representation of a protein with gaps" and "spatial protein structure" are used interchangeably in this application.

[0298] Go to Figure 37In action 3702, protein sequence accessor 3704 accesses the protein that has the corresponding amino acid at the corresponding position.

[0299] In action 3712, the missing amino acid designator 3714 designates a specific amino acid at a specific position in the protein as the missing amino acid and designates the remaining amino acids at the remaining positions in the protein as non-missing amino acids. In one embodiment, the specific amino acid is a reference amino acid that serves as the major allele of the protein.

[0300] In action 3722, the nicked space representation generator 3724 generates a nicked space representation of the protein, which includes the spatial configurations of un-nicked amino acids and spatial configurations excluding nicked amino acids. The spatial configurations of un-nicked amino acids are encoded as amino acid class distance channels. Each amino acid class distance channel in the amino acid class distance channels has a voxel-type distance value for a voxel in a plurality of voxels. The voxel-type distance value specifies the distance from the corresponding voxel in the plurality of voxels to the atom of the un-nicked amino acid. The spatial configuration of the un-nicked amino acid is determined based on the spatial proximity between the corresponding voxel and the atom of the un-nicked amino acid. When determining the voxel-type distance value, the spatial configuration of the nicked amino acid is excluded from the nicked space representation by ignoring the distance from the corresponding voxel to the atom of the nicked amino acid. The spatial configuration of the nicked amino acid is excluded from the nicked space representation by ignoring the spatial proximity between the corresponding voxel and the atom of the nicked amino acid.

[0301] Based on the pan-amino acid conservation frequencies of the amino acids with the closest atoms to the voxel, the spatial configurations of un-notched amino acids are encoded as evolutionary spectrum channels. In one embodiment, when determining the pan-amino acid conservation frequencies, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the closest atoms of the notched amino acids. Based on the per-amino acid conservation frequencies of the corresponding amino acids with the closest corresponding atoms to the voxel, the spatial configurations of un-notched amino acids are encoded as evolutionary spectrum channels. In one embodiment, when determining the per-amino acid conservation frequencies, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the closest corresponding atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as annotation channels. In one embodiment, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the atoms of the notched amino acids when determining the annotation channels. The spatial configurations of un-notched amino acids are encoded as structure confidence channels. In one embodiment, when determining the structure confidence channels, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as additional input channels. In one specific implementation, when determining additional input channels, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0302] In action 3732, pathogenicity determiner 3734 determines the pathogenicity of a nucleotide variant based at least in part on the nicked space representation and the representation of the alternative amino acid produced by the nucleotide variant at a specific position. The representation of the alternative amino acid may be a one-hot encoding of the alternative amino acid (e.g., see...). Figure 8 In some implementations, the alternative amino acid is the same as the reference amino acid. In other implementations, the alternative amino acid is a different amino acid from the reference amino acid.

[0303] Figure 38 An example of the spatial representation of a protein 3800 is shown. This protein contains the amino acid sequence 3804. The aspartic (D) amino acid at position 22 in the amino acid sequence 3804 is selected as the missing amino acid 3802. Figure 39 It shows Figure 38 The nicked space of the protein shown represents an instance of 3900. In Figure 39 In the middle, the missing amino acid 3802 is removed from the missing space representation 3900. Similarly, in... Figure 39 In the figure, the deletion of amino acid 3802 is shown as the deletion of amino acid 3902.

[0304] Figure 40 It shows Figure 38 The atomic space representation of the protein shown is 4000 instances. Figure 40 The atomic 4002 of the missing amino acid 3802 was also depicted. Figure 41 It shows Figure 38 The space representation of the nicked atoms in the protein shown is an example of 4100. In Figure 41 In the notched atomic space representation 4100, atoms 4002 of the notched amino acid 3802 are removed. Similarly, in... Figure 41 In the figure, the deletion of atom 4002 of the nicked amino acid 3802 is shown as the deletion of atom 4102 of the nicked amino acid 3802.

[0305] It should also be noted that the terms "pathogenicity determiner", "pathogenicity predictor", "pathogenicity classifier", "variant pathogenicity classifier", "evolutionary conservation predictor" and "evolutionary conservation determiner" are used interchangeably in this application.

[0306] Figure 42 A specific implementation of pathogenicity classifier 2108 / 2600 / 2700 is shown, which determines the pathogenicity of variants of target alternative amino acids 4200 based on the nicked protein space representation 4202 and the alternative amino acid representation 4212 processed by the target alternative amino acid.

[0307] The pathogenicity classifier 2108 / 2600 / 2700 determines the pathogenicity of nucleotide variants by processing the nicked space representation 4202 and the representation of the alternative amino acids as inputs, and generating a pathogenicity score 4208 of the alternative amino acids as outputs.

[0308] Figure 43 A specific embodiment of training data 4300 for training pathogenicity classifiers 2108 / 2600 / 2700 is described. The pathogenicity classifiers 2108 / 2600 / 2700 are trained on a benign training set 4302. The benign training set 4302 has corresponding benign protein samples 4322, 4342, and 4362 for corresponding reference amino acids at corresponding positions 4312, 4332, and 4352 in the proteome. The reference amino acid is the major allelic amino acid of the proteome. In one embodiment, the proteome has ten million positions, therefore the benign training set 4302 has ten million benign protein samples. Each benign protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid as the corresponding nicked amino acid. Each benign protein sample has a corresponding representation with the corresponding reference amino acid as the corresponding substitute amino acid. In various embodiments, the proteome includes human and non-human proteomes, including non-human primate proteomes.

[0309] Figure 44 A specific implementation is shown to generate the nicked space representations 4322G, 4342G, and 4362G of reference protein samples 4400, 4400, by using reference amino acids 4402, 4412, and 4422 as nicked amino acids, respectively. Figure 45 A specific implementation of training pathogenicity classifiers 2108 / 2600 / 2700 on benign protein samples 4500 is shown.

[0310] The pathogenicity classifiers 2108 / 2600 / 2700 are trained on specific benign protein samples and estimate the pathogenicity of a specific reference amino acid at a specific location in the specific benign protein sample by processing (i) a specific nicked spatial representation 4322G of the specific benign protein sample as input and (ii) a specific reference amino acid representation 4402 as a specific alternative amino acid (e.g., one-hot encoding) and generating a pathogenicity score of the specific reference amino acid as output. A specific nicked spatial representation is generated by using the specific reference amino acid as the nicked amino acid and by using the remaining amino acids at the remaining locations in the specific benign protein sample as the non-nicked amino acids.

[0311] Each of these benign protein samples has a baseline true benign label 4506, which indicates the absolute benignity of the benign protein sample. In one specific implementation, the baseline true benign label is 0, 1, or -1. The pathogenicity score 4502 of a specific reference amino acid is compared with the baseline true benign label to determine an error 4504, and the coefficients of the pathogenicity classifier 2108 / 2600 / 2700 are improved based on this error using training techniques (e.g., backpropagation 4512).

[0312] Pathogenicity classifiers 2108 / 2600 / 2700 are trained on a pathogenicity training set 4308. The pathogenicity training set 4308 contains corresponding pathogenic protein samples 4322A-N, 4342A-N, and 4362A-N, for amino acid substitutions generated by corresponding combinations of reference amino acids 4312, 4332, and 4352 at each corresponding position in the proteome at each of the corresponding positions 4318, 4338, and 4358. In one embodiment, the amino acid substitutions generated by the corresponding combinations are restricted by the accessibility of single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid into an alternative amino acid of an unreachable alternative amino acid class. The amino acid substitutions generated by combinations of specific reference amino acids of a specific amino acid class at a specific position in the proteome include corresponding alternative amino acids of the corresponding amino acid class that are different from the specific amino acid class.

[0313] In one specific implementation, the proteome has ten million positions, with nineteen combinations of amino acid substitutions for each of the ten million positions, thus the pathogenicity training set 4308 has 190 million pathogenic protein samples.

[0314] The corresponding pathogenic protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid as the corresponding nicked amino acid. The corresponding pathogenic protein sample has a corresponding representation of the amino acid substitutions generated by the corresponding combination, as the corresponding substitute amino acid generated at the corresponding position in the proteome by the nucleotide variant generated by the corresponding combination.

[0315] Figure 46A specific implementation of training pathogenicity classifiers 2108 / 2600 / 2700 on a pathogenic protein sample 4600 is shown. The pathogenicity classifiers 2108 / 2600 / 2700 are trained on a specific pathogenic protein sample and estimate the pathogenicity of amino acid substitutions generated by a specific combination of a specific reference amino acid at a specific position in the specific pathogenic protein sample by processing (i) a specific nicked spatial representation 4322G of the specific pathogenic protein sample as input and (ii) a representation 4622 of amino acid substitutions generated by a specific combination of specific alternative amino acids (e.g., one-hot encoding) and generating a pathogenicity score of the amino acid substitutions generated by the specific combination as output. A specific nicked spatial representation is generated by using a specific reference amino acid as the nicked amino acid and by using the remaining amino acids at the remaining positions in the specific pathogenic protein sample as the non-nicked amino acids.

[0316] Each of these pathogenic protein samples has a baseline true pathogenicity label, which indicates the absolute pathogenicity of the pathogenic protein sample. In one specific implementation, the baseline true pathogenicity label is 1, 0, or -1, as long as it differs from (e.g., opposite to) the baseline true benign label. The pathogenicity score 4602 of the amino acid substitutions generated by a specific combination is compared with the baseline true pathogenicity label 4606 to determine an error 4604, and the coefficients of the pathogenicity classifier 2108 / 2600 / 2700 are improved based on this error using training techniques (e.g., backpropagation 4612).

[0317] In one implementation, the pathogenicity classifiers 2108 / 2600 / 2700 are trained over 200 million training iterations. In this implementation, the 200 million training iterations include 10 million iterations with 10 million benign protein samples and 190 million iterations with 190 million pathogenic protein samples. In another implementation, the proteome has between 1 million and 10 million positions, thus the benign training set has between 1 million and 10 million benign protein samples. In yet another implementation, for each of the 1 million to 10 million positions, there are 19 combinatorial amino acid substitutions, thus the pathogenicity training set has between 19 million and 190 million pathogenic protein samples.

[0318] In one specific implementation, the pathogenicity classifiers 2108 / 2600 / 2700 are trained over 20 million to 200 million training iterations. In such implementations, the 20 million to 200 million training iterations include 1 million to 10 million training iterations with 1 million to 10 million benign protein samples, and 19 million to 190 million iterations with 19 million to 190 million pathogenic protein samples.

[0319] Figure 47 This demonstrates how certain unreachable amino acid classes (4700) are masked during training. In action 4702, those unreachable substitute amino acids restricted by single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid to a substitute amino acid of the unreachable substitute amino acid class are masked in the baseline truth marker. In action 4712, the masked amino acid classes result in 0 loss and contribute no to gradient updates. In action 4722, the masked amino acid classes are identified in a lookup table. In action 4732, the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0320] Figure 48 A specific implementation for determining the final pathogenicity score is illustrated. In action 4802, in one implementation, pathogenicity classifiers 2108 / 2600 / 2700 generate a first pathogenicity score for a first substitute amino acid that is the same as the first reference amino acid. In action 4812, in one implementation, pathogenicity classifiers 2108 / 2600 / 2700 generate a second pathogenicity score for a second substitute amino acid that is different from the first reference amino acid. In action 4822, in one implementation, the final pathogenicity score of the second substitute amino acid is the second pathogenicity score of the second substitute amino acid.

[0321] In other alternatives, the final pathogenicity score of the second alternative amino acid is based on a combination of the first and second pathogenicity scores. In the first alternative at 4822a, in one implementation, the final pathogenicity score of the second alternative amino acid is the ratio of the second pathogenicity score to the sum of the first and second pathogenicity scores. In the second alternative at 4822b, in one implementation, the final pathogenicity score of the second alternative amino acid is determined by subtracting the first pathogenicity score from the second pathogenicity score.

[0322] The discussion so far has covered Figure 49A The content described. Figure 49A The text illustrates the variant pathogenicity determination of a target substitute amino acid 4922 that fills a vacancy created by a reference gap amino acid 4902 at a given position in protein 4912. Specifically, this analysis is performed by spatially representing protein 4912 and the vacancy in a 3D format, for example, by calculating a voxelized amino acid categorical distance excluding the reference gap amino acid 4902 (or its atoms).

[0323] The discussion now turns to Figure 49B . Figure 49B The pathogenicity of corresponding variants of amino acids in class 4916 that fill the vacancy created by reference gap amino acid 4902 at a given position in protein 4912 is shown. Figure 49Aand Figure 49B The inputs are the same; only the outputs differ, and the spatial representations of protein 4912 and vacancies in 3D format are also different. Figure 49A In the middle, only one pathogenicity score is generated; while in the middle... Figure 49B In this process, a pathogenicity score is generated for each of the 20 amino acid classes / categories (e.g., by using 20-to-softmax classification).

[0324] Determination of pathogenicity based on nicked protein spatial representation using multiple alternative amino acids

[0325] Figure 50 This demonstrates a specific implementation for determining the pathogenicity of variants with over 5,000 alternative amino acids based on processing the spatial representation of nicked proteins. In action 5002, protein sequence accessor 3704 accesses the protein having the corresponding amino acid at the appropriate position.

[0326] In action 5012, the missing amino acid designator 3714 designates a specific amino acid at a specific position in the protein as the missing amino acid and designates the remaining amino acids at other positions in the protein as non-missing amino acids. In one embodiment, the specific amino acid is a reference amino acid that serves as the major allele of the protein.

[0327] In action 5022, the notched space representation generator 3724 generates a notched space representation of the protein, which includes the spatial configurations of unnotched amino acids and spatial configurations excluding notched amino acids. The spatial configurations of unnotched amino acids are encoded as amino acid class distance channels. Each amino acid class distance channel in the amino acid class distance channels has a voxel-type distance value for a voxel in a plurality of voxels. The voxel-type distance value specifies the distance from the corresponding voxel in the plurality of voxels to the atom of the unnotched amino acid. The spatial configuration of the unnotched amino acid is determined based on the spatial proximity between the corresponding voxel and the atom of the unnotched amino acid. When determining the voxel-type distance value, the spatial configuration of the notched amino acid is excluded from the notched space representation by ignoring the distance from the corresponding voxel to the atom of the notched amino acid. The spatial configuration of the notched amino acid is excluded from the notched space representation by ignoring the spatial proximity between the corresponding voxel and the atom of the notched amino acid.

