Computer-implemented method for identifying rare variants that cause extreme gene expression levels

The causal relationship between rare variants and extreme gene expression levels is solved through artificial intelligence technology, and the problem of difficulty in identifying and predicting rare variants in the prior art is improved, and the recognition accuracy and gene expression changes of pathogenic variants are improved. sensitivity.

CN118922887APending Publication Date: 2024-11-08ILLUMINA INC
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202380029488.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-08-05
Filing Date
2023-07-21
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art is difficult to effectively identify and predict rare variants that cause extreme gene expression levels, especially in the presence of multiple confounding factors, which lead to difficulties in identifying and diagnosing pathogenic variants.

Method used

Using artificial intelligence technology, by obtaining individual gene expression levels, normalizing and identifying outliers, selecting rare variants, and using causal relationship model fitting to determine the causal relationship between variants and extreme gene expression levels, controlling multiple confounding factors, generating causal scores to identify pathogenic variants.

Benefits of technology

Improves the accuracy of identification and pathogenic prediction of rare variants, enhances sensitivity and sensitivity to changes in gene expression, and simplifies understanding of disease diagnosis and treatment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118922887A_ABST
    Figure CN118922887A_ABST
Patent Text Reader

Abstract

The technology disclosed herein relates to reliably identifying variants that cause extreme gene expression levels. Extreme gene expression levels include expression insufficiency and expression overexpression. The variants can then be used to train an artificial intelligence-based model, thereby completing a variety of prediction tasks. One example of these predictive tasks is to produce per base resolution of chromatin sequences. Another example of the chromatin task is to produce gene expression changes caused by reliably identified variants.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The technology disclosed in the present invention relates to artificial intelligence type computers and digital data processing systems and corresponding data processing methods and products for simulating intelligence (i.e., knowledge-based systems, inference systems, and knowledge acquisition systems); and includes systems for uncertainty inference (e.g., fuzzy logic systems), adaptive systems, machine learning systems, and artificial neural networks. In particular, the technology disclosed in the present invention relates to epigenetics based on artificial intelligence at base resolution.

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application is related to concurrently filed U.S. Patent Application entitled “ARTIFICIAL INTELLIGENCE-BASED DETECTION OF GENE CONSERVATION AND EXPRESSION PRESERVATION AT BASE RESOLUTION” (Attorney Docket No. ILLM 1036-1 / IP-2045-PRV), which is incorporated by reference as if fully set forth herein for all purposes.

[0004] Literature Incorporation

[0005] The following documents are incorporated by reference as if fully set forth herein for all purposes:

[0006] U.S. Patent Application No. 62 / 903,700, entitled “ARTIFICIAL INTELLIGENCE-BASED EPIGENETICS,” filed on September 20, 2019 (Attorney Docket No. ILLM 1025-1 / IP-1898-PRV);

[0007] Sundaram, L. et al., Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018);

[0008] Jaganathan,K. et al., Predicting splicing from primary sequence with deeplearning. Cell 176,535-548(2019);

[0009] U.S. Patent Application No. 62 / 573,144, entitled “TRAINING A DEEP PATHOGENICITY CLASSIFIERUSING LARGE-SCALE BENIGN TRAINING DATA,” filed on October 16, 2017 (Attorney Docket No. ILLM 1000-1 / IP-1611-PRV);

[0010] U.S. Patent Application No. 62 / 573,149, entitled “PATHOGENICITY CLASSIFIER BASED ON DEEP CONVOLUTIONAL NEURAL NETWORKS (CNNs)”, filed on October 16, 2017 (Attorney Docket No. ILLM 1000-2 / IP-1612-PRV);

[0011] U.S. Patent Application No. 62 / 573,153, entitled “DEEP SEMI-SUPERVISED LEARNING THATGENERATES LARGE-SCALE PATHOGENIC TRAINING DATA,” filed on October 16, 2017 (Attorney Docket No. ILLM 1000-3 / IP-1613-PRV);

[0012] U.S. Patent Application No. 62 / 582,898, filed on November 7, 2017, entitled “PATHOGENICITY CLASSIFICATION OF GENOMICDATA USING DEEP CONVOLUTIONAL NEURAL NETWORKS (CNNs)” (Attorney Docket No. ILLM 1000-4 / IP-1618-PRV);

[0013] U.S. Patent Application No. 16 / 160,903, filed on October 15, 2018, entitled “DEEP LEARNING-BASED TECHNIQUES FORTRAINING DEEP CONVOLUTIONAL NEURAL NETWORKS” (Attorney Docket No. ILLM 1000-5 / IP-1611-US);

[0014] U.S. Patent Application No. 16 / 160,986, filed on October 15, 2018, entitled “DEEP CONVOLUTIONAL NEURAL NETWORKS FOR VARIANT CLASSIFICATION” (Attorney Docket No. ILLM 1000-6 / IP-1612-US);

[0015] U.S. Patent Application No. 16 / 160,968, filed on October 15, 2018, entitled “SEMI-SUPERVISED LEARNING FOR TRAINING ANENSEMBLE OF DEEP CONVOLUTIONAL NEURAL NETWORKS” (Attorney Docket No. ILLM 1000-7 / IP-1613-US);

[0016] U.S. Patent Application No. 16 / 407,149, entitled “DEEP LEARNING-BASED TECHNIQUES FOR PRE-TRAINING DEEP CONVOLUTIONAL NEURAL NETWORKS,” filed on May 8, 2019 (Attorney Docket No. ILLM 1010-1 / IP-1734-US);

[0017] U.S. Patent Application No. 17 / 232,056, filed on April 15, 2021, entitled “DEEP CONVOLUTIONAL NEURAL NETWORKS TOPREDICT VARIANT PATHOGENICITY USING THREE-DIMENSIONAL (3D) PROTEIN STRUCTURES” (Attorney Docket No. ILLM 1037-2 / 1P-2051-US);

[0018] U.S. Patent Application No. 63 / 175,495, entitled “MULTI-CHANNEL PROTEIN VOXELIZATION TOPREDICT VARIANT PATHOGENICITY USING DEEP CONVOLUTIONAL NEURAL NETWORKS,” filed on April 15, 2021 (Attorney Docket No. ILLM 1047-1 / IP-2142-PRV);

[0019] U.S. Patent Application No. 63 / 175,767, entitled “EFFICIENT VOXELIZATION FOR DEEP LEARNING,” filed on April 16, 2021 (Attorney Docket No. ILLM 1048-1 / IP-2143-PRV); and

[0020] U.S. patent application Ser. No. 17 / 468,411 (Agent Docket No. ILLM 1037-3 / IP-2051A-US), filed on Sept. 7, 2021, and entitled “ARTIFICIAL INTELLIGENCE-BASED ANALYSIS OF PROTEIN THREE-DIMENSIONAL (3D) STRUCTURES.” Background Art

[0021] The subject matter discussed in this section should not be considered to be prior art simply because it is mentioned in this section. Similarly, problems mentioned in this section or associated with the subject matter provided as background technology should not be considered to have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, which themselves may also correspond to specific implementations of the technology protected by the claims.

[0022] Genomics, also referred to in a broad sense as functional genomics, aims to characterize the function of each genomic element of an organism by 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 genotype and phenotype, discovering biomarkers for patient stratification, predicting gene function, and mapping biochemically active genomic regions and residues such as transcriptional enhancers and single nucleotide polymorphisms (SNPs).

[0023] Genomic data are too large and complex to be mined by visualizing pairwise correlations alone. For example, protein sequences can be classified into families of homologous proteins that originate from an ancestral protein and share similar structures and functions. Analyzing multiple sequence alignments (MSAs) of homologous proteins provides important information about functional and structural constraints. Statistics of MSA columns representing amino acid sites identify functional residues that have been conserved during evolution. The correlation of amino acid usage between MSA columns includes important information about functional segments and structural contacts.