[0328] Based on the pan-amino acid conservation frequencies of the amino acids with the closest atoms to the voxel, the spatial configurations of un-notched amino acids are encoded as evolutionary spectrum channels. In one embodiment, when determining the pan-amino acid conservation frequencies, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the closest atoms of the notched amino acids. Based on the per-amino acid conservation frequencies of the corresponding amino acids with the closest corresponding atoms to the voxel, the spatial configurations of un-notched amino acids are encoded as evolutionary spectrum channels. In one embodiment, when determining the per-amino acid conservation frequencies, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the closest corresponding atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as annotation channels. In one embodiment, when determining annotation channels, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as structure confidence channels. In one embodiment, when determining structure confidence channels, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as additional input channels. In one specific implementation, when determining additional input channels, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0329] In action 5032, pathogenicity determiner 3734 determines the pathogenicity of a corresponding alternative amino acid at a specific position based at least in part on a nicked spatial representation. The corresponding alternative amino acid is an alternative amino acid generated by a nucleotide variant produced at a specific position from a corresponding combination.

[0330] Figure 51 A specific embodiment of a pathogenicity classifier 2108 / 2600 / 2700 is shown, which determines the pathogenicity of variants of more than 5100 alternative amino acids based on processing a nicked protein spatial representation 5102. The pathogenicity classifier 2108 / 2600 / 2700 determines the pathogenicity of a corresponding alternative amino acid by processing the nicked spatial representation 5102 as input and generating a corresponding pathogenicity score 1-20 for the corresponding amino acid class as output. In some embodiments, the corresponding amino acid class corresponds to a corresponding twenty naturally occurring amino acids. In other embodiments, the corresponding amino acid class corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids. In one embodiment, the output is displayed along with a corresponding ranking of the corresponding pathogenicity scores 1-20 for the corresponding amino acid class.

[0331] Figure 52A concrete implementation of simultaneously training 5200 pathogenicity classifiers 2108 / 2600 / 2700 on both benign and pathogenic protein samples is demonstrated. The pathogenicity classifiers 2108 / 2600 / 2700 are trained on a training set. The training set contains corresponding protein samples for corresponding positions in the proteome. The proteome has ten million positions, therefore the training set contains ten million protein samples. Each corresponding protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid at the corresponding position in the proteome as the corresponding nicked amino acid. The reference amino acid is the major allelic amino acid of the proteome.

[0332] The pathogenicity classifiers 2108 / 2600 / 2700 are trained on specific protein samples and estimate the pathogenicity of the corresponding substitute amino acid for a specific reference amino acid at a specific position in the specific protein sample by processing a specific nicked spatial representation 5202 as input and generating a corresponding pathogenicity score 1-20 for the corresponding amino acid class as output. A specific nicked spatial representation is generated by using the specific reference amino acid as the nicked amino acid and by using the remaining amino acids at the remaining positions in the specific protein sample as the unnicked amino acids.

[0333] Each of these protein samples has a corresponding baseline truth value label for the corresponding amino acid class. The corresponding baseline truth value label includes an absolute benign label for a reference amino acid class within the corresponding amino acid class, and a corresponding absolute pathogenicity label for the corresponding alternative amino acid class within the corresponding amino acid class. In one embodiment, the absolute benign label is 0. The absolute pathogenicity label is the same for the corresponding alternative amino acid class. In one embodiment, the absolute pathogenicity label is 1.

[0334] In one specific implementation, the pathogenicity score of a reference amino acid class is compared with an absolute benign marker (e.g., Figure 52 The pathogenicity score of the reference missing amino acid 5212 (8) and the corresponding pathogenicity scores of the corresponding substitute amino acids are compared with the corresponding absolute pathogenicity markers (e.g., Figure 52 The pathogenicity scores 1-7 and 9-20 in the model are used to determine the error 5204. In one implementation, training techniques (e.g., backpropagation 5224) are used to improve the coefficients of the pathogenicity classifiers 2108 / 2600 / 2700 based on the error.

[0335] In one implementation, the pathogenicity classifier 2108 / 2600 / 2700 is trained with ten million protein samples over ten million training iterations. In some implementations, the proteome has one to ten million locations, thus the training set has one to ten million protein samples. In one implementation, the pathogenicity classifier 2108 / 2600 / 2700 is trained with one to ten million protein samples over one to ten million training iterations.

[0336] In one implementation, pathogenicity classifier 2108 / 2600 / 2700 generates a reference pathogenicity score for the first alternative amino acid of a reference amino acid class. In another implementation, pathogenicity classifier 2108 / 2600 / 2700 generates a corresponding substitution pathogenicity score for the corresponding alternative amino acid of the corresponding alternative amino acid class.

[0337] In one implementation, the corresponding final substitution pathogenicity score of the corresponding alternative amino acid is the corresponding substitution pathogenicity score. In one implementation, the corresponding final substitution pathogenicity score of the corresponding alternative amino acid is based on a corresponding combination of a reference pathogenicity score and the corresponding substitution pathogenicity score. In one implementation, the corresponding final substitution pathogenicity score of the corresponding alternative amino acid is a corresponding ratio of the corresponding substitution pathogenicity score to the sum of the reference pathogenicity score and the corresponding substitution pathogenicity score. In one implementation, the corresponding final substitution pathogenicity score of the corresponding alternative amino acid is determined by subtracting the reference pathogenicity score from the corresponding substitution pathogenicity score, respectively.

[0338] In one implementation, the pathogenicity classifier 2108 / 2600 / 2700 has an output layer that generates corresponding pathogenicity scores. In some implementations, the output layer is a normalization layer. In such implementations, the corresponding pathogenicity scores are normalized. In one implementation, the output layer is a softmax layer. In such implementations, the corresponding pathogenicity scores are exponentially normalized. In another implementation, the output layer has corresponding sigmoid units that generate corresponding pathogenicity scores respectively. In yet another implementation, the corresponding pathogenicity scores are unnormalized.

[0339] Spatial representation of proteins with multiple alternative amino acids based on gaps and pathogenicity determination based on evolutionary conservation

[0340] Evolutionary conservation refers to the presence of similar genes, gene segments, or chromosomal fragments across different species, reflecting both a common origin and the important functional characteristics of conserved elements. Mutations occur spontaneously in each generation, randomly altering amino acids at various locations within a protein. Individuals with mutations that impair key protein functions may experience reduced reproductive capacity. Harmful mutations are lost from the gene pool because individuals carrying these mutations are less efficient at reproduction. Due to the loss of harmful mutations, amino acids essential for protein function are retained in the gene pool. In contrast, harmless (or very rare beneficial) mutations are retained in the gene pool, resulting in variability in non-critical amino acids. Evolutionary conservation in proteins is identified by comparing the amino acid sequences of proteins with the same function from different taxa (orthologs). Predicting the functional consequences of variants relies at least in part on the assumption that key amino acids in a protein family are conserved during evolution due to negative selection (i.e., amino acid changes at these sites have been harmful in the past), and that mutations at these sites increase the likelihood of pathogenicity (causing disease) in humans. Generally, homologous sequences of the target protein are collected and aligned, and a conservation measure is calculated based on the weighted frequencies of different amino acids observed at the target position in the alignment. Figure 53 A specific implementation is shown that determines the pathogenicity of variants with more than 5,300 alternative amino acids based on processing the spatial representation of a nicked protein and generating an evolutionary conservation score for multiple alternative amino acids in response. In action 5302, the nicked amino acid designator 3714 designates a specific amino acid at a specific position in the protein as the nicked amino acid and designates the remaining amino acids at the remaining positions in the protein as non-nicked amino acids. In one implementation, the specific amino acid is a reference amino acid serving as the major allele of the protein.

[0341] In action 5312, the nicked space representation generator 3724 generates a nicked space representation of the protein, which includes the spatial configurations of un-nicked amino acids and spatial configurations excluding nicked amino acids. The spatial configurations of un-nicked amino acids are encoded as amino acid class distance channels. Each amino acid class distance channel in the amino acid class distance channel has a voxel-type distance value for a voxel in a plurality of voxels. The voxel-type distance value specifies the distance from the corresponding voxel in the plurality of voxels to the atom of the un-nicked amino acid. The spatial configuration of the un-nicked amino acid is determined based on the spatial proximity between the corresponding voxel and the atom of the un-nicked amino acid. When determining the voxel-type distance value, the spatial configuration of the nicked amino acid is excluded from the nicked space representation by ignoring the distance from the corresponding voxel to the atom of the nicked amino acid. The spatial configuration of the nicked amino acid is excluded from the nicked space representation by ignoring the spatial proximity between the corresponding voxel and the atom of the nicked amino acid.

[0342] Based on the pan-amino acid conservation frequencies of the amino acids with the closest atoms to the voxel, the spatial configurations of un-notched amino acids are encoded as evolutionary spectrum channels. In one embodiment, when determining the pan-amino acid conservation frequencies, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the closest atoms of the notched amino acids. Based on the per-amino acid conservation frequencies of the corresponding amino acids with the closest corresponding atoms to the voxel, the spatial configurations of un-notched amino acids are encoded as evolutionary spectrum channels. In one embodiment, when determining the per-amino acid conservation frequencies, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the closest corresponding atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as annotation channels. In one embodiment, when determining annotation channels, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as structure confidence channels. In one embodiment, when determining structure confidence channels, the spatial configurations of notched amino acids are excluded from the notched spatial representation by ignoring the atoms of the notched amino acids. The spatial configurations of un-notched amino acids are encoded as additional input channels. In one specific implementation, when determining additional input channels, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0343] In action 5322, the evolutionary conservation determiner 5324 determines the evolutionary conservation of a specific position of a corresponding amino acid in a corresponding amino acid class based at least in part on the gapped space representation.

[0344] Figure 54 An evolutionary conservation determiner 5324 is illustrated according to one embodiment of operation 5400. In some embodiments, the evolutionary conservation determiner 5324 has the same architecture as the pathogenicity classifiers 2108 / 2600 / 2700. The evolutionary conservation determiner 5324 determines evolutionary conservation by processing a nicked spatial representation 5402 as input and generating a corresponding evolutionary conservation score 5406 for the corresponding amino acid 5408 as output. The corresponding evolutionary conservation scores can be sorted by size. For the purposes of this disclosure, "classifier," "determiner," and "included term" may include one or more software modules, one or more hardware modules, or any combination thereof.

[0345] In action 5332, pathogenicity determiner 3734 determines the pathogenicity of corresponding nucleotide variants based at least in part on the evolutionary conservation of corresponding amino acids 5408, which in a protein substitution representation replace specific amino acids with corresponding amino acids 5408 respectively.

[0346] Figure 55One specific implementation of determining pathogenicity based on predicted evolutionary scores is illustrated. When the evolutionary conservation score generated by the evolutionary conservation determiner 5324 for a corresponding amino acid substitution is below a threshold, the classifier 5516 classifies the nucleotide variant as pathogenic 5508. In one specific implementation, when the evolutionary conservation score generated by the evolutionary conservation determiner 5324 for a corresponding amino acid substitution is 0 (i.e., an indication of non-conservatism), the classifier 5516 classifies the nucleotide variant as pathogenic 5508.

[0347] When the evolutionary conservation score generated by the evolutionary conservation determiner 5324 for the corresponding amino acid substitution is higher than a threshold, the classifier 5516 classifies the nucleotide variant as benign 5528. In one specific implementation, when the evolutionary conservation score generated by the evolutionary conservation determiner 5324 for the corresponding amino acid substitution is non-zero (i.e., an indicator of conservation), the classifier 5516 classifies the nucleotide variant as benign 5528.

[0348] Figure 56 A specific implementation of training data 5600 for training the evolutionary conservation determiner 5324 is shown. The evolutionary conservation determiner 5324 is trained on a conserved training set and a non-conserved training set. The conserved training set has corresponding conserved protein samples 5602 for corresponding conserved amino acids at corresponding positions in the proteome. The non-conserved training set has corresponding non-conserved (or unconserved) protein samples 5608 for corresponding non-conserved amino acids at corresponding positions. In various specific implementations, the proteome includes the human proteome and non-human proteomes, including the non-human primate proteome.

[0349] Each of these corresponding positions has a set of conserved amino acids and a set of non-conserved amino acids. A specific conserved amino acid set at a specific position of a specific protein in the proteome includes at least one major allele amino acid observed at that specific position in multiple species. In one embodiment, the major allele amino acid is a reference amino acid (e.g., REF allele 5612 spans benign protein sample 5622, and REF allele 5662 spans benign protein sample 5682). A specific conserved amino acid set includes one or more minor allele amino acids observed at that specific position in multiple species (e.g., ALT allele 5632 observed spanning benign protein samples 5642, 5652, and 5662, and ALT allele 5692 observed spanning benign protein samples 5695 and 5696).

[0350] A specific nonconserved amino acid set at a specific location includes amino acids that are not in a specific conserved amino acid set (e.g., unobserved ALT allele 5618 across pathogenic protein sample 5622A-N and unobserved ALT allele 5668 across pathogenic protein sample 5682A-N).

[0351] In one embodiment, each of these corresponding positions has C conserved amino acids from a conserved amino acid set. In another embodiment, each of these corresponding positions has NC nonconserved amino acids from a nonconserved amino acid set, where NC = 20 - C. The conserved training set has CP conserved protein samples, where CP = number of corresponding positions * C. The nonconserved training set has NCP nonconserved protein samples, where NCP = number of corresponding positions * (20 - C). In one embodiment, C ranges from 1 to 10. In another embodiment, C varies between corresponding positions. In yet another embodiment, C is the same for some of the corresponding positions.

[0352] In one embodiment, the proteome has 10 million to 10 million sites. In such an embodiment, each of the 10 million to 10 million sites has C conserved amino acids from a conserved amino acid set. Each of the 10 million to 10 million sites has NC non-conserved amino acids from a non-conserved amino acid set, where NC = 20 - C. The conserved training set has CP conserved protein samples, where CP = 10 million * C. The non-conserved training set has NCP non-conserved protein samples, where NCP = 10 million * (20 - C).

[0353] In one specific implementation, the evolutionary conservation determiner 5324 is trained over 20 million to 200 million training iterations. In such a specific implementation, the 20 million to 200 million training iterations include 1 million to 10 million training iterations with 1 million to 10 million conserved protein samples, and 19 million to 190 million iterations with 19 million to 190 million non-conserved protein samples.

[0354] In another implementation, the proteome has one million to ten million sites, and therefore the training set has one million to ten million protein samples. In such an implementation, the evolutionary conservation determiner 5324 is trained with one million to ten million protein samples over one million to ten million training iterations.

[0355] The corresponding conserved and non-conserved protein samples have corresponding nicked space representations generated by using the corresponding reference amino acid at the corresponding position as the corresponding nicked amino acid. An evolutionary conservation determiner 5324 is trained on a specific conserved protein sample and estimates the evolutionary conservation of a specific conserved amino acid at a specific position in the specific conserved protein sample by processing the specific nicked space representation of the specific conserved protein sample as input and generating an evolutionary conservation score for the specific conserved amino acid as output. A specific nicked space representation is generated by using the specific reference amino acid at the specific position as the nicked amino acid and by using the remaining amino acids at the remaining positions in the specific conserved protein sample as the non-nicked amino acids.

[0356] Each of these conserved protein samples has a baseline true conservation label. The baseline true conservation label is the frequency of evolutionary conservation. In one implementation, the baseline true conservation label is 1. The evolutionary conservation of a specific conserved amino acid is compared to the baseline true conservation label to determine the error, and a training technique is used to improve the coefficients of the evolutionary conservation determiner 5324 based on this error. In one implementation, the training technique is a gradient update technique based on a loss function (e.g., backpropagation).

[0357] In some implementations, when a particular conserved amino acid is a particular reference amino acid, the baseline truth value conservation marker is masked and not used to determine the error. In such implementations, the masking prevents the evolutionary conservation determiner 5324 from overfitting to the particular reference amino acid.

[0358] The evolutionary conservation determiner 5324 is trained on a specific nonconservative protein sample and estimates the evolutionary conservation of a specific nonconservative amino acid at a specific position in the specific nonconservative protein sample by processing a specific nicked spatial representation of the specific nonconservative protein sample as input and generating an evolutionary conservation score for the specific nonconservative amino acid as output. The specific nicked spatial representation is generated by using a specific reference amino acid at a specific position as the nicked amino acid and by using the remaining amino acids at the remaining positions in the specific nonconservative protein sample as the unnicked amino acids.

[0359] Each of these nonconservative protein samples has a baseline true nonconservative marker. The baseline true nonconservative marker is the frequency of evolutionary conservation. In one specific implementation, the baseline true nonconservative marker is 0. The evolutionary conservation score of a particular nonconservative amino acid is compared to the baseline true nonconservative marker to determine the error, and the coefficients of the evolutionary conservation determiner 5324 are improved based on this error using training techniques (e.g., backpropagation).

[0360] The evolutionary conservation determiner 5324 is trained on a training set. The training set contains corresponding protein samples for corresponding positions in the proteome. Each corresponding protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid at the corresponding position as the corresponding nicked amino acid.

[0361] Figure 57 A specific implementation of training an evolutionary conservation determiner 5700 on both benign and pathogenic protein samples is demonstrated. The evolutionary conservation determiner 5324 is trained on a specific protein sample and estimates the evolutionary conservation of a corresponding amino acid of a corresponding amino acid class at a specific position in the specific protein sample by processing a specific nicked spatial representation 5722 as input to the specific protein sample and generating corresponding evolutionary conservation scores 1-20 as output for the corresponding amino acid. A specific nicked spatial representation 5722 is generated by using a specific reference amino acid at a specific position as the nicked amino acid and by using the remaining amino acids at the remaining positions in the specific protein sample as the unnicked amino acids.

[0362] Each of these protein samples has a corresponding baseline truth marker for the corresponding amino acid. The corresponding baseline truth marker includes one or more conserved (benign) markers for one or more conserved amino acids 5732, 5702, and 5712 in the corresponding amino acid group, and one or more non-conserved (pathogenic) markers for one or more non-conserved amino acids in the corresponding amino acid group. The conserved and non-conserved markers have corresponding evolutionary conservation frequencies. These frequencies can be ordered by magnitude. In one specific implementation, the conserved marker is 1, and the non-conserved marker is 0.

[0363] In one implementation, error 5704 is determined based on a comparison of the evolutionary conservation score of a corresponding conserved amino acid with the corresponding conserved amino acid and a comparison of the evolutionary conservation score of a corresponding non-conserved amino acid with the corresponding non-conserved amino acid. The coefficients of the evolutionary conservation determiner 5324 are then improved using a training technique (e.g., backpropagation 5744) based on this error.

[0364] In one specific implementation, the conserved amino acid includes a specific reference amino acid, and the conservatism marker of the specific reference amino acid is masked and not used to determine the error. This masking prevents the evolutionary conservation determiner 5324 from overfitting to the specific reference amino acid.

[0365] Synonymous mutations are point mutations, meaning they are simply a DNA nucleotide that changes only one base pair in the RNA copy of the DNA. A codon in RNA is a set of three nucleotides that encodes a specific amino acid. Most amino acids have several RNA codons that translate into that specific amino acid. Most of the time, if the third nucleotide mutates, it will result in the same amino acid being encoded. This is called a synonymous mutation because, like grammatical synonyms, the mutated codon has the same meaning as the original codon and therefore does not change the amino acid. If the amino acid doesn't change, the protein is unaffected. Synonymous mutations don't change anything, and therefore nothing changes. This means they have no real role in the evolution of species because the gene or protein isn't altered in any way. Synonymous mutations are actually quite common, but because they have no effect, they often go unnoticed.

[0366] Nonsynonymous mutations have a much greater impact on individuals than synonymous mutations. In nonsynonymous mutations, a single nucleotide insertion or deletion typically occurs in the sequence during transcription, when messenger RNA replicates DNA. This single deletion or addition of a nucleotide results in a frameshift mutation, causing the loss of the entire reading frame of the amino acid sequence and codon confusion. This usually affects the encoded amino acid and alters the resulting protein. The severity of this mutation depends on how early in the amino acid sequence it occurs. If it occurs near the start, the entire protein is altered, which can be a lethal mutation. Another way nonsynonymous mutations occur is through point mutations, which change a single nucleotide to a codon that does not translate into the same amino acid. Often, a change in a single amino acid does not have a significant impact on the protein and is still functional. If it occurs early in the sequence and the codon is changed to translate into a stop signal, the protein will not be made, potentially leading to serious consequences. Sometimes, nonsynonymous mutations are actually positive changes. Natural selection may favor this new expression of the gene, and the individual may have developed an advantageous adaptation from the mutation. If this mutation occurs in the gamete, this adaptation can be inherited by the next generation. Nonsynonymous mutations increase the diversity of the gene pool, making it easier for natural selection to play a role at the micro-evolutionary level and drive evolution.

[0367] The triplet of nucleotides that encodes an amino acid is called a codon. Each set of three nucleotides encodes one amino acid. Because there are 64 possible combinations of four nucleotides, and only 20 amino acids can be represented by three at a time, the codon is degenerate (in most cases, each amino acid has more than one codon). An example of an unreachable class of substitute amino acids is those that are not encoded by synonymous SNPs. Another example of an unreachable class of substitute amino acids is those that are limited by the number of triplet nucleotide mutant combinations that deviate from the start codon at the triplet nucleotide position due to a single nucleotide polymorphism (SNP).

[0368] In one implementation, the unreachable alternative amino acids that are restricted by SNPs in the baseline truth label are masked to transform the reference codon of the reference amino acid into an unreachable alternative amino acid class. In this implementation, the masked amino acid classes result in zero loss and do not contribute to gradient updates. In another implementation, the masked amino acid classes are identified in a lookup table. In yet another implementation, the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0369] Specific conserved and non-conserved amino acid sequences are identified based on the evolutionary conservation profiles of homologous proteins across multiple species. In one embodiment, a position-specific frequency matrix (PSFM) is used to determine the evolutionary conservation profile of homologous proteins. In another embodiment, a position-specific score matrix (PSSM) is used to determine the evolutionary conservation profile of homologous proteins.

[0370] Figure 58 Different specific implementations of the baseline truth label encoding used to train the evolutionary conservation determiner 5324 are described. The baseline truth label encoding 5802 uses evolutionary conservation frequencies (e.g., PSFM or PSSM) to label conserved amino acid classes A, C, and F, and uses a "0 value" to label the remaining non-conserved amino acid classes. The baseline truth label encoding 5812 is identical to the baseline truth label encoding 5802, except that the baseline truth label encoding 5812 "masks" the REF major allele / most conserved amino acid class F, so that the REF major allele / most conserved amino acid class F does not contribute to the training of the evolutionary conservation determiner 5324 (e.g., by zeroing out the loss calculated by the loss function of the REF major allele / most conserved amino acid class F).

[0371] The baseline truth label code 5822 uses "one value" to label conserved amino acid classes A, C, and F, and uses "0 value" to label the remaining non-conserved amino acid classes. The baseline truth label code 5832 is the same as the baseline truth label code 5822, except that the baseline truth label code 5832 "masks" the REF major allele / most conserved amino acid class F, so that the REF major allele / most conserved amino acid class F does not contribute to the training of the evolutionary conservation determiner 5324 (e.g., by zeroing out the loss calculated by the loss function of the REF major allele / most conserved amino acid class F).

[0372] Figure 59 The example PSFM 5900 is shown. Figure 60 An example PSSM 6000 is depicted. Figure 61 A specific implementation for generating PSFM and PSSM is shown. Figure 62 This demonstrates an example of PSFM 6200 encoding. Figure 63 Example PSSM 6300 encoding is depicted.

[0373] Multiple sequence alignment (MSA) is the alignment of multiple homologous protein sequences with a target protein sequence. MSA is an important step in the comparative analysis and property prediction of biological sequences because a great deal of information (e.g., evolutionary clusters and co-evolutionary clusters) is generated from MSA and can be mapped to the selected target sequence or protein structure.

[0374] The sequence profile of a protein sequence X of length L is an L×20 matrix in the form of a PSSM or PSFM. The columns of both PSSM and PSFM are alphabetically indexed by amino acid, with each row corresponding to a position within the protein sequence. PSSM and PSFM contain the substitution fraction and frequency of amino acids at different positions within the protein sequence, respectively. Each row of a PSFM is normalized to a total of 1. The sequence profile of protein sequence X is calculated by comparing X with multiple sequences in a protein database that have statistically significant sequence similarity to X. Therefore, the sequence profile contains more general evolutionary and structural information about the protein family to which protein sequence X belongs, thus providing valuable information for long-range homology detection and fold identification.

[0375] A protein sequence (called the query sequence, e.g., the reference amino acid sequence of the protein) can be used as a seed to search for and align homologous sequences from a protein database (e.g., SWISSPROT) using a procedure such as PSI-BLAST. Aligned sequences share some homologous fragments and belong to the same protein family. The aligned sequences are further transformed into two spectra to express their homology information: PSSM and PSFM. Both PSSM and PSFM are matrices with 20 rows and L columns, where L is the total number of amino acids in the query sequence. Each column of the PSSM represents the log-likelihood of residue substitution at the corresponding position in the query sequence. The (i,j)th element of the PSSM matrix represents the probability that the amino acid at position j in the query sequence mutates to amino acid type i during evolution. The PSFM contains the weighted observation frequency for each position in the aligned sequence. Specifically, the (i,j)th element of the PSFM matrix represents the probability that position j in the query sequence has amino acid type i.

[0376] Given a query sequence, we first obtain its sequence profile by presenting it to PSI-BLAST to search and align homologous protein sequences from protein databases (e.g., the Swiss-Prot database). Figure 61 The procedure for obtaining sequence profiles using the PSI-BLAST program is shown. The parameters h and j in PSI-BLAST are typically set to 0.001 and 3, respectively. The protein sequence profile encapsulates homology information related to the queried protein sequence. In PSI-BLAST, homology information is represented by two matrices: PSFM and PSSM. Examples of PSFM and PSSM are shown below. Figure 62 and Figure 63 As shown.

[0377] exist Figure 62 In this context, the (1, u)th element (I∈{1,2,…,Li}, u∈{1,2,…,20}) represents the probability of having the uth amino acid at the first position of the query protein. For example, the probability of having amino acid M at the first position of the query protein is 0.36.

[0378] exist Figure 63 In the expression, the (1,u)th element (I∈{1,2,…,Li},u∈{1,2,…,20}) represents the possible fractions of the amino acid at position 1 of the query protein mutating to the uth amino acid during evolution. For example, the amino acid V at position 1 of the query protein mutated to H during evolution has a fraction of -3, while the fraction at position 8 is -4.

[0379] Combinatorial learning and transfer learning

[0380] Figure 64Two datasets are shown, and the model disclosed in this paper can be learned, for example, through a combination learning approach. Figures 65A to 65B ) or through transfer learning ( Figures 66A to 66B Training was conducted on these two datasets. The first training dataset is called Jigsaw AI dataset 6406. The second training dataset is called Primate AI dataset 6408. As mentioned above, Jigsaw AI dataset 6406 is characterized by voxel input 6412 having a missing central residue that is identified as a missing amino acid. Primate AI dataset 6408 is characterized by voxel input 6412 having no missing residues and a complete input.

[0381] For the JigsawAI dataset 6406, the baseline ground truth marker 6422 has missing or masked markers 6426 for the nicked amino acids (e.g., REF amino acids). For the PrimateAI dataset 6408, the baseline ground truth marker 6422 has nineteen missing or masked markers 6436 for the remaining amino acids that are different from the alternative amino acids (benign or pathogenic) to be analyzed. In one specific implementation, the number of samples 6432 in the JigsawAI dataset 6406 is ten million 6436, and the number in the PrimateAI dataset 6408 is one million 6438.

[0382] Figures 65A to 65B This paper demonstrates a specific implementation of the combinatorial learning 6500 of the model disclosed herein. In action 6502, a gapped training set is accessed. This gapped training set is also referred to herein as the Jigsaw AI dataset 6406. The gapped training set comprises corresponding gapped protein samples for corresponding positions in the proteome. Each corresponding gapped protein sample is labeled with a corresponding gapped benchmark truth sequence. The specific gapped benchmark truth sequence of a particular gapped protein sample has a benign label corresponding to a reference amino acid at a specific position in the particular gapped protein, and a corresponding pathogenic label corresponding to the remaining amino acid class corresponding to the substitute amino acid at the specific position.

[0383] In action 6512, the non-gap training set is accessed. The non-gap training set is also referred to in this paper as the PrimateAI dataset 6408. The non-gap training set includes non-gap benign protein samples and non-gap pathogenic protein samples. Specific non-gap benign protein samples include benign alternative amino acids substituted by benign nucleotide variants at specific positions. Specific non-gap pathogenic protein samples include pathogenic alternative amino acids substituted by pathogenic nucleotide variants at specific positions. Specific non-gap benign protein samples are labeled with a benign benchmark truth sequence having a benign label for a specific amino acid class corresponding to a benign alternative amino acid and a corresponding masking label for the corresponding remaining amino acid class corresponding to an amino acid different from the benign alternative amino acid. Specific non-gap pathogenic protein samples are labeled with a pathogenic benchmark truth sequence having a pathogenic label for a specific amino acid class corresponding to a pathogenic alternative amino acid and a corresponding masking label for the corresponding remaining amino acid class corresponding to an amino acid different from the pathogenic alternative amino acid.

[0384] In one implementation, a benign marker corresponding to a reference amino acid at a specific position in a specific nicked protein is masked. In one implementation, the non-nicked benign protein sample is derived from common human and non-human primate nucleotide variants. In one implementation, the non-nicked pathogenic protein sample is derived from combinatorially mimicked nucleotide variants.