[0024] Instead, analytical tools are needed to support the discovery of unexpected relationships, to derive new hypotheses and models, and to make predictions. Unlike some algorithms where assumptions and domain expertise are hard-coded, machine learning algorithms are designed to automatically detect patterns in data. As a result, machine learning algorithms are well suited to data-driven science, and are particularly well suited to genomics. However, the performance of machine learning algorithms can strongly depend on how the data is represented, that is, how each variable (also called a feature) is calculated. For example, to classify a tumor as malignant or benign from a fluorescence microscopy image, a preprocessing algorithm may detect cells, identify cell types, and generate a list of cell counts for each cell type.

[0025] Machine learning models can use estimated cell counts (an example of handcrafted features) as input features to classify tumors. The core problem is that classification performance heavily depends on the quality and relevance of these features. For example, relevant visual features (such as cell morphology, distance between cells, or localization within an organ) are not captured in cell counts, and this incomplete representation of the data may reduce classification accuracy.

[0026] Deep learning, a subdiscipline of machine learning, addresses this problem by embedding the computation of features into the machine learning model itself to produce an end-to-end model. This has been achieved by developing deep neural networks, which are machine learning models that consist of successive elementary operations that compute increasingly complex features by taking the results of previous operations as input. Deep neural networks are able to improve prediction accuracy by discovering relevant features of high complexity, such as cell morphology and the spatial organization of cells in the above example. The construction and training of deep neural networks has been made possible by the explosion of data, advances in algorithms, and a significant increase in computing power, particularly through the use of graphics processing units (GPUs).

[0027] 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 an intron is spliced ​​out given a (target) feature on RNA, such as the presence or absence of a typical splice site sequence, the location of a splice branch point, or the length of an intron. Training a machine learning model refers to learning its parameters, which usually involves minimizing a loss function on the training data, with the goal of making accurate predictions on unseen data.

[0028] For many supervised learning problems in computational biology, the input data can be represented as a table with multiple columns or features, each containing numerical or categorical data that can potentially be used to make predictions. Some input data are naturally represented as features in a table (such as temperature or time), while other input data need to be first transformed (such as transforming deoxyribonucleic acid (DNA) sequences into k-mer counts) using a process called feature extraction to fit a tabular representation. For the intron splicing prediction problem, the presence or absence of typical splice site sequences, the location of splicing branch points, and the length of introns can be preprocessed features collected in a 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), as well as many other models.

[0029] 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 computing the weighted sum of input features mapped to the interval [0,1] using a sigmoid function (a first-class activation function). The parameters of logistic regression or other linear classifiers using different activation functions are the weights in the weighted sum. Linear classifiers fail when the weighted sum of the input features cannot distinguish the classes well (e.g., the classes of spliced ​​or non-spliced ​​introns). To improve the prediction performance, new input features can be added manually by transforming or combining existing features in new ways (e.g., by exponentiation or pairwise products).

[0030] 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 organize the input features into related complex patterns that help with the task of distinguishing between two classes.

[0031] Deep neural networks use many hidden layers, where each neuron receives input from all neurons in the previous layer, and is called a fully connected layer. Neural networks are typically trained using stochastic gradient descent, an algorithm suitable 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 for many genomics applications, including predicting the percentage of exons spliced ​​for a given sequence from sequence features (such as the presence of binding motifs for splicing factors 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 marks, gene expression, and evolutionary conservation.

[0032] For effective prediction, local dependencies of both spatial and longitudinal data must be accounted for. For example, shuffling the pixels of a DNA sequence or an image can severely disrupt information patterns. These local dependencies set spatial or longitudinal data apart from tabular data, for which the ordering of features is arbitrary. Consider the problem of classifying genomic regions as bound or not bound by a specific transcription factor, where bound 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 examples in the sequence or position weight matrix (PWM) matching, can be used for this task. Since k-mer or PWM example frequencies are robust to shifting motifs within a sequence, such models generalize well to sequences with the same motif at different positions. However, they fail to recognize patterns where transcription factor binding depends on the combination of multiple motifs with well-defined spacing. Furthermore, the number of possible k-mers increases exponentially with the k-mer length, which poses challenges in terms of both storage and overfitting.

[0033] A convolutional layer is a special form of a fully connected layer in which the same fully connected layer is applied locally (e.g., in a 6bp window) to all sequence positions. This approach can also be viewed as using multiple PWMs to scan a sequence, for example, for the transcription factors GATA1 and TALI. By using the same model parameters at different positions, the total number of parameters is dramatically reduced, and the network is able to detect motifs at positions not seen during training. Each convolutional layer scans the sequence with several filters by producing a scalar value at each position that quantifies the fit between the filter and the sequence. As in a fully connected neural network, a nonlinear activation function (usually ReLU) is applied at each layer. Next, a pooling operation is applied that aggregates activations into continuous bins across the position axis, typically taking the maximum activation or average activation for each channel. Pooling reduces the effective sequence length and makes the signal coarser. Subsequent convolutional layers compose the output of the previous layer and are able to detect whether the GATA1 motif and the TAL1 motif are present 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.

[0034] Convolutional neural networks (CNNs) are able to predict a variety of molecular phenotypes based on DNA sequences alone. Applications include classification of transcription factor binding sites, and prediction of molecular phenotypes such as chromatin features, DNA contact maps, DNA methylation, gene expression, translation efficiency, RBP binding and microRNA (miRNA) targets. In addition to predicting molecular phenotypes from sequences, convolutional neural networks can also be applied to more technical tasks that are traditionally solved by manual bioinformatics pipelines. For example, convolutional neural networks can predict the specificity of guide RNAs, denoise ChIP-seq, improve the resolution of Hi-C data, predict the laboratory of origin from DNA sequences, and detect genetic variants. Convolutional neural networks have also been used to model long-range dependencies in the genome. Although interacting regulatory elements may be located far from each other on unfolded linear DNA sequences, these elements are usually adjacent to each other in the actual 3D chromatin conformation. Therefore, although modeling molecular phenotypes by linear DNA sequences is a crude 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 dilated convolutions with a receptive field of up to 32 kb. Dilated convolutions also allow prediction of splice sites from sequences using a receptive field of 10 kb, enabling integration of gene sequences across distances as long as typical human introns (see Jaganathan, K. et al., Predicting splicing from primary sequence with deep learning. Cell 176, 535-548 (2019)).

[0035] Different types of neural networks can be characterized by their parameter sharing schemes. For example, fully connected layers do not have parameter sharing, while convolutional layers impose translation invariance by applying the same filter at each position of their input. Recurrent neural networks (RNNs) are alternatives to convolutional neural networks for processing sequential data (such as DNA sequences or time series) that implement different parameter sharing schemes. Recurrent neural networks apply the same operation to each sequence element. The operation takes the previous sequence element and the new input as input to the memory. The operation updates the memory and optionally emits an output, which is passed to subsequent layers or is directly used as a model prediction result. Since the same model is applied at each sequence element, the recurrent neural network remains unchanged for the position index in the sequence being processed. For example, a recurrent neural network can detect an open reading frame in a DNA sequence, regardless of the position in the sequence. This task requires identifying a specific series of inputs, such as an in-frame stop codon followed by a start codon.

[0036] The main advantage of recurrent neural networks over convolutional neural networks is that, in theory, they can carry information through sequences of infinite length via memory. In addition, recurrent neural networks can naturally process sequences with large length variations, such as mRNA sequences. However, in terms of sequence modeling tasks (such as audio synthesis and machine translation), convolutional neural networks combined with various techniques (such as dilated convolutions) can achieve performance comparable to or even better than recurrent neural networks. Recurrent neural networks can aggregate the outputs of convolutional neural networks for predicting single-cell DNA methylation status, RBP binding, transcription factor binding, and DNA accessibility. In addition, since recurrent neural networks apply sequential operations, they cannot be easily parallelized, so the calculation speed is much slower than that of convolutional neural networks.

[0037] Although each person has a unique genetic code, most of the human genetic code is shared by all people. In some cases, the human genetic code may include outliers, called genetic variants, which may be shared among individuals in a relatively small group of the population. For example, a particular human protein may contain a particular amino acid sequence, while a variant of that protein may differ by one amino acid in an otherwise identical particular sequence.