[0385] In action 6522, the corresponding notched space representation of the notched protein sample is generated, and the corresponding unnotched space representation of the unnotched benign protein sample and the unnotched pathogenic protein sample are generated.

[0386] In action 6532, the pathogenicity classifier 2108 / 2600 / 2700 is trained in one or more training loops, and a trained pathogenicity classifier 2108 / 2600 / 2700 is generated as the parameters / coefficients / weights of the trained pathogenicity classifier 2108 / 2600 / 2700 are optimized. Each of these training loops uses a gleaned space representation from the corresponding gleaned space representation and a non-gleaned space representation from the corresponding non-gleaned space representation as training examples.

[0387] In action 6542, the trained pathogenicity classifiers 2108 / 2600 / 2700 are used to determine the pathogenicity of the variant.

[0388] In one specific implementation, the sample indicator is used to indicate to the pathogenicity classifier 2108 / 2600 / 2700 whether the current training example is a notched spatial representation of a notched protein sample or a non-notched spatial representation of a non-notched protein sample.

[0389] In one implementation, the pathogenicity classifier 2108 / 2600 / 2700 generates an amino acid-type output sequence in response to processing training examples. The amino acid-type output sequence has an amino acid-type pathogenicity score.

[0390] In one implementation, the performance of the trained pathogenicity classifier 2108 / 2600 / 2700 is measured between training cycles on a validation set. In some implementations, the validation set comprises a pair of notched and unnotched spatial representations for each retained protein sample.

[0391] In one embodiment, a trained pathogenicity classifier 2108 / 2600 / 2700 generates a first amino acid class output sequence for the nicked space representation of the pair, and a second amino acid class output sequence for the unnicked space representation of the pair. In some embodiments, a final pathogenicity score for the nucleotide variant causing the amino acid substitution in the retained protein sample is determined based on a combination of first and second pathogenicity scores for amino acid substitutions in the first and second amino acid class output sequences. In other embodiments, the final pathogenicity score is based on the average of the first and second pathogenicity scores.

[0392] In some implementations, at least some of these training cycles use the same number of gapped and ungapped space representations. In other implementations, at least some of these training cycles use training example batches with the same number of gapped and ungapped space representations.

[0393] In one implementation, the masked labels do not contribute to error determination and therefore do not contribute to the training of the pathogenicity classifiers 2108 / 2600 / 2700. In some implementations, the masked labels are cleared to zero.

[0394] In some implementations, the gapped and non-gapped spatial representations are weighted differently, such that the contribution of the gapped spatial representation to the gradient updates of the parameters of the pathogenicity classifiers 2108 / 2600 / 2700 in response to the processing of the non-gapped spatial representation differs from the contribution of the non-gapped spatial representation to the gradient updates of the parameters of the pathogenicity classifiers 2108 / 2600 / 2700 in response to the processing of the non-gapped spatial representation. In one implementation, the variation is determined by predefined weights.

[0395] Figures 66A to 66B Demonstrates the use of transfer learning 6600 and... Figure 64The two datasets shown are used to train a specific implementation of the model disclosed in this paper. In action 6602, the pathogenicity classifiers 2108 / 2600 / 2700 are first trained on the gapped training set (i.e., the Jigsaw AI dataset 6406) to generate trained pathogenicity classifiers 2108 / 2600 / 2700.

[0396] In action 6612, the trained pathogenicity classifiers 2108 / 2600 / 2700 are further trained on the non-gap training set (i.e., the PrimateAI dataset 6408) to generate retrained pathogenicity classifiers 2108 / 2600 / 2700.

[0397] In action 6622, the retrained pathogenicity classifiers 2108 / 2600 / 2700 are used to determine the pathogenicity of the variant.

[0398] In action 6632, the performance of the trained pathogenicity classifiers 2108 / 2600 / 2700 is measured between training loops on a first validation set, which includes only the un-notched spatial representations of the retained protein samples. In another embodiment, the performance of the retrained pathogenicity classifiers 2108 / 2600 / 2700 is measured between training loops on a second validation set, which includes both notched and unnotched spatial representations of the retained protein samples.

[0399] In action 6642, the retrained pathogenicity classifiers 2108 / 2600 / 2700 generate a first amino acid class output sequence for the pair in response to processing the pair. In one embodiment, based on the first amino acid class output sequence, a final pathogenicity score is determined for the nucleotide variant that causes amino acid substitutions in the corresponding retained protein sample.

[0400] Generate training data and training labels

[0401] Figure 67 A specific implementation for generating 6700 training data points and labels to produce the model disclosed herein is shown.

[0402] The proteome accessor 6704 accesses multiple amino acid positions in a proteome containing multiple proteins.

[0403] Reference designator 6714 designates the major allele amino acid at multiple amino acid positions as the reference amino acid for multiple proteins.

[0404] For each of the multiple amino acid positions, benign labeler 6724 classifies those nucleotide substitutions as benign variants, which replace a specific reference amino acid at a specific amino acid position in a specific substitution representation of a specific protein.

[0405] For each of the multiple amino acid positions, pathogenicity marker 6734 classifies those nucleotide substitutions as pathogenic variants, which replace a specific reference amino acid with an alternative amino acid at that specific amino acid position. The alternative amino acid is different from the specific reference amino acid.

[0406] The trainer 6744 trains the variant pathogenicity classifier 2108 / 2600 / 2700 on training data including spatial representations of protein samples, such that the spatial representations are assigned a baseline ground truth benign label corresponding to benign variants and a baseline ground truth pathogenic label corresponding to pathogenic variants.

[0407] In one embodiment, a variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a substitution of a second amino acid for a first amino acid at a given amino acid position in a protein is pathogenic or benign. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a pathogenicity score for the substitution. In another embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a corresponding substitution of a corresponding amino acid for a first amino acid at a given amino acid position in a protein is pathogenic or benign. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a corresponding pathogenicity score for the corresponding substitution. In some embodiments, the corresponding amino acid corresponds to one of twenty naturally occurring amino acids. In other embodiments, the corresponding amino acid corresponds to a subset of the twenty naturally occurring amino acids.

[0408] In one embodiment, a variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether an insertion of an amino acid at a given empty amino acid position in a protein is pathogenic or benign. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a pathogenicity score for the insertion. In another embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a corresponding insertion of a corresponding amino acid at a given empty amino acid position in a protein is pathogenic or benign. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a corresponding pathogenicity score for the corresponding insertion. In some embodiments, the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids. In other embodiments, the corresponding amino acid corresponds to a subset of the twenty naturally occurring amino acids.

[0409] In one embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a substitution of a second amino acid for a first amino acid at a given amino acid position in a protein is spatially tolerated by other amino acids in the protein. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a spatial tolerance score for the substitution. In another embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a corresponding substitution of a corresponding amino acid for a first amino acid at a given amino acid position in a protein is spatially tolerated by other amino acids in the protein. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a corresponding spatial tolerance score for the corresponding substitution. In some embodiments, the corresponding amino acid corresponds to one of twenty naturally occurring amino acids. In other embodiments, the corresponding amino acid corresponds to a subset of the twenty naturally occurring amino acids.

[0410] In one embodiment, a variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether an insertion of an amino acid at a given empty amino acid position in a protein is spatially tolerated by other amino acids in that protein. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a spatial tolerance score for the insertion. In another embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a corresponding insertion of a corresponding amino acid at a given empty amino acid position in a protein is spatially tolerated by other amino acids in that protein. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a corresponding spatial tolerance score for the corresponding insertion. In some embodiments, the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids. In other embodiments, the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0411] In one embodiment, a variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a substitution of a second amino acid for a first amino acid at a given amino acid position in a protein is evolutionarily conserved or non-conserved. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate an evolutionary conservation score for the substitution. In another embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a corresponding substitution of a corresponding amino acid for a first amino acid at a given amino acid position in a protein is evolutionarily conserved or non-conserved. In such an embodiment, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a corresponding evolutionary conservation score for the corresponding substitution. In some embodiments, the corresponding amino acid corresponds to one of twenty naturally occurring amino acids. In other embodiments, the corresponding amino acid corresponds to a subset of the twenty naturally occurring amino acids.

[0412] In one implementation, a variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether an insertion of an amino acid at a given empty amino acid position in a protein is evolutionarily conserved or non-conserved. In such an implementation, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate an evolutionary conservation score for the insertion.

[0413] In one implementation, a variant pathogenicity classifier 2108 / 2600 / 2700 is trained to determine whether a corresponding insertion of a corresponding amino acid at a given empty amino acid position in a protein is evolutionarily conserved or non-conserved. In such implementations, the variant pathogenicity classifier 2108 / 2600 / 2700 is trained to generate a corresponding evolutionary conservation score for the corresponding insertion. In some implementations, the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids. In other implementations, the corresponding amino acid corresponds to a subset of the twenty naturally occurring amino acids.

[0414] In different embodiments, steric tolerance corresponds to structural tolerance, and steric intolerance corresponds to structural intolerance. In different embodiments, the number of amino acid positions ranges from one million to ten million amino acid positions. In different embodiments, the number of amino acid positions ranges from ten million to one hundred million amino acid positions. In different embodiments, the number of amino acid positions ranges from one hundred million to one billion amino acid positions. In different embodiments, the number of amino acid positions ranges from one million to one million amino acid positions.

[0415] In one implementation, unreachable substitution amino acids that are restricted by single nucleotide polymorphisms (SNPs) to convert the reference codon of a reference amino acid into an unreachable substitution amino acid class are masked in the baseline truth marker. In this implementation, the masked amino acid classes result in zero loss and do not contribute to gradient updates. In another implementation, the masked amino acid classes are identified in a lookup table. In yet another implementation, the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0416] In different implementations, spatial representation is a structural representation of the protein structure of a protein sample. In different implementations, voxelization is used to encode the spatial representation.

[0417] Pathogenicity determined

[0418] Figure 68 A specific embodiment of method 6800 for determining the pathogenicity of nucleotide variants is illustrated. The method includes, in action 6802, accessing a spatial representation of a protein. The spatial representation of a protein specifies the corresponding spatial conformation of a corresponding amino acid at a corresponding position in the protein.

[0419] The method includes, in action 6812, removing a specific spatial conformation of a specific amino acid at a specific location from the spatial representation of the protein, thereby generating a nicked spatial representation of the protein. In one implementation, the removal of the specific spatial conformation is performed by a script (or automatically).

[0420] The method includes, in action 6822, determining the pathogenicity of a nucleotide variant based at least in part on the nicked spatial representation and the representation of the alternative amino acid produced by the nucleotide variant at a specific position.

[0421] Structural tolerance prediction

[0422] Figure 69 A specific implementation of system 6900 for predicting the structural tolerance of amino acid substitutions is shown. In action 6902, cleavage logic is configured to remove a specific amino acid at a specific position from the spatial representation of the protein and to create an amino acid vacancy at the specific position in the spatial representation of the protein.

[0423] In action 6912, the structural tolerance prediction logic is configured to process the spatial representation of proteins with amino acid vacancies and rank the structural tolerance of alternative amino acids as candidates for filling amino acid vacancies based on the co-occurrence patterns of amino acids near the amino acid vacancies.

[0424] Performance results as objective indicators of creativity and non-obviousness

[0425] The variant pathogenicity classifier disclosed in this paper predicts pathogenicity based on 3D protein structure and is referred to as "PrimateAI 3D". "Primate AI" is a co-owned and previously disclosed variant pathogenicity classifier that predicts pathogenicity based on protein sequence. Additional details about PrimateAI can be found in co-owned U.S. patent applications 16 / 160,903, 16 / 160,986, 16 / 160,968 and 16 / 407,149, and Sundaram, L. et al., Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018).

[0426] A variant pathogenicity classifier trained using the transfer learning techniques disclosed in this paper. Figures 66A to 66B This is called "transfer learning." A variant pathogenicity classifier trained using the combinatorial learning technique disclosed in this paper ( Figures 65A to 65B This is called "combinatorial learning".

[0427] Figure 70A , Figure 70B and Figure 70CThe performance results were generated on a classification task that accurately distinguishes benign variants from pathogenic variants across multiple validation sets. Novel Developmental Delay Disorder (Novel DDD) is an example of a validation set used to compare the classification accuracy of transfer learning and combinatorial learning with Primate AI 3D and Primate AI. The Novel DDD validation set labels variants from individuals with DDD as pathogenic and the same variants from healthy relatives of individuals with DDD as benign. A similar labeling scheme was used for the Autism Spectrum Disorder (ASD) validation set.

[0428] BRCA1 is another example of a validation set used to compare the classification accuracy of transfer learning and combinatorial learning with that of Primate AI 3D and Primate AI. The BRCA1 validation set labels reference amino acid sequences generated from the synthesis of proteins simulating the BRCA1 gene as benign variants and alleles with altered amino acid sequences from the synthesis of proteins simulating the BRCA1 gene as pathogenic variants. Similar labeling schemes are used for the TP53 gene, TP53S3 gene and its variants, and... Figure 70A , Figure 70B and Figure 70C The other genes and their variants shown are different validation sets.

[0429] exist Figure 70A , Figure 70B and Figure 70C In this context, the y-axis has p-values, and the x-axis has different validation sets. For example... Figure 70A , Figure 70B and Figure 70C As shown by the p-values, ensemble learning generally outperforms other methods, followed by transfer learning, and then PrimateAI 3D. A larger p-value, i.e., a longer bar, indicates higher accuracy in distinguishing between benign and pathogenic variants. Figure 70A , Figure 70B and Figure 70C In this study, the bar for combinatorial learning is consistently longer than the bars for other methods.

[0430] In addition, Figure 70A , Figure 70B and Figure 70C In the graph, the individual “average” chart is calculated as the average of the p-values ​​determined for each of these validation sets. Also in the average chart, ensemble learning generally outperforms other methods, followed by transfer learning, and then PrimateAI 3D, as indicated by the fact that the horizontal bar for ensemble learning is consistently longer than the horizontal bars for other methods.

[0431] The average statistic may be biased due to outliers. To address this issue, Figure 70A , Figure 70B and Figure 70C The document also depicts a separate "Method Rank" chart. Higher ranks indicate lower classification accuracy. Also in the Method Rank chart, ensemble learning typically outperforms other methods, followed by transfer learning, and then PrimateAI 3D. In the Method Rank chart, more counts with lower ranks 1 and 2 outperform counts with higher ranks 3.

[0432] Terms and Conditions

[0433] The technology disclosed in this invention can be practiced as a system, method, or article of manufacture. One or more features of a specific embodiment can be combined with a basic embodiment. Non-exclusive embodiments are taught to be composable. One or more features of a specific embodiment can be combined with other embodiments. This disclosure periodically reminds the user of these options. The omission of statements repeating these options from some embodiments should not be considered as limiting the combinations taught in the foregoing, and these statements are hereby incorporated by reference into each of the following embodiments.

[0434] One or more embodiments and terms or elements thereof of the disclosed technology may be implemented in the form of a computer product including a non-transitory computer-readable storage medium having computer-usable program code for performing the indicated method steps. Furthermore, one or more embodiments and terms or elements thereof of the disclosed technology may be implemented in the form of an apparatus including memory and at least one processor coupled to the memory, and operate to perform the exemplary method steps. Further, in another aspect, one or more embodiments and terms or elements thereof of the disclosed technology may be implemented in the form of an apparatus for performing one or more method steps described herein; the apparatus may include (i) a hardware module, (ii) a software module executing on one or more hardware processors, or (iii) a combination of hardware and software modules; any one of (i) to (iii) implements the particular technology set forth herein, and the software module is stored in a computer-readable storage medium (or multiple such media).

[0435] The clauses described in this section can be combined to form features. For the sake of brevity, combinations of features are not listed separately and are not repeated with each group of basic features. The reader will understand how easily the features identified in the clauses described in this section can be combined with the set of basic features identified as specific implementations in other parts of this application. These clauses are not intended to be mutually exclusive, exhaustive, or limiting; and the disclosed technology is not limited to these clauses but includes all possible combinations, modifications, and variations within the scope of the claimed technology and its equivalents.

[0436] Other specific implementations of the provisions described in this section may include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the provisions described in this section. Yet another specific implementation of the provisions described in this section may include a system comprising a memory and one or more processors operable to execute instructions stored in the memory, thereby performing any of the provisions described in this section.

[0437] This invention discloses the following provisions:

[0438] Clause Set 1 (ILLM 1050-2)

[0439] 1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, the computer-implemented method comprising: accessing a protein having a corresponding amino acid at a corresponding position;

[0440] A specific amino acid at a specific position in the protein is designated as a missing amino acid, and the remaining amino acids at the remaining positions in the protein are designated as non-missing amino acids;

[0441] The nicked space representation of the protein is generated, the nicked space representation

[0442] Including the spatial configuration of the non-notched amino acids, and

[0443] Excluding the spatial configuration of the missing amino acid; and

[0444] At least in part based on

[0445] The gapped space representation and

[0446] The pathogenicity of a nucleotide variant is determined by the representation of the alternative amino acid produced by the nucleotide variant at the specific position.

[0447] 2. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an amino acid-type distance channel.

[0448] Each amino acid type distance channel in the aforementioned amino acid type distance channels has voxel-type distance values ​​for voxels in a plurality of voxels, and

[0449] The voxel-based distance value specifies the distance from the corresponding voxel among the plurality of voxels to the atom of the non-notched amino acid.

[0450] 3. The computer-implemented method according to Clause 2, wherein the spatial configuration of the unnotched amino acid is determined based on the spatial proximity between the corresponding voxel and the atoms of the unnotched amino acid.

[0451] 4. The computer-implemented method according to Clause 2, wherein when determining the voxel distance value, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the distance from the corresponding voxel to the nicked amino acid atom.

[0452] 5. The computer-implemented method according to Clause 4, wherein the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the spatial proximity between the corresponding voxel and the atoms of the nicked amino acid.

[0453] 6. The computer-implemented method according to Clause 1, wherein the specific amino acid is a reference amino acid serving as the major allele of the protein.

[0454] 7. The computer-implemented method according to Clause 1, wherein the pathogenicity predictor determines the pathogenicity of the nucleotide variant by:

[0455] Processing the nicked spatial representation as input and

[0456] The representation of the alternative amino acids; and

[0457] The pathogenicity fraction of the alternative amino acid is generated as an output.

[0458] 8. The computer-implemented method according to Clause 7, wherein the pathogenicity predictor is trained on a benign training set.

[0459] 9. The computer-implemented method according to Clause 8, wherein the benign training set has corresponding benign protein samples for corresponding reference amino acids at corresponding positions in the proteome.

[0460] 10. The computer-implemented method according to Clause 9, wherein the reference amino acid is the major allele amino acid of the proteome.

[0461] 11. The computer-implemented method according to Clause 10, wherein the proteome has ten million locations, and thus the benign training set has ten million benign protein samples.

[0462] 12. The computer-implemented method according to Clause 11, wherein the corresponding benign protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid as the corresponding nicked amino acid.

[0463] 13. The computer-implemented method according to Clause 12, wherein the corresponding benign protein sample has a corresponding representation of the corresponding reference amino acid as a corresponding alternative amino acid.

[0464] 14. The computer-implemented method according to Clause 13, wherein the pathogenicity predictor is trained on a specific benign protein sample and estimates the pathogenicity of a specific reference amino acid at a specific location in the specific benign protein sample by:

[0465] Processing the input

[0466] (i) A specific nicked space representation of the particular benign protein sample.

[0467] The specific notched spatial representation is generated in the following manner.

[0468] Using the specific reference amino acid as the missing amino acid, and

[0469] The remaining amino acids at the remaining positions in the specific benign protein sample are used as non-missing amino acids, and

[0470] (ii) Representation of the specific reference amino acid as a specific alternative amino acid; and generation of the pathogenicity score of the specific reference amino acid as an output.

[0471] 15. The computer-implemented method according to Clause 14, wherein each of the benign protein samples has a baseline truth benign marker indicating the absolute benignity of the benign protein sample.

[0472] 16. The computer-implemented method according to Clause 15, wherein the baseline truth value benign flag is 0.

[0473] 17. The computer-implemented method according to Clause 16, wherein the pathogenicity score of the particular reference amino acid is compared with the baseline true benign label to determine an error, and a training technique is used to improve the coefficients of the pathogenicity predictor based on the error.

[0474] 18. The computer-implemented method according to Clause 1, wherein the pathogenicity predictor is trained on a pathogenicity training set.

[0475] 19. The computer-implemented method according to Clause 18, wherein the pathogenicity training set has amino acid substitutions generated for each reference amino acid of the reference amino acids at each of the corresponding positions in the proteome.

[0476] 20. The computer-implemented method according to Clause 19, wherein the combination of a specific reference amino acid of a specific amino acid class at a specific location in the proteome generates an amino acid substitution comprising a corresponding alternative amino acid of a corresponding amino acid class different from the specific amino acid class.

[0477] 21. The computer-implemented method according to Clause 20, wherein the proteome has ten million positions, wherein for each of the ten million positions there are nineteen combined amino acid substitutions, and thus the pathogenicity training set has one hundred and ninety million pathogenic protein samples.

[0478] 22. The computer-implemented method according to Clause 21, wherein the corresponding pathogenic protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid as the corresponding nicked amino acid.

[0479] 23. The computer-implemented method according to Clause 22, wherein the corresponding pathogenic protein sample has a corresponding representation of the amino acid substitutions generated by the corresponding combination, as corresponding alternative amino acids generated at the corresponding positions in the proteome by nucleotide variants generated by the corresponding combination.

[0480] 24. The computer-implemented method according to Clause 23, wherein the pathogenicity predictor is trained on a specific pathogenic protein sample and estimates the pathogenicity of amino acid substitutions resulting from a specific combination of reference amino acids at a specific location in the specific pathogenic protein sample by:

[0481] Processing the input

[0482] (i) A specific nicked space representation of the specific pathogenic protein sample.

[0483] The specific notched spatial representation is generated in the following manner.

[0484] Using the specific reference amino acid as the missing amino acid, and

[0485] The remaining amino acids at the remaining positions in the specific pathogenic protein sample are used as non-missing amino acids, and

[0486] (ii) Representation of amino acid substitutions resulting from the specific combination of amino acids as specific alternatives; and

[0487] The pathogenicity fraction of the amino acid substitutions generated by the specific combination is generated as output.

[0488] 25. The computer-implemented method according to Clause 24, wherein each pathogenic protein sample in the pathogenic protein sample has a baseline truth pathogenicity marker indicating the absolute pathogenicity of the pathogenic protein sample.

[0489] 26. The computer-implemented method according to Clause 25, wherein the baseline truth value pathogenicity marker is 1.

[0490] 27. The computer-implemented method according to Clause 26, wherein the pathogenicity score of the amino acid substitutions generated by the specific combination is compared with the baseline true pathogenicity label to determine an error, and the training technique is used to improve the coefficients of the pathogenicity predictor based on the error.

[0491] 28. The computer-implemented method according to Clause 27, wherein the pathogenicity predictor is trained over two hundred million training iterations.

[0492] The two hundred million training iterations mentioned above include

[0493] Ten million training iterations using the ten million benign protein samples, and

[0494] 190 million iterations were performed using the aforementioned 190 million pathogenic protein samples.

[0495] 29. The computer-implemented method according to Clause 10, wherein the proteome has one million to ten million sites, and therefore the benign training set has one million to ten million benign protein samples.

[0496] There are nineteen combinations of amino acid substitutions at each of the one million to ten million positions, so the pathogenicity training set has nineteen million to one hundred and ninety million pathogenic protein samples.

[0497] 30. The computer-implemented method according to Clause 29, wherein the pathogenicity predictor is trained over 20 million to 200 million training iterations.

[0498] The aforementioned 20 million to 200 million training iterations include

[0499] One million to ten million training iterations using the aforementioned one million to ten million benign protein samples, and

[0500] Nineteen million to one hundred and ninety million iterations were performed using the aforementioned nineteen million to one hundred and ninety million pathogenic protein samples.

[0501] 31. The computer-implemented method according to Clause 6, wherein the alternative amino acid is the same amino acid as the reference amino acid.

[0502] 32. The computer-implemented method according to Clause 31, wherein the alternative amino acid is an amino acid different from the reference amino acid.

[0503] 33. The computer-implemented method according to Clause 32, wherein the pathogenicity predictor generates a first pathogenicity score for a first substitute amino acid that is identical to a first reference amino acid.

[0504] The pathogenicity predictor generates a second pathogenicity score for a second alternative amino acid that is different from the first reference amino acid.

[0505] 34. The computer-implemented method according to Clause 33, wherein the final pathogenicity fraction of the second alternative amino acid is the second pathogenicity fraction.

[0506] 35. The computer-implemented method according to Clause 34, wherein the final pathogenicity score of the second alternative amino acid is based on a combination of the first pathogenicity score and the second pathogenicity score.

[0507] 36. The computer-implemented method according to Clause 35, wherein the final pathogenicity fraction of the second alternative amino acid is the ratio of the second pathogenicity fraction to the sum of the first pathogenicity fraction and the second pathogenicity fraction.

[0508] 37. The computer-implemented method according to Clause 36, wherein the final pathogenicity fraction of the second alternative amino acid is determined by subtracting the first pathogenicity fraction from the second pathogenicity fraction.

[0509] 38. The computer-implemented method according to Clause 1, wherein the spatial configuration of the non-notched amino acid is encoded as an evolutionary spectrum channel based on the pan-amino acid conserved frequency of the amino acid having the atom closest to the voxel.

[0510] 39. The computer-implemented method according to Clause 38, wherein when determining the conservative frequency of the ubiquitin amino acid, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the nearest atom of the nicked amino acid.

[0511] 40. The computer-implemented method according to Clause 1, wherein the spatial configuration of the non-notched amino acid is encoded as an evolutionary spectrum channel based on the per-amino acid conserved frequency of the corresponding amino acid having the corresponding atom closest to the voxel.

[0512] 41. The computer-implemented method according to Clause 40, wherein when determining the conservative frequency per amino acid, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the corresponding nearest atom of the nicked amino acid.

[0513] 42. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an annotation channel.

[0514] 43. The computer-implemented method according to Clause 42, wherein when the annotation channel is determined, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0515] 44. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as a structure confidence channel.

[0516] 45. The computer-implemented method according to Clause 44, wherein when determining the structural confidence channel, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0517] 46. ​​The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an additional input channel.

[0518] 47. The computer-implemented method according to Clause 46, wherein when the additional input channel is determined, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0519] 48. The computer-implemented method according to Clause 9, wherein the proteome includes human proteomes and non-human proteomes, including non-human primate proteomes.

[0520] 49. The computer-implemented method according to Clause 7, wherein in the reference truth mark, those unreachable alternative amino acids of the reference amino acid class are masked by accessibility restrictions imposed by single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid class into the alternative amino acid class of the unreachable alternative amino acid class.

[0521] 50. The computer-implemented method according to Clause 1, wherein the masked amino acid class results in zero loss and makes no contribution to gradient updates.

[0522] 51. The computer-implemented method according to Clause 50, wherein the masked amino acid class is identified in a lookup table.

[0523] 52. The computer-implemented method according to Clause 51, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0524] Clause Group 2

[0525] 1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, the computer-implemented method comprising:

[0526] Access proteins that contain the corresponding amino acids at the corresponding locations;

[0527] A specific amino acid of a specific amino acid class at a specific position in the protein is designated as a missing amino acid, and the remaining amino acids at the remaining positions in the protein are designated as non-missing amino acids;

[0528] The nicked space representation of the protein is generated, the nicked space representation

[0529] Including the spatial configuration of the non-notched amino acids, and

[0530] Excluding the spatial configuration of the missing amino acid; and

[0531] The pathogenicity of the corresponding alternative amino acid at the specific position is determined at least in part based on the notched spatial representation.

[0532] 2. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an amino acid-type distance channel.

[0533] Each amino acid type distance channel in the aforementioned amino acid type distance channels has voxel-type distance values ​​for voxels in a plurality of voxels, and

[0534] The voxel-based distance value specifies the distance from the corresponding voxel among the plurality of voxels to the atom of the non-notched amino acid.

[0535] 3. The computer-implemented method according to Clause 2, wherein the spatial configuration of the unnotched amino acid is determined based on the spatial proximity between the corresponding voxel and the atoms of the unnotched amino acid.

[0536] 4. The computer-implemented method according to Clause 2, wherein when determining the voxel distance value, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the distance from the corresponding voxel to the nicked amino acid atom.

[0537] 5. The computer-implemented method according to Clause 4, wherein the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the spatial proximity between the corresponding voxel and the atoms of the nicked amino acid.

[0538] 6. The computer-implemented method according to Clause 1, wherein the specific amino acid is a reference amino acid serving as the major allele of the protein.

[0539] 7. The computer-implemented method according to Clause 1, wherein the corresponding alternative amino acid is an alternative amino acid generated by a nucleotide variant generated by a corresponding combination at the specific position.

[0540] 8. The computer-implemented method according to Clause 1, wherein the pathogenicity predictor determines the pathogenicity of the corresponding alternative amino acid by:

[0541] Processing the nicked spatial representation as input; and

[0542] Generate the corresponding pathogenicity score for the corresponding amino acid class as output.

[0543] 9. The computer-implemented method according to Clause 8, wherein the pathogenicity predictor is trained on a training set.

[0544] 10. The computer-implemented method according to Clause 9, wherein the training set has corresponding protein samples for corresponding locations in the proteome.