[0038] Genetic variants can be pathogenic, thus causing disease. Although most such genetic variants have been depleted from the genome by natural selection, the ability to identify which genetic variants are likely to be pathogenic can help researchers focus on these genetic variants to gain an understanding of the corresponding diseases and their diagnosis, treatment, or cure. The clinical interpretation of the millions of human genetic variants remains unclear. Some of the most common pathogenic variants are single nucleotide missense mutations that change the amino acid of a protein. However, not all missense mutations are pathogenic.

[0039] Models that can predict molecular phenotypes directly from biological sequences can be used as computer perturbation tools to detect associations between genetic and phenotypic variation, and have become new methods for quantitative trait loci identification and variant prioritization. These methods are very important because most variants identified by genome-wide association analyses of complex phenotypes are non-coding, which makes it challenging to estimate their effects and contributions to the phenotype. In addition, linkage disequilibrium causes blocks of variants to be co-inherited, which creates difficulties in pinpointing single causal variants. Therefore, sequence-based deep learning models that can be used as interrogation tools to assess the effects of such variants provide a promising approach to discover potential drivers of complex phenotypes. One example includes 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 predictions. Another example includes predicting the generation of new splice sites based on sequence or based on the quantitative effects of genetic variants on splicing.

[0040] An end-to-end deep learning method for predicting variant effects is 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 herein as "PrimateAI"). PrimateAI uses a deep neural network trained on variants known to be pathogenic, where cross-species information is used for data enhancement. In particular, PrimateAI uses the sequences of wild-type proteins and mutant proteins to compare differences, and uses trained deep neural networks to determine the pathogenicity of mutations. This method of using protein sequences for pathogenicity prediction is promising because it can avoid circularity problems and overfitting of prior knowledge. However, compared with the number of data for effectively training deep neural networks, the number of clinical data available in ClinVar is relatively small. In order to overcome this lack of data, PrimateAI uses common human variants and primate variants as benign data, and simulated variants based on trinucleotide backgrounds are used as unlabeled data.

[0041] PrimateAI outperforms existing methods when trained directly on sequence alignments. PrimateAI learns important protein domains, conserved amino acid positions, and sequence dependencies directly from training data consisting of approximately 120,000 human samples. PrimateAI significantly outperforms 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 is an important advancement in variant classification tools that can reduce the reliance of clinical reports on prior knowledge.

[0042] At the core of protein biology is understanding how structural elements give rise to observed functions. The glut of protein structural data has enabled the development of computational methods to systematically derive the rules governing structure-function relationships. However, the performance of these methods depends critically on the choice of protein structural representation.

[0043] Protein sites are microenvironments within protein structures that are distinguished by their structural or functional roles. Sites can be defined by a three-dimensional (3D) position and a local neighborhood around the position where structure or function exists. The core of rational protein engineering is to understand 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 helps to engineer and alter protein function. Identifying functionally or structurally important amino acids allows for focused engineering efforts, such as site-directed mutagenesis for altering the functional properties of a target protein. Alternatively, this knowledge can help to avoid engineering designs that would destroy the desired function.

[0044] Since it has been established that structure is much more conserved than sequence, the increase in protein structural data provides an opportunity to systematically study the underlying patterns that govern structure-function relationships using data-driven methods. A fundamental aspect of any computational analysis of proteins 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. A good representation efficiently captures the most critical information, while a poor representation produces a noisy distribution without underlying patterns.

[0045] The plethora of protein structures and the recent success of deep learning algorithms provide opportunities to develop tools for automatically extracting task-specific representations of protein structures.

[0046] Computational analysis of genomic studies is challenged by confounding variants that are unrelated to the genetic factor of interest. Identifying variants that cause extreme gene expression levels (high or low) is critical for diagnosing the pathogenicity of genetic diseases. However, there are many confounding factors that may interfere with the identification of pathogenic variants. Isolating variants by examining rare variants that may be associated with a specific pathology can simplify this problem. Additionally, removing the noise introduced by confounding factors can improve the signal-to-noise ratio.

[0047] Therefore, an opportunity arises to apply artificial intelligence to epigenetics to greatly increase the sensitivity of recovering genetic correlations between variable loci and individual gene expression levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] This patent or patent application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee. Color drawings are also available in pairs via the Supplemental Content tab.

[0049] In the drawings, similar reference symbols generally refer to similar parts throughout the different views. In addition, the drawings are not necessarily drawn to scale, but rather focus on illustrating the principles of the disclosed technology. In the following description, various embodiments of the technology disclosed in the present invention are described with reference to the following drawings, in which:

[0050] Figure 1 One specific implementation of a processing pipeline is shown that identifies rare variants that cause extreme gene expression levels.

[0051] Figure 2 A specific implementation of a processing system is presented that generates causality scores for outlier variants that cause extreme gene expression levels.

[0052] Figure 3 Depicts a specific implementation of a fitted causal model that determines causal relationships between rare variants and extreme gene expression levels while controlling for multiple confounding factors.

[0053] Figure 4 An example of a causality score generated by the disclosed techniques for a rare variant sample is shown.

[0054] Figure 5 It was shown that performance outcomes (measured as z scores across individuals and genes) gradually improved by successively adjusting for multiple confounders using a fitted causal model.

[0055] Figure 6 The counts of true rare variants and shuffled (or randomly selected) variants identified as causing gene under-expression and gene over-expression for a given p-value from the fitted causal model are compared.

[0056] Figure 7 Shown are the odds ratios comparing causality of true rare variants and shuffled variants in terms of gene under-expression and gene over-expression for a given p-value from the fitted causality model.

[0057] Figure 8 A first example architecture of the chromatin model disclosed in the present invention is shown.

[0058] Fig. 9 A second example architecture of the chromatin model disclosed in the present invention is shown.

[0059] Fig.10 A third example architecture of the chromatin model disclosed in the present invention is shown.

[0060] Fig.11 A fourth example architecture of the chromatin model disclosed in the present invention is shown.

[0061] Fig.12 A fifth example architecture of the chromatin model disclosed in the present invention is shown.

[0062] Fig.13 Shows the input generation logic components that access the sequence database and generate the input base sequence.

[0063] Fig.14 Describes a specific implementation of base-resolution evolutionary conservation prediction using chromatin models.

[0064] Fig.15 An example of an output sequence corresponding to a target base sequence is shown.

[0065] Fig.16 A specific implementation of the gene expression model disclosed in the present invention is shown.

[0066] Fig.17 One example of a reference sequence and an alternative sequence is shown.

[0067] Fig.18 A specific implementation of the variant classification logic component disclosed in the present invention is shown.

[0068] Fig.19 A specific implementation of the pathogenicity prediction logic component disclosed in the present invention is shown.

[0069] Fig. 20 is an example computer system that can be used to implement various aspects of the disclosed technology. DETAILED DESCRIPTION

[0070] The following discussion is presented to enable any person skilled in the art to make and use the disclosed technology, and is provided in the context of a specific application and its requirements. Various modifications to the disclosed implementations will be apparent to those skilled in the art, and the general principles defined herein may be applied to other implementations 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 implementations shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0071] The detailed description of various implementations will be better understood when read in conjunction with the accompanying drawings. To the extent that the accompanying drawings illustrate diagrams of functional blocks of various implementations, the functional blocks do not necessarily indicate a division between hardware circuits. Thus, for example, one or more of the functional blocks (e.g., modules, processors, or memories) may be implemented in a single piece of hardware (e.g., a general purpose signal processor or random access memory block, a hard disk, etc.) or in multiple pieces of hardware. Similarly, a program may be a stand-alone program, may be incorporated into an operating system as a subroutine, may be a function in an installed software package, etc. It should be understood that the various implementations are not limited to the arrangements and tools shown in the accompanying drawings.

[0072] The processing engines and databases designated as modules in the accompanying drawings may be implemented in hardware or software and need not be divided into identical blocks exactly as shown in the accompanying drawings. Some of these modules may also be implemented on different processors, computers, or servers, or distributed between multiple different processors, computers, or servers. In addition, it should be understood that some of the modules may be combined, operated synchronously, or operated in a sequence different from that shown in the drawings without affecting the functions implemented. The modules in the accompanying drawings may also be considered as flow chart steps in the method. A module also does not necessarily need to have all of its code placed continuously in memory; some parts of the code may be separated from other parts of the code, with code from other modules or other functions arranged between the two.