[0545] 11. The computer-implemented method according to Clause 10, wherein the proteome has ten million locations, and therefore the training set has ten million protein samples.

[0546] 12. The computer-implemented method according to Clause 11, wherein the corresponding protein sample has a corresponding nicked space representation generated by using a corresponding reference amino acid at the corresponding position in the proteome as a corresponding nicked amino acid.

[0547] 13. The computer-implemented method according to Clause 12, wherein the reference amino acid is the major allelic amino acid of the proteome.

[0548] 14. The computer-implemented method according to Clause 13, wherein the pathogenicity predictor is trained on a specific protein sample and estimates the pathogenicity of a corresponding substitute amino acid at a specific location of a specific reference amino acid in the specific protein sample by:

[0549] Process the specific nicked spatial representation of the specific protein sample as input.

[0550] The specific notched spatial representation is generated in the following manner.

[0551] Using the specific reference amino acid as the missing amino acid, and

[0552] The remaining amino acids at the remaining positions in the specific protein sample are used as non-missing amino acids; and

[0553] Generate the corresponding pathogenicity score of the corresponding amino acid class as output.

[0554] 15. The computer-implemented method according to Clause 14, wherein each protein sample in the protein sample has a corresponding benchmark truth label for the corresponding amino acid class.

[0555] 16. The computer-implemented method according to Clause 15, wherein the corresponding benchmark truth mark includes an absolutely benign mark of a reference amino acid class in the corresponding amino acid class, and includes a corresponding absolutely pathogenic mark of a corresponding alternative amino acid class in the corresponding amino acid class.

[0556] 17. The computer-implemented method according to Clause 16, wherein the absolute benignity marker is 0.

[0557] 18. The computer-implemented method according to Clause 17, wherein the absolute pathogenicity marker is the same in the corresponding alternative amino acid class.

[0558] 19. The computer-implemented method according to Clause 18, wherein the absolute pathogenicity marker is 1.

[0559] 20. The computer-implemented method according to Clause 1, wherein the error is determined based on:

[0560] The comparison between the pathogenicity score of the reference amino acid class and the absolute benign marker, and

[0561] The corresponding pathogenicity score of the corresponding alternative amino acid class is compared with the corresponding absolute pathogenicity marker.

[0562] 21. The computer-implemented method according to Clause 20, wherein training techniques are used to improve the coefficients of the pathogenicity predictor based on the error.

[0563] 22. The computer-implemented method according to Clause 21, wherein the pathogenicity predictor is trained with the ten million protein samples in ten million training iterations.

[0564] 23. The computer-implemented method according to Clause 8, wherein the corresponding amino acid class corresponds to the corresponding twenty naturally occurring amino acids.

[0565] 24. The computer-implemented method according to Clause 23, wherein the corresponding amino acid class corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0566] 25. The computer-implemented method according to Clause 11, wherein the proteome has one million to ten million locations, and therefore the training set has one million to ten million protein samples.

[0567] The pathogenicity predictor is trained using one million to ten million protein samples in one million to ten million training iterations.

[0568] 26. The computer-implemented method according to Clause 8, wherein the pathogenicity predictor generates a reference pathogenicity score for a first alternative amino acid of the reference amino acid class.

[0569] The pathogenicity predictor generates the corresponding pathogenicity score of the corresponding alternative amino acid for the corresponding alternative amino acid class.

[0570] 27. The computer-implemented method according to Clause 26, wherein the corresponding final substitution pathogenicity fraction of the corresponding alternative amino acid is the corresponding substitution pathogenicity fraction.

[0571] 28. The computer-implemented method according to Clause 27, wherein the corresponding final substitution pathogenicity score of the corresponding alternative amino acid is based on a corresponding combination of the reference pathogenicity score and the corresponding substitution pathogenicity score.

[0572] 29. The computer-implemented method according to Clause 28, wherein the corresponding final substitution pathogenicity fraction of the corresponding alternative amino acid is a corresponding ratio of the corresponding substitution pathogenicity fraction to the sum of the reference pathogenicity fraction and the corresponding substitution pathogenicity fraction.

[0573] 30. The computer-implemented method according to Clause 29, wherein the corresponding final substitution pathogenicity score of the corresponding substitution amino acid is determined by subtracting the reference pathogenicity score from the corresponding substitution pathogenicity score, respectively.

[0574] 31. The computer-implemented method according to Clause 8, wherein the pathogenicity predictor has an output layer that generates the corresponding pathogenicity score.

[0575] 32. The computer-implemented method according to Clause 31, wherein the output layer is a normalization layer.

[0576] 33. The computer-implemented method according to Clause 32, wherein the corresponding pathogenicity score is normalized.

[0577] 34. The computer-implemented method according to Clause 31, wherein the output layer is a softmax layer.

[0578] 35. The computer-implemented method according to Clause 34, wherein the corresponding pathogenicity score is exponentially normalized.

[0579] 36. The computer-implemented method according to Clause 31, wherein the output layer has corresponding S-type units that generate the respective pathogenicity scores.

[0580] 37. The computer-implemented method according to Clause 31, wherein the corresponding pathogenicity score is non-normalized.

[0581] 38. The computer-implemented method according to Clause 1, wherein the spatial configuration of the non-notched amino acid is encoded as an evolutionary spectrum channel based on the pan-amino acid conserved frequency of the amino acid having the atom closest to the voxel.

[0582] 39. The computer-implemented method according to Clause 38, wherein when determining the conservative frequency of the ubiquitin amino acid, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the nearest atom of the nicked amino acid.

[0583] 40. The computer-implemented method according to Clause 1, wherein the spatial configuration of the non-notched amino acid is encoded as an evolutionary spectrum channel based on the per-amino acid conserved frequency of the corresponding amino acid having the corresponding atom closest to the voxel.

[0584] 41. The computer-implemented method according to Clause 40, wherein when determining the conservative frequency per amino acid, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the corresponding nearest atom of the nicked amino acid.

[0585] 42. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an annotation channel.

[0586] 43. The computer-implemented method according to Clause 42, wherein when the annotation channel is determined, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0587] 44. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as a structure confidence channel.

[0588] 45. The computer-implemented method according to Clause 44, wherein when determining the structural confidence channel, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0589] 46. ​​The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an additional input channel.

[0590] 47. The computer-implemented method according to Clause 46, wherein when the additional input channel is determined, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0591] 48. The computer-implemented method according to Clause 10, wherein the proteome includes human proteomes and non-human proteomes, including non-human primate proteomes.

[0592] 49. The computer-implemented method according to Clause 8, wherein in the reference truth mark, those unreachable alternative amino acids of the reference amino acid class are masked by accessibility restrictions imposed by single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid class into the alternative amino acid class of the unreachable alternative amino acid class.

[0593] 50. The computer-implemented method according to Clause 1, wherein the masked amino acid class results in zero loss and makes no contribution to gradient updates.

[0594] 51. The computer-implemented method according to Clause 50, wherein the masked amino acid class is identified in a lookup table.

[0595] 52. The computer-implemented method according to Clause 51, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0596] Clause Group 3

[0597] 1. A computer-implemented method for generating training data for training a variant pathogenicity classifier, the computer-implemented method comprising:

[0598] Access multiple amino acid positions in a proteome containing multiple proteins;

[0599] The major allele amino acid at the specified amino acid position is designated as the reference amino acid for the specified proteins.

[0600] For each of the plurality of amino acid positions

[0601] Those nucleotide substitutions are classified as benign variants, which replace a specific reference amino acid at a specific amino acid position in a specific substitution representation of a specific protein, and

[0602] Those nucleotide substitutions are classified as pathogenic variants, which are those in which a specific reference amino acid is replaced by an alternative amino acid at the specific amino acid position, wherein the alternative amino acid is different from the specific reference amino acid; and

[0603] The benign variant and the pathogenic variant are used as training data to train a variant pathogenicity classifier.

[0604] 2. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a second amino acid for a first amino acid at a given amino acid position in a protein is pathogenic or benign.

[0605] 3. The computer-implemented method according to Clause 2, wherein the variant pathogenicity classifier is trained to generate the alternative pathogenicity score.

[0606] 4. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a corresponding substitution of a first amino acid at a given amino acid position in a protein is pathogenic or benign.

[0607] 5. The computer-implemented method according to Clause 4, wherein the variant pathogenicity classifier is trained to generate the corresponding pathogenicity score of the corresponding substitution.

[0608] 6. The computer-implemented method according to Clause 5, wherein the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids.

[0609] 7. The computer-implemented method according to Clause 6, wherein the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0610] 8. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether the insertion of an amino acid at a given empty amino acid position in a protein is pathogenic or benign.

[0611] 9. The computer-implemented method according to Clause 8, wherein the variant pathogenicity classifier is trained to generate the pathogenicity score of the insertion.

[0612] 10. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a corresponding insertion of a corresponding amino acid at a given empty amino acid position in a protein is pathogenic or benign.

[0613] 11. The computer-implemented method according to Clause 10, wherein the variant pathogenicity classifier is trained to generate a corresponding pathogenicity score for the corresponding insertion.

[0614] 12. The computer-implemented method according to Clause 11, wherein the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids.

[0615] 13. The computer-implemented method according to Clause 12, wherein the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0616] 14. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a second amino acid for a first amino acid at a given amino acid position in a protein is spatially tolerant to other amino acids of the protein.

[0617] 15. The computer-implemented method according to Clause 14, wherein the variant pathogenicity classifier is trained to generate the alternative spatial tolerance score.

[0618] 16. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a corresponding substitution of a first amino acid at a given amino acid position in a protein is spatially tolerant to other amino acids of the protein.

[0619] 17. The computer-implemented method according to Clause 16, wherein the variant pathogenicity classifier is trained to generate the corresponding spatial tolerance score for the corresponding substitution.

[0620] 18. The computer-implemented method according to Clause 17, wherein the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids.

[0621] 19. The computer-implemented method according to Clause 18, wherein the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0622] 20. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether the insertion of an amino acid at a given empty amino acid position in a protein is spatially tolerant to other amino acids of the protein.

[0623] 21. The computer-implemented method according to Clause 20, wherein the variant pathogenicity classifier is trained to generate the spatial tolerance score of the insertion.

[0624] 22. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a corresponding insertion of a corresponding amino acid at a given empty amino acid position in a protein is spatially tolerant to other amino acids of the protein.

[0625] 23. The computer-implemented method according to Clause 22, wherein the variant pathogenicity classifier is trained to generate a corresponding spatial tolerance score for the corresponding insertion.

[0626] 24. The computer-implemented method according to Clause 23, wherein the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids.

[0627] 25. The computer-implemented method according to Clause 24, wherein the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0628] 26. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a second amino acid for a first amino acid at a given amino acid position in a protein is evolutionarily conserved or non-conserved.

[0629] 27. The computer-implemented method according to Clause 26, wherein the variant pathogenicity classifier is trained to generate the evolutionary conservation score of the substitution.

[0630] 28. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether a corresponding substitution of a first amino acid at a given amino acid position in a protein is evolutionarily conserved or non-conserved.

[0631] 29. The computer-implemented method according to Clause 28, wherein the variant pathogenicity classifier is trained to generate a corresponding evolutionary conservation score for the corresponding substitution.

[0632] 30. The computer-implemented method according to Clause 29, wherein the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids.

[0633] 31. The computer-implemented method according to Clause 30, wherein the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0634] 32. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether the insertion of an amino acid at a given empty amino acid position in a protein is evolutionarily conserved or non-conserved.

[0635] 33. The computer-implemented method according to Clause 32, wherein the variant pathogenicity classifier is trained to generate an evolutionary conservation score for the insertion.

[0636] 34. The computer-implemented method according to Clause 1, wherein the variant pathogenicity classifier is trained to determine whether the corresponding insertion of a corresponding amino acid at a given empty amino acid position in a protein is evolutionarily conserved or non-conserved.

[0637] 35. The computer-implemented method according to Clause 34, wherein the variant pathogenicity classifier is trained to generate a corresponding evolutionary conservation score for the corresponding insertion.

[0638] 36. The computer-implemented method according to Clause 35, wherein the corresponding amino acid corresponds to one of the twenty naturally occurring amino acids.

[0639] 37. The computer-implemented method according to Clause 36, wherein the corresponding amino acid corresponds to a corresponding naturally occurring amino acid from a subset of the twenty naturally occurring amino acids.

[0640] 38. The computer-implemented method according to Clause 14, wherein spatial tolerance corresponds to structural tolerance, and spatial intolerance corresponds to structural intolerance.

[0641] 39. The computer-implemented method according to Clause 1, wherein the plurality of amino acid positions ranges from one million to ten million amino acid positions.

[0642] 40. The computer-implemented method according to Clause 1, wherein the plurality of amino acid positions ranges from ten million to one hundred million amino acid positions.

[0643] 41. The computer-implemented method according to Clause 1, wherein the plurality of amino acid positions ranges from one hundred million to one billion amino acid positions.

[0644] 42. The computer-implemented method according to Clause 1, wherein the plurality of amino acid positions ranges from one to one million amino acid positions.

[0645] 43. The computer-implemented method according to Clause 1, wherein in the reference truth mark, those unreachable alternative amino acids of the reference amino acid class are masked by accessibility restrictions imposed by single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid class into the alternative amino acid class of the unreachable alternative amino acid class.

[0646] 44. The computer-implemented method according to Clause 1, wherein the masked amino acid class results in zero loss and makes no contribution to gradient updates.

[0647] 45. The computer-implemented method according to Clause 44, wherein the masked amino acid class is identified in a lookup table.

[0648] 46. ​​The computer-implemented method according to Clause 45, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0649] Clause Set 4 (ILLM 1060-1)

[0650] 1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, the computer-implemented method comprising:

[0651] A specific amino acid at a specific position in a protein is designated as a missing amino acid, and the remaining amino acids at the remaining positions in the protein are designated as non-missing amino acids;

[0652] The nicked space representation of the protein is generated, the nicked space representation

[0653] Including the spatial configuration of the non-notched amino acids, and

[0654] Exclude the spatial configuration of the missing amino acid;

[0655] The evolutionary conservation at the specific position of the corresponding amino acid in the corresponding amino acid class is determined at least in part based on the gapped spatial representation; and

[0656] The pathogenicity of a corresponding nucleotide variant is determined at least in part based on the evolutionary conservation of the corresponding amino acid, wherein the corresponding nucleotide variant replaces the specific amino acid in the alternative representation of the protein with the corresponding amino acid, respectively.

[0657] 2. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an amino acid-type distance channel.

[0658] Each amino acid type distance channel in the aforementioned amino acid type distance channels has voxel-type distance values ​​for voxels in a plurality of voxels, and

[0659] The voxel-based distance value specifies the distance from the corresponding voxel among the plurality of voxels to the atom of the non-notched amino acid.