[0073] Example applications of the disclosed technology

[0074] Gene expression is the process of converting instructions in DNA into functional products (such as RNA molecules or proteins). The technology disclosed in the present invention identifies rare variants that cause extreme gene expression levels, wherein extreme gene expression levels include both underexpression and overexpression. These rare variants are identified by being associated with neighboring genes with extreme gene expression levels. In a specific implementation, the technology disclosed in the present invention identifies those individuals in which there is a specific variant in the promoter region of a gene and the gene also has significantly different gene expression compared to individuals without the specific variant. Based on the identification of such individuals, the technology disclosed in the present invention classifies specific variants as gene expression change variants. The technology disclosed in the present invention also uses artificial intelligence to train multiple models, wherein the rare variants identified are used as training data, and the underexpression and overexpression phenotypes of these rare variants are used as ground truth labels.

[0075] In this application, the terms "chromatin sequence", "chromatin input sequence" and "output sequence" are used throughout this application. Chromatin is DNA bound to proteins and / or RNA. By using the term "chromatin sequence", we refer to the DNA sequence of chromatin. A DNA sequence in a section of chromatin may be protected by proteins and RNA and then sequenced like DNA footprinting. Chromatin sequences can also be chemically modified. For example, DNA sequences typically have methyl groups attached to the nucleotides in the sequence and are therefore methylated.

[0076] As highlighted by numerous comparisons of various implementations of the chromatin model 124, many implementations share overlap in architectural components. Each element of the chromatin model 124 has multiple implementations that can be combined in multiple configurations. The multiple arrangements that can be implemented for the technology disclosed in the present invention provide not only a wide range of practicality, performance efficiency, but also a wide range of performance accuracy. The data transformation applied to the input base sequence in many implementations of the technology disclosed in the present invention, generating a variety of additional sequence formats from the perspective of nucleic acid sequence and the perspective of chromatin structure, is an innovative strategy that causes the output of a large number of output signals and has wide applicability to a wide range of genomics, protein analysis and pathogenicity research problems. Previous versions of PrimateAI have used multiple tools to classify variant pathogenicity with high performance. The chromatin model 124 introduces another tool in this approach, as well as an additional dimension when studying epigenetic signals that affect biological replication and transcription processes.

[0077] Although the additional vantage point of chromatin structure has obvious utility for studying gene expression using the various tools provided by PrimateAI, the real impact provided by the technology disclosed in the present invention is to add epigenetic signals to the entire gene expression prediction logic component. Both the DNA sequence and the histone protein components of chromatin can undergo a large number of chemical modifications. Enzymes that directly bind to chromatin components and catalyze chemical modifications of chromatin components can change the chromatin structure, and changes in chromatin structure can also change the ability of chromatin-interacting enzymes to approach their target ligands and the function of chromatin-interacting enzymes. The structure of chromatin and the enzymes that change the structure directly affect the accessibility of gene transcription and expression. DNA variants can cause changes in chromatin structure, which can subsequently change epigenetic effects, such as transcription factor binding, and enzymatic reactions necessary for proper regulation of gene expression and gene repression.

[0078] In contrast, epigenetic effects on chromatin (such as methylation and protein binding events) can affect mutation rates, potentially introducing variants that may be silent or pathogenic. As demonstrated in many specific implementations of the technology disclosed in the present invention, studies of evolutionary constraints on genes and the pathogenicity of gene variants are significantly more comprehensive and accurate when enhanced by epigenetic features. In summary, the technology disclosed in the present invention has several arrangements that are subject to a series of training and learning strategies to generate several outputs that can be applied to predict gene expression and gene pathogenicity of target genetic sequences. The chromatin focusing strategy disclosed in the present invention can be used to study genetic diseases and environmental exposure-related diseases, develop drugs, and study the effects of epigenetics in the transcription and translation of nucleic acid sequences into proteins.

[0079] Rare variant identification

[0080] Processing pipeline

[0081] Figure 1 A specific implementation of the processing pipeline 100 is shown, which identifies rare variants that cause extreme gene expression levels. At act 110, gene expression levels for a set of individuals are obtained. In a specific implementation, the gene expression levels are obtained from genotype-tissue expression (GTEx).

[0082] At act 120 , the gene expression levels are normalized, for example, by calculating a mean and a plurality of standard deviations relative to the mean.

[0083] At action 130, outlier individuals with extreme gene expression levels are identified from the group of individuals. The extreme gene expression levels are determined from the tail quantiles 124 of the normalized gene expression levels 122. Examples of tail quantiles 124 include one or more standard deviations relative to the mean in both positive and negative directions. For example, the z-score of the gene expression level of the outlier individual relative to the mean is at least |1.2|.

[0084] At action 140, rare variants are selected from the gene sequence of the outlier individual. The rare variants are selected based on an allele frequency cutoff. For example, the rare variant has a minor allele frequency (MAF) of less than 0.1%.

[0085] At action 150, a causal relationship model is fitted to determine the causal relationship between the rare variants and the extreme gene expression levels in the outlier individuals while controlling for multiple confounding factors. In one specific implementation, the fitted causal relationship model determines the causal relationship by predicting the specific gene expression level of a specific gene in a specific chromosome based on the variant-driven gene expression level caused by the specific rare variant. In one specific implementation, the fitted causal relationship model measures the contribution of the variant-driven gene expression level as a variant effect covariate.

[0086] Examples of causal models include logistic regression models, linear regression models, analysis of covariance (ANCOVA) models, and / or multivariate analysis of covariance (MANCOVA) models. Examples of multiple confounding factors include distal trans-expression quantitative trait loci (eQTL) effects, local cis-eQTL effects, genotype-based principal components (gPCs), expression residual (PEER) effects, environmental effects, population structure and ancestral effects, as well as sex effects, batch effects, genotyping platform effects, and library construction scheme effects. PEER stands for "probability estimation of expression residuals." This is a collection of Bayesian methods that use factor analysis methods to infer hidden determinants and their effects from gene expression profiles.

[0087] At action 160, a causality score for a rare variant is generated based on the determined causality. The specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant. In a specific implementation, the causality score is a probability value (p-value). In a specific implementation, the p-value is determined by the Pearson correlation coefficient.

[0088] Processing system

[0089] Figure 2 A specific implementation of the display processing system 200 is shown, which generates a causal relationship score for an outlier variant causing an extreme gene expression level. A data storage 202 includes gene expression data, which for example derives from GTEx, RNA-Seq or whole genome sequencing (WGS). A normalizer 204 normalizes the gene expression data, and stores the normalized gene expression data in a data storage 206. The normalized gene expression data can be measured by a mean value and one or more standard deviations relative to the mean value. In a specific implementation, the extreme gene expression level is determined from the tail quantile of the normalized gene expression data.

[0090] The causal relationship model 212 is fitted to determine the causal relationship between the variant 216 and the extreme gene expression level while controlling for multiple confounding factors 226. In one embodiment, the fitted causal relationship model 212 determines the causal relationship by predicting the specific gene expression level of a specific gene in a specific chromosome based on the variant-driven gene expression level caused by the specific variant (which is referred to herein as the gene expression caused by the variant 214). In one embodiment, the fitted causal relationship model 212 measures the variant-driven gene expression level 214 (caused by the variant 216) as a variant effect covariate (later in Figure 3 In one embodiment, the fitted causal model 212 measures the confounder-driven gene expression level 224 (caused by the confounder 226) as a multiple confounder effect covariate (later in Figure 3 The contributions are shown as 304, 306 and 310).

[0091] Examples of causal relationship models 212 include logistic regression models, linear regression models, analysis of covariance (ANCOVA) models, and / or multivariate analysis of covariance (MANCOVA) models. Examples of confounding factors 226 include distal trans-expression quantitative trait loci (eQTL) effects, local cis-eQTL effects, genotype-based principal components (gPCs), expression residual (PEER) effects, environmental effects, population structure and ancestral effects, as well as sex effects, batch effects, genotyping platform effects, and library construction protocol effects.

[0092] The causal model 212 generates as output the confounder-corrected normalized gene expression 232 caused by the variant 216 , in which the effect of the confounder 226 is regressed out.

[0093] Rare variant identifier 234 identifies rare variant 236 from among variants 216. Rare variant 216 can be selected based on an allele frequency cutoff value. For example, rare variant 236 can have a minor allele frequency (MAF) of less than 0.1%.

[0094] The causality score generator 242 generates a causality score 246 for a rare variant 236 based on the normalized gene expression 232 corrected for confounding factors. The specific causality score of a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant. In a specific implementation, the causality score 246 is a probability value (p-value). In a specific implementation, the p-value is determined by the Pearson correlation coefficient.

[0095] Rare variant ranker 256 ranks rare variants 236 based on causality scores 246 and then stores the ranked rare variants in data store 252 .

[0096] Causal Model

[0097] Figure 3 A specific implementation of a fitted causal relationship model 300 is depicted, which determines a causal relationship between rare variants and extreme gene expression levels while controlling for multiple confounding factors. The fitted causal relationship model 300 has a dependent variable 302 named "G" (gene expression), and multiple independent variables 304, 306, 308, and 310 having the following names, respectively:

[0098] “G1” (gene expression caused by distal trans-eQTL effects),

[0099] “G2” (gene expression caused by local cis-eQTL effects),

[0100] "G3" (gene expression caused by rare variants), and

[0101] “G4” (gene expression caused by other effects, such as genotype-based principal components (gPC) 320, expression residual (PEER) effects 330, environmental effects 340, population structure and ancestry effects 350, as well as sex effects, batch effects, genotyping platform effects, and library construction protocol effects 360).

[0102] The technology disclosed in the present invention removes G1, G2 and G4 from G, that is, G-G1-G2-G4=G3, to eliminate (or correct) the confounding effect and accurately determine the gene expression caused by rare variants.

[0103] The fitted causal model 300 controls for distal trans eQTL effects by predicting specific gene expression levels (G) 302 from trans gene expression levels caused by other genes in other chromosomes (G1) 304. In one implementation, the fitted causal model 300 measures the contribution of trans gene expression levels as a trans effect covariate.

[0104] The fitted causal model 300 controls for local cis-eQTL effects by predicting a specific gene expression level (G) 302 based on the cis-gene expression levels (G2) 306 caused by the presence of multiple common variants in the neighborhood of the specific gene. In some implementations, the neighborhood is defined by an offset relative to the transcription start site (TSS) in the specific gene. In one implementation, the fitted causal model 300 measures the contribution of the cis-gene expression level as a cis-effect covariate.

[0105] The fitted causal model 300 controls for population structure and ancestral effects by predicting specific gene expression levels (G) 302 based on gPC gene expression levels (G4) 310 caused by gPC. In one implementation, the fitted causal model 300 measures the contribution of gPC gene expression levels (G4) 310 as a covariate for population structure and ancestral effects.

[0106] The fitted causal model 300 controls for the PEER effect by predicting a specific gene expression level (G) 302 based on the PEER gene expression level (G4) 310 caused by PEER. In one implementation, the fitted causal model 300 measures the contribution of the PEER gene expression level (G4) 310 as a covariate of the PEER effect.

[0107] The fitted causal model 300 controls for environmental effects by predicting specific gene expression levels (G) 302 based on environmental gene expression levels (G4) 310 caused by the environmental effects. In one implementation, the fitted causal model 300 measures the contribution of environmental gene expression levels (G4) 310 as a covariate of environmental effects.

[0108] In some implementations, extreme gene expression levels include gene overexpression and gene underexpression.

[0109] Overexpression

[0110] In one implementation, the fitted causal relationship model 300 determines the causal relationship between rare variants and gene overexpression while controlling for multiple confounding factors. In some implementations, the fitted causal relationship model 300 generates a so-called "excess causality score" for rare variants. The specific excess causality score of a particular rare variant indicates the likelihood that the particular rare variant causes gene overexpression in those outlier individuals whose gene sequences contain the particular rare variant.

[0111] In one implementation, the excess causality score is an excess probability value (excess p-value). In one implementation, the excess p-value is determined by the Pearson correlation coefficient. In some implementations, the excess p-value specifies the statistically unconfounded likelihood that a rare variant increases gene expression in a gene that otherwise has lower gene expression relative to other genes in the gene set.

[0112] Insufficient expression

[0113] In one implementation, the fitted causal relationship model 300 determines the causal relationship between rare variants and insufficient gene expression while controlling for multiple confounding factors. In some implementations, the fitted causal relationship model 300 generates a so-called "insufficient causality score" for rare variants. The specific insufficient causality score for a particular rare variant indicates the likelihood that the particular rare variant causes insufficient gene expression in those outlier individuals whose gene sequences contain the particular rare variant.

[0114] In one implementation, the insufficient causality score is an insufficient probability value (insufficient p-value). In one implementation, the insufficient p-value is determined by the Pearson correlation coefficient. In some implementations, the insufficient p-value specifies a statistically unconfounded likelihood that a rare variant reduces gene expression in a gene that otherwise has higher gene expression relative to other genes in the gene set.

[0115] In some implementations, the rare variants are non-coding variants. Non-coding variants may include 5' untranslated region (UTR) variants, 3'UTR variants, enhancer variants, and promoter variants.

[0116] In some implementations, gene expression levels are further stratified into tissue-specific gene expression levels of multiple tissues. In one implementation, quantile normalization is used to normalize the gene expression levels of each gene in each tissue. In some implementations, a causal relationship model is fitted separately for each tissue. In some implementations, stratification is used to fit the causal relationship model.

[0117] Ranking

[0118] In some implementations, a ranking of rare variants is generated based on a causality score. In one implementation, a ranking of rare variants is generated based on an excessive causality score. In one implementation, a ranking of rare variants is generated based on an insufficient causality score.

[0119] In some implementations, the rare variant is a single-occurring variant. In one implementation, the single-occurring variant occurs in only one outlier individual from the outlier individuals.

[0120] Causality score

[0121] Figure 4 An example of a causality score 400 generated by the disclosed techniques for a rare variant sample is shown. Figure 4 A "Chromosome" column 402 is shown, identifying the chromosome on which the rare variant is located. Figure 4 Also shown is a "Position" column 404, which identifies the position of the rare variant. Figure 4 Also shown is a "REF" column 406, which identifies the reference nucleotide (or base) corresponding to the rare variant. Figure 4 Also shown is an "ALT" column 408, which identifies alternative nucleotides (or bases) representing rare variants.

[0122] Figure 4 Also shown is a "p_deficient" column 410, which identifies the deficient causality score for rare variants. The p_deficient value is negatively correlated with the deficient causality score, i.e., the higher the p_deficient value, the less likely the rare variant is to cause deficient gene expression; the lower the p_deficient value, the more likely the rare variant is to cause deficient gene expression.

[0123] Figure 4 Also shown is a "p_over" column 412, which identifies the over-causality score of the rare variant. The p_over value is negatively correlated with the over-causality score, that is, the higher the p_over value, the less likely the rare variant is to cause over-expression of the gene; the lower the p_over value, the greater the likelihood that the rare variant is to cause over-expression of the gene.

[0124] Performance results as objective indicators of non-obviousness and inventive step

[0125] Comparing and detecting differences between a sample distribution and a reference distribution, or sample outliers relative to a reference distribution, can include the use of parametric and nonparametric statistical tests, such as (one-tailed or two-tailed) t-tests, Mann-Whitney RankSum tests, and other tests, including the use of z-scores, such as z-scores based on median absolute deviations (e.g., such as used by Stumm et al., 2014, Prenat Diagn 34:185). When comparing a distribution to a reference distribution (or an outlier relative to a reference distribution), in certain embodiments, the comparison is distinguished (and / or identified as significantly different) if: the mean, median, or individual samples are separated from the reference distribution by more than about 1.5, 1.6, 1.7, 1.8, 1.9, 1.95, 1.97, 2.0 standard distributions ("SD"), or more than about 2.0 standard distributions; and / or the individual samples are separated from the reference distribution by a z-score of greater than about 1.5, 1.6, 1.7, 1.8, 1.9, 2.0, 2.5, 2.7, 2.8, 2.9, 3.0, 3.1, 3.2, 3.3, 3.4, 3.5, 3.75, 4.0, 4.5, 5.0, or more than about 5.0.