[0660] 3. The computer-implemented method according to Clause 2, wherein the spatial configuration of the unnotched amino acid is determined based on the spatial proximity between the corresponding voxel and the atoms of the unnotched amino acid.

[0661] 4. The computer-implemented method according to Clause 2, wherein when determining the voxel distance value, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the distance from the corresponding voxel to the nicked amino acid atom.

[0662] 5. The computer-implemented method according to Clause 4, wherein the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the spatial proximity between the corresponding voxel and the atoms of the nicked amino acid.

[0663] 6. The computer-implemented method according to Clause 1, wherein the specific amino acid is a reference amino acid serving as the major allele of the protein.

[0664] 7. The computer-implemented method according to Clause 1, wherein the evolutionary conservation predictor determines the evolutionary conservation in such a way as:

[0665] Processing the nicked spatial representation as input; and

[0666] Generate the corresponding evolutionary conservation score of the corresponding amino acid as output.

[0667] 8. The computer-implemented method according to Clause 7, wherein the corresponding evolutionary conservation scores are sorted by size.

[0668] 9. The computer-implemented method according to Clause 7 further includes classifying a nucleotide variant as pathogenic when the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is below a threshold.

[0669] 10. The computer-implemented method according to Clause 7 further includes classifying a nucleotide variant as pathogenic when the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is 0.

[0670] 11. The computer-implemented method according to Clause 7 further includes classifying a nucleotide variant as benign when the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is higher than a threshold.

[0671] 12. The computer-implemented method according to Clause 7 further includes classifying a nucleotide variant as benign when the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is non-zero.

[0672] 13. The computer-implemented method according to Clause 7, wherein the evolutionary conservation predictor is trained on a conservative training set and a non-conservative training set.

[0673] 14. The computer-implemented method according to Clause 13, wherein the conserved training set comprises corresponding conserved protein samples for corresponding conserved amino acids at corresponding positions in the proteome.

[0674] The nonconservative training set refers to the nonconservative protein sample with the corresponding nonconservative amino acid at the corresponding position.

[0675] 15. The computer-implemented method according to Clause 14, wherein each of the respective positions has a set of conserved amino acids and a set of non-conserved amino acids.

[0676] 16. The computer-implemented method according to Clause 15, wherein a specific conserved amino acid set at a specific location of a specific protein in the proteome includes at least one major allelic amino acid observed at the specific location in multiple species.

[0677] 17. The computer-implemented method according to Clause 16, wherein the particular conserved amino acid set comprises one or more minor allele amino acids observed at the particular position in the plurality of species.

[0678] 18. The computer-implemented method according to Clause 17, wherein the specific non-conservative amino acid group at the specific position includes amino acids not in the specific conserved amino acid group.

[0679] 19. The computer-implemented method according to Clause 18, wherein the specific conserved amino acid set and the specific non-conserved amino acid set are identified based on the evolutionary conservation profile of homologous proteins of the plurality of species.

[0680] 20. The computer-implemented method according to Clause 18, wherein the evolutionary conservation spectrum of the homologous protein is determined using a position-specific frequency matrix (PSFM).

[0681] 21. The computer-implemented method according to Clause 18, wherein the evolutionary conservation spectrum of the homologous protein is determined using a position-specific scoring matrix (PSSM).

[0682] 22. The computer-implemented method according to Clause 16, wherein the major allele amino acid is a reference amino acid.

[0683] 23. The computer-implemented method according to Clause 14, wherein each of the corresponding positions has C conserved amino acids from the conserved amino acid group.

[0684] Each of the corresponding positions has NC nonconserved amino acids from the nonconserved amino acid group, where NC = 20-C.

[0685] The conserved training set contains CP conserved protein samples, where CP = the number of the corresponding positions * C.

[0686] The non-conservative training set contains NCP non-conservative protein samples, where NCP = the number of the corresponding positions * (20 - C).

[0687] 24. The computer-implemented method according to Clause 23, wherein the range of C is 1 to 10.

[0688] 25. The computer-implemented method according to Clause 24, wherein C varies between the respective positions.

[0689] 26. The computer-implemented method according to Clause 25, wherein the C is the same for some of the corresponding positions.

[0690] 27. The computer-implemented method according to Clause 14, wherein the respective conserved and non-conserved protein samples have a respective nicked space representation generated by using the respective reference amino acid at the respective position as the respective nicked amino acid.

[0691] 28. The computer-implemented method according to Clause 27, wherein the evolutionary conservation predictor is trained on a particular conserved protein sample and estimates the evolutionary conservation of a particular conserved amino acid at a particular position in the particular conserved protein sample by:

[0692] Process the specific nicked spatial representation of the specific conserved protein sample as input.

[0693] The specific notched spatial representation is generated in the following manner.

[0694] Using the specific reference amino acid at the specified position as the missing amino acid, and

[0695] The remaining amino acids at the remaining positions in the specific conserved protein sample are used as non-missing amino acids; and

[0696] Generate the evolutionary conservation score of the specific conserved amino acid as output.

[0697] 29. The computer-implemented method according to Clause 28, wherein each of the conserved protein samples has a baseline truth value conserved label.

[0698] 30. The computer-implemented method according to Clause 29, wherein the baseline truth value conservation marker is an evolutionary conservation frequency.

[0699] 31. The computer-implemented method according to Clause 29, wherein the base truth value conservative label is 1.

[0700] 32. The computer-implemented method according to Clause 29, wherein the evolutionary conservation of the particular conserved amino acid is compared with the baseline true value conservation label to determine an error, and a training technique is used to improve the coefficients of the evolutionary conservation predictor based on the error.

[0701] 33. The computer-implemented method according to Clause 32, wherein when the specific conserved amino acid is the specific reference amino acid, the baseline truth value conservatism marker is masked and not used to determine the error.

[0702] The masking prevents the evolutionary conservation predictor from overfitting the specific reference amino acid.

[0703] 34. The computer-implemented method according to Clause 32, wherein the training technique is a gradient update technique based on a loss function.

[0704] 35. The computer-implemented method according to Clause 27, wherein the evolutionary conservation predictor is trained on a specific nonconservative protein sample and estimates the evolutionary conservation of a specific nonconservative amino acid at a specific position in the specific nonconservative protein sample by:

[0705] Process the specific nicked spatial representation of the specific nonconservative protein sample as input.

[0706] The specific notched spatial representation is generated in the following manner.

[0707] Using a specific reference amino acid at the specified location as the missing amino acid, and

[0708] The remaining amino acids at the remaining positions in the specific nonconservative protein sample are used as the non-gap amino acids; and

[0709] Generate the evolutionary conservation score of the specific nonconservative amino acid as output.

[0710] 36. The computer-implemented method according to Clause 35, wherein each of the nonconservative protein samples has a baseline true nonconservative label.

[0711] 37. The computer-implemented method according to Clause 35, wherein the baseline truth nonconservative label is an evolutionary conservation frequency.

[0712] 38. The computer-implemented method according to Clause 35, wherein the non-conservative flag of the baseline truth value is 0.

[0713] 39. The computer-implemented method according to Clause 35, wherein the evolutionary conservation score of the particular nonconservative amino acid is compared with the baseline true nonconservative label to determine an error, and the training technique is used to improve the coefficients of the evolutionary conservation predictor based on the error.

[0714] 40. The computer-implemented method according to Clause 7, wherein the evolutionary conservation predictor is trained on a training set.

[0715] 41. The computer-implemented method according to Clause 40, wherein the training set has corresponding protein samples for the corresponding positions in the proteome.

[0716] 42. The computer-implemented method according to Clause 41, wherein the corresponding protein sample has a corresponding nicked space representation generated by using the corresponding reference amino acid at the corresponding position as the corresponding nicked amino acid.

[0717] 43. The computer-implemented method according to Clause 42, wherein the evolutionary conservation predictor is trained on a specific protein sample and estimates the evolutionary conservation of a corresponding amino acid of a corresponding amino acid class at a specific position in the specific protein sample by:

[0718] Process the specific nicked spatial representation of the specific protein sample as input.

[0719] The specific notched spatial representation is generated in the following manner.

[0720] Using the specific reference amino acid at the specified position as the missing amino acid, and

[0721] The remaining amino acids at the remaining positions in the specific protein sample are used as non-missing amino acids; and

[0722] Generate the corresponding evolutionary conservation score of the corresponding amino acid as output.

[0723] 44. The computer-implemented method according to Clause 43, wherein each protein sample in the protein sample has a corresponding benchmark truth label for the corresponding amino acid.

[0724] 45. The computer-implemented method according to Clause 44, wherein the corresponding reference truth mark comprises one or more conservative marks of one or more conserved amino acids in the corresponding amino acid, and comprises one or more non-conserved marks of one or more non-conserved amino acids in the corresponding amino acid.

[0725] 46. ​​The computer-implemented method according to Clause 45, wherein the conservative markers and the non-conservative markers have corresponding evolutionary conservation frequencies.

[0726] 47. The computer-implemented method according to Clause 46, wherein the corresponding evolutionary conservation frequencies can be sorted according to size.

[0727] 48. The computer-implemented method according to Clause 46, wherein the conservative flag is 1 and the non-conservative flag is 0.

[0728] 49. The computer-implemented method according to Clause 46, wherein the error is determined based on:

[0729] The corresponding evolutionary conservation score of the corresponding conserved amino acid and the corresponding comparison of the corresponding conserved amino acid and

[0730] The evolutionary conservation score of the corresponding non-conservative amino acid is compared with that of the corresponding non-conservative amino acid.

[0731] 50. The computer-implemented method according to Clause 49, wherein the training technique is used to improve the coefficients of the evolutionary conservation predictor based on the error.

[0732] 51. The computer-implemented method according to claim 50, wherein the conserved amino acid includes the specific reference amino acid, and the conserved label of the specific reference amino acid is masked and not used to determine the error.

[0733] The masking prevents the evolutionary conservation predictor from overfitting the specific reference amino acid.

[0734] 52. The computer-implemented method according to Clause 14, wherein the proteome has one to ten million sites.

[0735] Each of the ten million positions mentioned herein has the C conserved amino acids from the conserved amino acid set.

[0736] Each of the ten million positions described herein has NC nonconserved amino acids from the nonconserved amino acid group, where NC = 20-C.

[0737] The conserved training set contains CP conserved protein samples, where CP = 10 million to 10 million * C, and

[0738] The non-conservative training set contains NCP non-conservative protein samples, where NCP = one to ten million * (20-C).

[0739] 53. The computer-implemented method according to Clause 14, wherein the evolutionary conservation predictor is trained over 20 million to 200 million training iterations.

[0740] The aforementioned 20 million to 200 million training iterations include

[0741] One million to ten million training iterations using the aforementioned one million to ten million conserved protein samples, and

[0742] Nineteen million to one hundred and ninety million iterations were performed using the aforementioned nineteen million to one hundred and ninety million non-conservative protein samples.

[0743] 54. The computer-implemented method according to Clause 14, wherein the proteome has one million to ten million sites, and therefore the training set has one million to ten million protein samples.

[0744] The evolutionary conservation predictor is trained with one million to ten million protein samples in one million to ten million training iterations.

[0745] 55. The computer-implemented method according to Clause 1, wherein the spatial configuration of the non-notched amino acid is encoded as an evolutionary spectrum channel based on the pan-amino acid conserved frequency of the amino acid having the atom closest to the voxel.

[0746] 56. The computer-implemented method according to Clause 55, wherein when determining the conservative frequency of the ubiquitin amino acid, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the nearest atom of the nicked amino acid.

[0747] 57. The computer-implemented method according to Clause 1, wherein the spatial configuration of the non-missing amino acid is encoded as an evolutionary spectrum channel based on the per-amino acid conserved frequency of the corresponding amino acid having the corresponding atom closest to the voxel.

[0748] 58. The computer-implemented method according to Clause 57, wherein when determining the conservative frequency per amino acid, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the corresponding nearest atom of the nicked amino acid.

[0749] 59. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an annotation channel.

[0750] 60. The computer-implemented method according to Clause 59, wherein when the annotation channel is determined, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0751] 61. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as a structure confidence channel.

[0752] 62. The computer-implemented method according to Clause 61, wherein when determining the structural confidence channel, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0753] 63. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as a structure confidence channel.

[0754] 64. The computer-implemented method according to Clause 63, wherein when determining the structural confidence channel, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0755] 65. The computer-implemented method according to Clause 1, wherein the spatial configuration of the unnotched amino acid is encoded as an additional input channel.

[0756] 66. The computer-implemented method according to Clause 65, wherein when the additional input channel is determined, the spatial configuration of the nicked amino acid is excluded from the nicked spatial representation by ignoring the atoms of the nicked amino acid.

[0757] 67. The computer-implemented method according to Clause 14, wherein the proteome includes human proteomes and non-human proteomes, including non-human primate proteomes.

[0758] 68. The computer-implemented method according to Clause 1, wherein in the reference truth mark, those unreachable alternative amino acids of the reference amino acid class are masked by accessibility restrictions of single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid class into the alternative amino acid class of the unreachable alternative amino acid class.

[0759] 69. The computer-implemented method according to Clause 1, wherein the masked amino acid class results in zero loss and makes no contribution to gradient updates.

[0760] 70. The computer-implemented method according to Clause 69, wherein the masked amino acid class is identified in a lookup table.

[0761] 71. The computer-implemented method according to Clause 70, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

[0762] Clause Set 5 (ILLM 1061-1)

[0763] 1. A computer-implemented method for training a pathogenicity predictor, the computer-implemented method comprising:

[0764] Access a gapped training set, which includes corresponding gapped protein samples for corresponding locations in the proteome;

[0765] Access the non-gap training set, which includes non-gap benign protein samples and non-gap pathogenic protein samples;

[0766] Generate the corresponding notched space representation of the notched protein sample, and generate the corresponding unnotched space representation of the unnotched benign protein sample and the unnotched pathogenic protein sample;

[0767] A pathogenicity predictor is trained and generated in one or more training loops, wherein each training loop uses a gleaned spatial representation from the corresponding gleaned spatial representation and a non-gleaned spatial representation from the corresponding non-gleaned spatial representation as training examples; and

[0768] The trained pathogenicity classifier is used to determine the pathogenicity of the variant.

[0769] 2. The computer-implemented method according to Clause 1, wherein the corresponding nicked protein sample is labeled with a corresponding nicked baseline truth sequence.

[0770] 3. The computer-implemented method according to Clause 2, wherein a specific defective reference true sequence of a specific defective protein sample has a benign label of a specific amino acid class corresponding to a reference amino acid at a specific position in the specific defective protein.

[0771] 4. The computer-implemented method according to Clause 3, wherein the particular nicked protein sample has a corresponding pathogenicity marker of the corresponding remaining amino acid class corresponding to the substitute amino acid at the particular position.

[0772] 5. The computer-implemented method according to Clause 1, wherein a particular non-gap benign protein sample comprises a benign alternative amino acid substituted at a particular position by a benign nucleotide variant.