[0126] In some implementations, parameters (such as mean, median, standard deviation, median absolute deviation, or z-score) are calculated with respect to a set of samples. In some such implementations, such calculated parameters are used to identify outliers from those test samples detected / analyzed. In particular implementations, such parameters are calculated from all test samples without knowing the identity of any outliers (e.g., "masked" analysis). In other particular implementations, such parameters are calculated from a set of reference samples known to be (non-outlier) standards or test samples assumed to be (or unlikely to be) such standards.

[0127] In some implementations, in the context of a data set, a z-score (or an equivalent statistic based on a repeated distribution pattern of a parameter) can be calculated to identify outlier data points (e.g., representing extreme gene expression levels (under- or over-expression)), data representing such data points removed from the data set, and then the data set subjected to a z-score analysis in an attempt to identify additional outliers. Such an iterative z-score analysis can sometimes be particularly useful when two or more samples may deviate from a single z-score analysis, and / or in situations where subsequent testing may be used to confirm false positives and thus avoiding false negatives may be more important than the (initial) identification of false positives.

[0128] Figure 5It is shown that the performance results 500 (measured by z scores across individuals and genes) gradually improve by successively correcting for multiple confounding factors using the fitted causal relationship model 300. In particular, an under-z score is determined to measure the correlation between rare variants and under-expression of a gene, and an over-z score is determined to measure the correlation between rare variants and over-expression of a gene.

[0129] At 502, the under-z score is 1.7 and the over-z score is 1.2. At this stage, none of the confounders were corrected.

[0130] At 512, the under-z score increases to 2.25 and the over-z score increases to 2. At this stage, the PEER factor is corrected.

[0131] At 522, the under-z score increased to 2.6 and the over-z score increased to 2.2. At this stage, 30 gPCs were corrected.

[0132] At 532, the deficiency z-score increases to 3 and the excess z-score decreases to 2. At this stage, the local cis-eQTL effects are corrected.

[0133] At 542, the deficiency z score increased to 3.5 and the excess z score increased to 3. At this stage, the distal trans-eQTL effect was corrected.

[0134] Figure 6 The counts 611, 613 of true rare variants 622 and shuffled (or randomly selected) variants 632 identified as causing gene underexpression 612 and gene overexpression 616 are compared for a given p-value 604 from the fitted causal model 300. Figure 6 As shown, the count-based correlation of the true rare variant 622 is always higher than that of the shuffled variant 632. Figure 6 In FIG, the x-axis is the distance 662, 666 of the true rare variant 622 and the shuffled variant 632 relative to the transcription start site (TSS). Figure 6 , the y-axis is the counts 611 , 613 of the true rare variant 622 and the shuffled variant 632 .

[0135] Figure 7 614. The odds ratios 711, 713 comparing the causality of the true rare variant 622 and the shuffled variant 632 for gene underexpression 612 and gene overexpression 616 are shown for a given p-value 604 from the fitted causality model 300. Figure 7 As shown, the consistently high odds ratios 711 , 713 at different TSS distances 662 , 666 demonstrate a strong causal relationship for the true rare variant 622 .

[0136] Base-resolution evolutionary conservation prediction

[0137] Chromatin Model

[0138] Figure 8 A first example architecture 800 of a disclosed chromatin model 802 is shown. Fig. 9 A second example architecture 900 of the disclosed chromatin model 802 is shown. Fig.10 A third example architecture 1000 of the disclosed chromatin model 802 is shown. Fig.11 A fourth example architecture 1100 of the disclosed chromatin model 802 is shown. Fig.12 A fifth example architecture 1200 of the disclosed chromatin model 802 is shown.

[0139] like Figure 8 As shown, the chromatin model 802 includes multiple groups of residual blocks arranged in order from low to high. Each group of residual blocks is parameterized by the number of convolution filters in the residual block, the convolution window size of the residual block, and the dilated convolution rate of the residual block. In some embodiments, the dilated convolution rate progresses in a non-exponential manner from a lower residual block group to a higher residual block group. In other embodiments, the dilated convolution rate progresses in an exponential manner. The convolution window size varies between multiple groups of residual blocks, and each residual block includes at least one batch normalization layer, at least one rectified linear unit (abbreviated as ReLU) layer, at least one dilated convolution layer, and at least one residual connection.

[0140] like Fig. 9 As shown, the dimension of the input is (Cu+L+Cd)×4, where Cu is the number of upstream flanking background bases, Cd is the number of downstream flanking background bases, and L is the number of bases in the input promoter sequence. The dimension of the output is 4×L. In some specific implementations, each residual block group produces an intermediate output by processing the previous input, and the dimension of the intermediate output is (I-[{(W-1)*D}*A])×N, where I is the dimension of the previous input, W is the convolution window size of the residual block, D is the dilated convolution rate of the residual block, A is the number of dilated convolution layers in the group, and N is the number of convolution filters in the residual block.

[0141] The example architecture 1000 is used when the input has 200 upstream flanking background bases (Cu) on the left side of the input sequence and 200 downstream flanking background bases (Cd) on the right side of the input sequence. The length (L) of the input sequence can be any value, such as 3001. In the example architecture 1000, each residual block in the first group has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 1; each residual block in the second group has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 4. In other architectures, each residual block has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 1.

[0142] The example architecture 1100 is used when the input has 1000 upstream flanking background bases (Cu) on the left side of the input sequence and 1000 downstream flanking background bases (Cd) on the right side of the input sequence. The length (L) of the input sequence can be any value, such as 3001. In the example architecture 1100, there are at least three groups of four residual blocks and at least three skip connections. Each residual block in the first group has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 1; each residual block in the second group has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 4; each residual block in the third group has 32 convolution filters, a convolution window size of 21, and a dilated convolution rate of 19.

[0143] The example architecture 1200 is used when the input has 5000 upstream flanking background bases (Cu) on the left side of the input sequence and 5000 downstream flanking background bases (Cd) on the right side of the input sequence. The length (L) of the input sequence can be any value, such as 3001. In the example architecture 1200, there are at least four groups of four residual blocks and at least four skip connections. Each residual block in the first group has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 1; each residual block in the second group has 32 convolution filters, a convolution window size of 11, and a dilated convolution rate of 4; each residual block in the third group has 32 convolution filters, a convolution window size of 21, and a dilated convolution rate of 19; each residual block in the fourth group has 32 convolution filters, a convolution window size of 41, and a dilated convolution rate of 25.

[0144] In general, the chromatin model 802 can be a rule-based model, a tree-based model, or a machine learning model. Examples include multi-layer perceptron (MLP), feed-forward neural network, fully connected neural network, fully convolutional neural network, ResNet, sequence to sequence (Seq2Seq) model (such as WaveNet), semantic segmentation neural network and generative adversarial network (GAN) (e.g., CycleGAN, StyleGAN, pixelRNN, text-2-image, DiscoGAN, IsGAN).

[0145] In some implementations, the chromatin model 802 can include a self-attention mechanism such as a Transformer, a Visual Transformer (ViT), a Bidirectional Transformer (BERT), a Detection Transformer (DETR), a Deformable DETR, UP-DETR, DeiT, Swin, GPT, iGPT, GPT-2, GPT-3, BERT, SpanBERT, RoBERTa, XLNet, ELECTRA, UniLM, BART, T5, ERNIE (THU), KnowBERT, DeiT-Ti, DeiT-S, DeiT-B, T2T-ViT-14, T2T-ViT-19 , T2T-ViT-24, PVT-Small, PVT-Medium, PVT-Large, TNT-S, TNT-B, CPVT-S, CPVT-S-GAP, CPVT-B, Swin-T, Swin-S, Swin-B, Twins-SVT-S, Twins-SVT-B, Twins-SVT-L, Shuffle-T, Shuffle-S, Shuffle-B, XCiT-S12 / 16, CMT-S, CMT-B, VOLO-D1, VOLO-D2, VOLO-D3, VOLO-D4, MoCo v3, ACT, TSP, Max-DeepLab, VisTR, SETR, Manual Transformer, HOT-Net, METRO, Image Transformer, Tamed Transformer, TransGAN, IPT, TTSR, STTN, Masked Transformer, CLIP, DALL-E, Cogview, UniT, ASH, TinyBert, FullyQT, ConvBert, FCOS, Faster R-CNN+FPN, DETR-DC5, TSP-FCOS, TSP-RCNN, ACT+MKDD(L=32), ACT+MKDD(L=16), SMCA, efficient DETR, UP-DETR, UP-DETR, ViTB / 16-FRCNN, ViT -B / 16-FRCNN, PVT-Small+RetinaNet, Swin-T+RetinaNet, Swin-T+ATSS, PVT-Small+DETR, TNT-S+DETR, YOLOS-Ti, YOLOS-S and YOLOS-B.