[0773] 6. The computer-implemented method according to Clause 5, wherein a particular non-gap pathogenic protein sample comprises a pathogenic alternative amino acid substituted at a particular position by a pathogenic nucleotide variant.

[0774] 7. The computer-implemented method according to Clause 6, wherein the specific non-gap benign protein sample is labeled with a benign benchmark truth sequence having a benign label of a specific amino acid class corresponding to the benign alternative amino acid.

[0775] 8. The computer-implemented method according to Clause 7, wherein the benign baseline truth sequence has a corresponding masking label for the corresponding remaining amino acid class that corresponds to an amino acid different from the benign alternative amino acid.

[0776] 9. The computer-implemented method according to Clause 8, wherein the specific non-gap pathogenic protein sample is labeled with a pathogenicity benchmark truth sequence having a pathogenicity marker for a specific amino acid class corresponding to the pathogenic alternative amino acid.

[0777] 10. The computer-implemented method according to Clause 9, wherein the pathogenicity benchmark truth sequence has a corresponding masking label for the corresponding remaining amino acid class corresponding to an amino acid different from the pathogenic alternative amino acid.

[0778] 11. The computer-implemented method according to Clause 1 further includes using a sample indicator to indicate to the pathogenicity predictor whether the current training example is a notched spatial representation of a notched protein sample or a non-notched spatial representation of a non-notched protein sample.

[0779] 12. The computer-implemented method according to Clause 1 further includes masking the benign marker of the specific amino acid class corresponding to the reference amino acid at the specific location in the specific nicked protein.

[0780] 13. The computer-implemented method according to Clause 1, wherein the non-gap benign protein sample is derived from common human and non-human primate nucleotide variants.

[0781] 14. The computer-implemented method according to Clause 1, wherein the non-gap pathogenic protein sample is derived from a combinatorially simulated nucleotide variant.

[0782] 15. The computer-implemented method according to Clause 1, wherein the pathogenicity predictor generates an amino acid-like output sequence in response to processing a training example.

[0783] The amino acid class output sequence has an amino acid class pathogenicity score.

[0784] 16. The computer-implemented method according to Clause 1 further includes measuring the performance of the trained pathogenicity predictor between training cycles on a validation set.

[0785] 17. The computer-implemented method according to Clause 16, wherein the validation set comprises a pair of notched and unnotched spatial representations for each retained protein sample.

[0786] 18. The computer-implemented method according to Clause 1, wherein the trained pathogenicity predictor generates a first amino acid class output sequence for the nicked space representation of the pair, and a second amino acid class output sequence for the unnicked space representation of the pair.

[0787] The final pathogenicity score of the nucleotide variant causing the amino acid substitution in the retained protein sample is determined based on a combination of the first and second pathogenicity scores of the amino acid substitutions in the first and second amino acid class output sequences.

[0788] 19. The computer-implemented method according to Clause 18, wherein the final pathogenicity score is based on the average of the first and second pathogenicity scores.

[0789] 20. The computer-implemented method according to Clause 1, wherein at least some of the training cycles use the same number of gapped and ungauge space representations.

[0790] 21. The computer-implemented method according to Clause 1, wherein at least some of the training cycles use training example batches having the same number of gapped and ungapped spatial representations.

[0791] 22. The computer-implemented method according to Clause 1, wherein the masked markers do not contribute to error determination and therefore do not contribute to the training of the pathogenicity predictor.

[0792] 23. The computer-implemented method according to Clause 22, wherein the masking marker is cleared.

[0793] 24. The computer-implemented method according to Clause 1, wherein the gapped spatial representation and the non-gapped spatial representation are weighted differently such that the contribution of the gapped spatial representation to gradient updates of parameters applied to the pathogenicity predictor in response to the pathogenicity predictor processing the non-gapped spatial representation is different from the contribution of the non-gapped spatial representation to gradient updates of the parameters applied to the pathogenicity predictor in response to the pathogenicity predictor processing the non-gapped spatial representation.

[0794] 25. The computer-implemented method according to Clause 24, wherein the change is determined by predefined weights.

[0795] 26. A computer-implemented method for training a pathogenicity predictor, the computer-implemented method comprising:

[0796] Starting with training a pathogenicity classifier on a gapped training set, and generating the trained pathogenicity classifier;

[0797] The trained pathogenicity classifier is further trained on the non-gap training set, and a retrained pathogenicity classifier is generated; and

[0798] The pathogenicity of the variant is determined using the retrained pathogenicity classifier.

[0799] 27. The computer-implemented method of claim 26 further includes measuring the performance of the trained pathogenicity predictor between training cycles on a first validation set, the first validation set comprising only the non-gap spatial representation of the retained protein samples.

[0800] 28. The computer-implemented method of claim 27 further includes measuring the performance of the retrained pathogenicity predictor between training cycles on a second validation set, the second validation set comprising notched and unnotched spatial representations of retained protein samples.

[0801] 29. The computer-implemented method according to Clause 28, wherein the retrained pathogenicity predictor generates a first amino acid-like output sequence of the pair in response to processing the pair.

[0802] Based on the first amino acid type output sequence, the final pathogenicity score of the nucleotide variant that causes amino acid substitution in the corresponding retained protein sample is determined.

[0803] 30. A computer-implemented method for training a pathogenicity predictor, the computer-implemented method comprising:

[0804] Access includes a nicked training set of corresponding nicked protein samples for corresponding positions in the proteome, wherein the corresponding nicked protein samples are labeled with corresponding nicked reference truth sequences, wherein a specific nicked reference truth sequence of a particular nicked protein sample has a benign label of a specific amino acid class corresponding to a reference amino acid at a specific position in the particular nicked protein, and has a corresponding pathogenic label of a corresponding remaining amino acid class corresponding to a substitute amino acid at the specific position;

[0805] Access is made to a non-gap training set comprising non-gap benign protein samples and non-gap pathogenic protein samples, wherein a specific non-gap benign protein sample includes a benign alternative amino acid substituted by a benign nucleotide variant at a specific position, wherein a specific non-gap pathogenic protein sample includes a pathogenic alternative amino acid substituted by a pathogenic nucleotide variant at a specific position, wherein the specific non-gap benign protein sample is labeled with a benign benchmark truth sequence having a benign label for a specific amino acid class corresponding to the benign alternative amino acid and a corresponding masking label for a corresponding remaining amino acid class corresponding to an amino acid different from the benign alternative amino acid, and wherein the specific non-gap pathogenic protein sample is labeled with a pathogenic benchmark truth sequence having a pathogenic label for a specific amino acid class corresponding to the pathogenic alternative amino acid and a corresponding masking label for a corresponding remaining amino acid class corresponding to an amino acid different from the pathogenic alternative amino acid;

[0806] Generate the corresponding notched space representation of the notched protein sample, and generate the corresponding unnotched space representation of the unnotched benign protein sample and the unnotched pathogenic protein sample;

[0807] A pathogenicity predictor is trained and generated in one or more training loops, wherein each training loop uses a gleaned spatial representation from the corresponding gleaned spatial representation and a non-gleaned spatial representation from the corresponding non-gleaned spatial representation as training examples; and

[0808] The trained pathogenicity classifier is used to determine the pathogenicity of the variant.

[0809] Clause Group 6

[0810] 1. A computer-implemented method for generating training data for training a variant pathogenicity classifier, the computer-implemented method comprising:

[0811] Access multiple amino acid positions in a proteome containing multiple proteins;

[0812] The major allele amino acid at the specified amino acid position is designated as the reference amino acid for the specified proteins.

[0813] For each of the plurality of amino acid positions

[0814] Those nucleotide substitutions are classified as benign variants, which replace a specific reference amino acid at a specific amino acid position in a specific substitution representation of a specific protein, and

[0815] Those nucleotide substitutions are classified as pathogenic variants, which are those in which a specific reference amino acid is replaced by an alternative amino acid at the specific amino acid position, wherein the alternative amino acid is differe...

Claims

1. A computer-implemented method for training a pathogenicity predictor, the computer-implemented method comprising: Access a gapped training set, which includes a corresponding gapped protein sample having a corresponding gapped amino acid at a corresponding position of a corresponding reference amino acid; Access a non-gap training set, which includes non-gap benign protein samples having a corresponding reference amino acid at the corresponding position and non-gap pathogenic protein samples having a substitute amino acid at the corresponding position; Generate a corresponding nicked spatial representation that excludes the corresponding nicked amino acid from the spatial configuration of the corresponding nicked protein sample in the nicked training set; Generate corresponding unnotched spatial representations of unnotched benign protein samples and unnotched pathogenic protein samples from the unnotched training set; as well as The pathogenicity predictor is trained in one or more training loops using the following steps: Based on the nicked space representation from the corresponding nicked space representation, a first pathogenicity score for the substituted amino acid for the corresponding nicked amino acid is generated; and Based on the non-notch space representation from the corresponding non-notch space representation, a second pathogenicity fraction of the substituted amino acids at the corresponding positions in the non-notch benign protein sample or the non-notch pathogenic protein sample is generated.

2. The computer-implemented method of claim 1, further comprising using the pathogenicity predictor to determine the pathogenicity of variants causing amino acid substitutions within an amino acid sequence.

3. The computer-implemented method of claim 1, wherein the corresponding nicked protein sample is labeled with a corresponding nicked reference truth sequence, the corresponding nicked reference truth sequence comprising a corresponding benign label and a pathogenic label for the corresponding amino acid at the corresponding position of the corresponding nicked amino acid.

4. The computer-implemented method of claim 3, wherein the specific missing reference truth sequence of the specific missing protein sample in the corresponding missing protein sample has a benign label for a specific reference amino acid at a specific position of the corresponding missing amino acid in the specific missing protein sample.

5. The computer-implemented method of claim 4, wherein the particular nicked protein sample has a corresponding pathogenicity marker for the corresponding substitute amino acid at the particular location.

6. The computer-implemented method of claim 5, wherein a specific un-notched benign protein sample in the un-notched benign protein sample includes a benign alternative amino acid caused by a corresponding benign nucleotide variant at a specific position.

7. The computer-implemented method of claim 6, wherein a specific non-negative pathogenic protein sample in the non-negative pathogenic protein sample includes a pathogenic alternative amino acid caused by a corresponding pathogenic nucleotide variant at a specific position.

8. The computer-implemented method of claim 7, wherein the specific unnotched benign protein sample is labeled with a benign benchmark truth sequence having a benign label for the benign alternative amino acid.

9. The computer-implemented method of claim 8, wherein the benign benchmark truth sequence includes a corresponding masking label for the corresponding alternative amino acid that is different from the benign alternative amino acid.

10. The computer-implemented method according to any one of claims 7 to 9, wherein the specific non-gap pathogenic protein sample is labeled with a pathogenicity benchmark truth sequence having a pathogenicity label for the pathogenic alternative amino acid.

11. The computer-implemented method of claim 10, wherein the pathogenicity benchmark truth sequence has a corresponding masking label for the corresponding alternative amino acid that is different from the pathogenic alternative amino acid.

12. The computer-implemented method of claim 11, further comprising using a sample indicator to indicate to the pathogenicity predictor whether the current training example is a notched spatial representation of a notched protein sample or a non-notched spatial representation of a non-notched protein sample.

13. The computer-implemented method of claim 12, further comprising masking the benign label for the specific reference amino acid at the specific location in the specific nicked protein sample.

14. A system for training a pathogenicity predictor, the system comprising: At least one processor; and A non-transitory computer-readable medium, the non-transitory computer-readable medium including instructions that, when executed by the at least one processor, cause the system to: Access a gapped training set, which includes a corresponding gapped protein sample having a corresponding gapped amino acid at a corresponding position of a corresponding reference amino acid; Access a non-gap training set, which includes non-gap benign protein samples having a corresponding reference amino acid at the corresponding position and non-gap pathogenic protein samples having a substitute amino acid at the corresponding position; Generate a corresponding nicked spatial representation that excludes the corresponding nicked amino acid from the spatial configuration of the corresponding nicked protein sample in the nicked training set; Generate corresponding un-notched spatial representations of the un-notched benign protein samples and the un-notched pathogenic protein samples from the un-notched training set; as well as The pathogenicity predictor is trained in one or more training cycles using the following steps: Based on the nicked space representation from the corresponding nicked space representation, a first pathogenicity score for the substituted amino acid for the corresponding nicked amino acid is generated; and Based on the non-notch space representation from the corresponding non-notch space representation, a second pathogenicity fraction of the substituted amino acids at the corresponding positions in the non-notch benign protein sample or the non-notch pathogenic protein sample is generated.

15. The system of claim 14, wherein the non-notched benign protein sample is derived from common human and non-human primate nucleotide variants.

16. The system according to any one of claims 14 to 15, wherein the non-gap pathogenic protein sample is derived from a combinatorially simulated nucleotide variant.

17. The system of claim 15, wherein the pathogenicity predictor generates an amino acid-like output sequence in response to processing a training example. The amino acid class output sequence has an amino acid class pathogenicity score.

18. A non-transitory computer-readable storage medium comprising instructions that, when executed by at least one processor, cause a computing device to: Access a gapped training set, which includes a corresponding gapped protein sample having a corresponding gapped amino acid at a corresponding position of a corresponding reference amino acid; Access a non-gap training set, which includes non-gap benign protein samples having a corresponding reference amino acid at the corresponding position and non-gap pathogenic protein samples having a substitute amino acid at the corresponding position; Generate a corresponding nicked spatial representation that excludes the corresponding nicked amino acid from the spatial configuration of the corresponding nicked protein sample in the nicked training set; Generate corresponding un-notched spatial representations of the un-notched benign protein samples and the un-notched pathogenic protein samples from the un-notched training set; as well as The pathogenicity predictor is trained in one or more training loops using the following steps: Based on the nicked space representation from the corresponding nicked space representation, a first pathogenicity score for the substituted amino acid for the corresponding nicked amino acid is generated; and Based on the non-notch space representation from the corresponding non-notch space representation, a second pathogenicity fraction of the substituted amino acids at the corresponding positions in the non-notch benign protein sample or the non-notch pathogenic protein sample is generated.

19. The non-transitory computer-readable storage medium according to claim 18, wherein: At least some of the training cycles in the one or more training cycles use the same number of gapped space representations and ungapped space representations; or At least some of the training cycles use training example batches with the same number of gapped and ungapped space representations.

20. The non-transitory computer-readable storage medium of claim 19, wherein the notched spatial representation and the non-notched spatial representation are weighted differently such that the contribution of the notched spatial representation to gradient updates of parameters applied to the pathogenicity predictor in response to the pathogenicity predictor processing the non-notched spatial representation is different from the contribution of the non-notched spatial representation to gradient updates of the parameters applied to the pathogenicity predictor in response to the pathogenicity predictor processing the non-notched spatial representation.

Images