[0146] In some embodiments, examples of the chromatin model 802 include: a convolutional neural network (CNN) having multiple convolutional layers; a recurrent neural network (RNN), such as a long short-term memory network (LSTM), a bidirectional LSTM (Bi-LSTM), or a gated recurrent unit); and a combination of both CNN and RNN.

[0147] In some implementations, the chromatin model 802 may use 1D convolution, 2D convolution, 3D convolution, 4D convolution, 5D convolution, dilated or atrous 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 chromatin model 802 may use one or more loss functions, such as logistic regression / logarithmic loss function, multi-class cross entropy / softmax loss function, binary cross entropy loss function, mean square error loss function, L1 loss function, L2 loss function, smooth L1 loss function, and Huber loss function. The chromatin model 802 may use any parallelism scheme, efficiency scheme, and compression scheme, such as TFRecord, compression encoding (e.g., PNG), sharding, parallel detection of map transformations, batching, prefetching, model parallelism, data parallelism, and synchronous / asynchronous stochastic gradient descent (SGD). The chromatin model 802 may include upsampling layers, downsampling layers, recurrent connections, gates and gated memory units (such as LSTM or GRU), residual blocks, residual connections, high-speed connections, skip connections, peephole connections, activation functions (e.g., non-linear transformation functions such as rectified linear units (ReLU), leaky ReLU, exponential linear units (ELU), sigmoid, and hyperbolic tangent (tanh)), batch normalization layers, regularization layers, dropout layers, pooling layers (e.g., max or average pooling), global average pooling layers, and attention mechanisms.

[0148] In some implementations, the chromatin model 802 can be a linear regression model, a logistic regression model, an elastic net model, a support vector machine (SVM), a random forest (RF), a decision tree or a boosted decision tree (e.g., XGBoost), or some other tree-based logic (e.g., a metric tree, a kd tree, an R tree, a general B tree, an X tree, a ball tree, a locality-sensitive hash, and an inverted index). In some implementations, the chromatin model 802 can be an ensemble of multiple models.

[0149] In some implementations, a back-propagation based gradient update technique can be used to train the chromatin model 802. Example gradient descent techniques that can be used to train these models include stochastic gradient descent, batch gradient descent, and mini-batch gradient descent. Some examples of gradient descent optimization algorithms that can be used to train these models are Momentum, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, Adam, AdaMax, Nadam, and AMSGrad.

[0150] Input base sequence

[0151] Fig.13An input generation logic component 1304 is shown that accesses a sequence database 1302 (e.g., GTEx, RNA-Seq, WGS) and generates an input base sequence 1314. The input base sequence 1314 includes a target base sequence 1324. The target base sequence 1324 is flanked by a right end base sequence 1322 with downstream background bases, and a left end base sequence 1326 with upstream background bases.

[0152] Base-resolution output sequence

[0153] Fig.14 Depicted is one implementation of base-resolution evolutionary conservation prediction 1400 by chromatin model 802. Chromatin model 802 processes input base sequence 1314 and generates alternative representations 1406 of input base sequence 1314. In one implementation, when input base sequence 1314 is processed by a cascade of convolutional layers of chromatin model 802, alternative representations 1406 are convolutional representations of input base sequence 1314.

[0154] The chromatin output generation logic component 1408 processes the alternative representations 1406 of the input base sequence 1314 and generates an output sequence 1410 of corresponding per-base chromatin outputs for corresponding target bases in the target base sequence 1324 .

[0155] Fig.15 An example of an output sequence 1410 corresponding to a target base sequence 1324 is shown. For a given target base at a given position in the target base sequence 1324, a per-base chromatin output given in the output sequence 1410 specifies a measure of evolutionary conservation of the given target base across multiple species. The multiple species may include homologous species.

[0156] 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 were deleterious in the past) and that mutations at these sites increase the likelihood of being pathogenic (causing disease) in humans. Homologous sequences of the target protein are collected and aligned in a multiple sequence alignment (MSA). Conservation metrics are calculated based on weighted scores and / or frequencies of different amino acids observed at the target position in the MSA.

[0157] MSA is usually an alignment of three or more biological sequences, proteins or nucleic acids of similar length. From this alignment, the degree of homology can be inferred and the evolutionary relationships between sequences can be studied. MSA is also a tool for identifying evolutionary relationships and common patterns between genes. Computational algorithms are used to generate the alignment and analyze it. Dynamic and heuristic methods are used in most MSA algorithms. One of the purposes of alignment is to detect structural or functional identity and similarity between residues in a protein sequence relative to residues in other protein sequences.

[0158] The homology information about the aligned sequences in MSA can be represented by the following two matrices (evolutionary conservation metrics): position-specific score matrix (PSSM) and position-specific frequency matrix (PSFM). PSSM and PSFM reflect the conservation of residues at specific positions of protein chains based on evolutionary information.

[0159] In some implementations, a given per-base chromatin output further specifies a transcription start measure for a given target base at a given position.

[0160] In one embodiment, the measure of evolutionary conservation is a phylogenetic P-value (phyloP) score that specifies the deviation from a neural substitution null model so that a decrease in the substitution rate of a given target base at a given position is detected as conservation and an increase in the substitution rate of a given target base at a given position is detected as acceleration.

[0161] In one implementation, the measure of evolutionary conservation is a phastCons score, which specifies the posterior probability that the given target base at the given position has a conserved state or a non-conserved state.

[0162] In one implementation, the measure of evolutionary conservation is a Genome Evolutionary Rate Analysis (GERP) score, which specifies the decrease in the number of substitutions for a given target base at a given position across multiple species.

[0163] In one implementation, the measure of transcription initiation is a Cap Analysis Gene Expression (CAGE) score, which specifies the frequency of transcription initiation for a given target base at a given position.

[0164] In some implementations, a given per-base chromatin output further specifies the confounding factor signal level of a given target base at a given position. In one implementation, the confounding factor signal level specifies a DNA enzyme I-hypersensitive site (DHS). In one implementation, the confounding factor signal level specifies transposase accessible chromatin sequencing analysis (ATAC-Seq). In another implementation, the confounding factor signal level specifies transcription factor (TF) binding. In yet another implementation, the confounding factor signal level specifies histone modification (HM) marks. In yet another implementation, the confounding factor signal level specifies DNA methylation marks.

[0165] Base-resolution gene expression level prediction

[0166] Gene expression model

[0167] Fig.16 A specific implementation of the gene expression model 1600 disclosed in the present invention is shown. In general, the gene expression model 1602 can be a rule-based model, a tree-based model, or a machine learning model. Examples include multi-layer perceptron (MLP), feedforward neural network, fully connected neural network, fully convolutional neural network, ResNet, sequence to sequence (Seq2Seq) model (such as WaveNet), semantic segmentation neural network and generative adversarial network (GAN) (e.g., CycleGAN, StyleGAN, pixelRNN, text-2-image, DiscoGAN, IsGAN).

[0168] In some implementations, the gene expression model 1602 can include a self-attention mechanism such as a Transformer, a Visual Transformer (ViT), a Bidirectional Transformer (BERT), a Detection Transformer (DETR), a Deformable DETR, UP-DETR, DeiT, Swin, GPT, iGPT, GPT-2, GPT-3, BERT, SpanBERT, RoBERTa, XLNet, ELECTRA, UniLM, BART, T5, ERNIE (THU), KnowBERT, DeiT-Ti, DeiT-S, DeiT-B, T2T-ViT-14, T2T-ViT-15, 9. T2T-ViT-24, PVT-Small, PVT-Medium, PVT-Large, TNT-S, TNT-B, CPVT-S, CPVT-S-GAP, CPVT-B, Swin-T, Swin-S, Swin-B, Twins-SVT-S, Twins-SVT-B, Twins-SVT-L, Shuffle-T, Shuffle-S, Shuffle-B, XCiT-S12 / 16, CMT-S, CMT-B, VOLO-D1, VOLO-D2, VOLO-D3, VOLO-D4, MoCo v3, ACT, TSP, Max-DeepLab, VisTR, SETR, Manual Transformer, HOT-Net, METRO, Image Transformer, Tamed Transformer, TransGAN, IPT, TTSR, STTN, Masked Transformer, CLIP, DALL-E, Cogview, UniT, ASH, TinyBert, FullyQT, ConvBert, FCOS, Faster R-CNN+FPN, DETR-DC5, TSP-FCOS, TSP-RCNN, ACT+MKDD(L=32), ACT+MKDD(L=16), SMCA, efficient DETR, UP-DETR, UP-DETR, ViTB / 16-FRCNN, ViT -B / 16-FRCNN, PVT-Small+RetinaNet, Swin-T+RetinaNet, Swin-T+ATSS, PVT-Small+DETR, TNT-S+DETR, YOLOS-Ti, YOLOS-S and YOLOS-B.

[0169] In some embodiments, examples of gene expression model 1602 include: a convolutional neural network (CNN) with multiple convolutional layers; a recurrent neural network (RNN), such as a long short-term memory network (LSTM), a bidirectional LSTM (Bi-LSTM), or a gated recurrent unit); and a combination of both CNN and RNN.

[0170] In some implementations, the gene expression model 1602 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 gene expression model 1602 may use one or more loss functions, such as logistic regression / logarithmic loss function, multi-class cross entropy / softmax loss function, binary cross entropy loss function, mean square error loss function, L1 loss function, L2 loss function, smooth L1 loss function, and Huber loss function. The gene expression model 1602 may use any parallelism scheme, efficiency scheme, and compression scheme, such as TFRecord, compression encoding (e.g., PNG), sharding, parallel detection of map conversion, batching, prefetching, model parallelism, data parallelism, and synchronous / asynchronous stochastic gradient descent (SGD). The gene expression model 1602 may include upsampling layers, downsampling layers, recurrent connections, gates and gated memory units (such as LSTM or GRU), residual blocks, residual connections, high-speed connections, skip connections, peephole connections, activation functions (e.g., non-linear transformation functions such as rectified linear units (ReLU), leaky ReLU, exponential linear units (ELU), sigmoid, and hyperbolic tangent (tanh)), batch normalization layers, regularization layers, dropout layers, pooling layers (e.g., max or average pooling), global average pooling layers, and attention mechanisms.

[0171] In some implementations, the gene expression model 1602 may be a linear regression model, a logistic regression model, an elastic net model, a support vector machine (SVM), a random forest (RF), a decision tree or an enhanced decision tree (e.g., XGBoost), or some other tree-based logic (e.g., a metric tree, a kd tree, an R tree, a general B tree, an X tree, a ball tree, a locality-sensitive hash, and an inverted index). In some implementations, the gene expression model 1602 may be an ensemble of multiple models.

[0172] In some implementations, the gene expression model 1602 can be trained using a back-propagation based gradient update technique. Example gradient descent techniques that can be used to train these models include stochastic gradient descent, batch gradient descent, and mini-batch gradient descent. Some examples of gradient descent optimization algorithms that can be used to train these models are Momentum, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, Adam, AdaMax, Nadam, and AMS Grad.

[0173] The gene expression model 1602 processes the output sequence 1410 and generates an alternative representation 1612 of the output sequence 1410. In one specific implementation, the alternative representation 1612 is a convolutional representation of the output sequence 1410 when the output sequence 1410 is processed by the convolutional layers of the gene expression model 1602 in cascade.

[0174] The gene expression model output generation logic component 1622 processes the alternative representations 1612 of the output sequence 1410 and generates a gene expression output sequence 1632 of corresponding per-base gene expression outputs for corresponding target bases in the target base sequence 1324 .

[0175] For a given target base at a given position in the gene expression output sequence 1632, a given per base gene expression output specifies a measure of the gene expression level of a given target base at a given position. In one implementation, the gene expression level is measured per base metric (such as CAGE transcription start site (CTSS)). In another implementation, the gene expression level is measured per gene metric (such as transcripts per million (TPM) or reads per kilobase of transcripts (RPKM)). In yet another implementation, the gene expression level is measured per gene metric (such as fragments per kilobase per million (FPKM)).

[0176] Variant pathogenicity classification

[0177] Fig.17 One example of a reference sequence 1702 and an alternative sequence 1722 (or alternative sequence) is shown. The alternative sequence 1712 differs from the reference sequence 1702 by a variant nucleotide 1722.

[0178] Fig.18 A specific implementation of the variant classification logic component 1800 disclosed in the present invention is shown. The variant classification logic component 1800 is further configured to include a reference input generation logic component 1802, which accesses the sequence database 1302 and generates a reference base sequence 1702. The reference base sequence 1702 includes a reference target base sequence. The reference target base sequence includes a reference base at the position under analysis. The reference base is flanked by a right-end base sequence with downstream background bases, and a left-end base sequence with upstream background bases.

[0179] The variant classification logic 1800 is further configured to include an alternative input generation logic 1812 that accesses the sequence database 1302 and generates an alternative base sequence 1712. The alternative base sequence 1712 includes an alternative target base sequence. The alternative target base sequence includes an alternative base 1722 at the position under analysis. The alternative base 1722 is flanked by a right end base sequence with downstream background bases, and a left end base sequence with upstream background bases.

[0180] The variant classification logic component 1800 is further configured to include a reference processing logic component 1822, which causes the chromatin model 802 to process the reference base sequence 1702 and generate an alternative representation 1832 of the reference base sequence 1702, and also causes the chromatin output generation logic component 1408 to process the alternative representation 1832 of the reference base sequence 1702 and generate a reference output sequence 1842 of a corresponding per-base reference chromatin output for the corresponding reference target base in the reference target base sequence.

[0181] For a given reference target base at a given position in a reference target base sequence, the per-base reference chromatin output given in reference output sequence 1842 specifies a measure of evolutionary conservation of the given reference target base across multiple species.

[0182] The variant classification logic component 1800 is further configured to include an alternative processing logic component 1852, which causes the chromatin model 802 to process the alternative base sequence 1712 and generate an alternative representation 1862 of the alternative base sequence 1712, and also causes the chromatin output generation logic component 1408 to process the alternative representation 1862 of the alternative base sequence 1712 and generate an alternative output sequence 1872 of a corresponding per-base alternative chromatin output of the corresponding alternative target base in the alternative target base sequence.

[0183] In alternative output sequence 1872, for a given alternative target base at a given position in the alternative target base sequence, a given per-base alternative chromatin output specifies a measure of the evolutionary conservation of the given alternative target base across multiple species.

[0184] Fig.19 A specific implementation of the pathogenicity prediction logic component 1900 disclosed in the present invention is shown. The variant classification logic component 1800 is further configured to include a pathogenicity prediction logic component 1900, which compares the reference output sequence 1842 and the alternative output sequence 1872 by position, and then generates an incremental sequence 1912 having position-by-position sequence differences between the positions in the reference output sequence 1842 and the alternative output sequence 1872.

[0185] The pathogenicity prediction logic 1900 is further configured to generate a pathogenicity prediction 1922 for the alternative base 1722 based on the incremental sequence 1912. In one implementation, the pathogenicity prediction logic 1900 is further configured to accumulate the sequence difference values ​​by position into a cumulative sequence value, and then generate a pathogenicity prediction 1922 for the alternative base 1722 based on the cumulative sequence value. In some implementations, the cumulative sequence value is the average or maximum value of the sequence difference values ​​by position. In other implementations, the cumulative sequence value is the sum of the sequence difference values ​​by position.

[0186] In some implementations, pathogenicity prediction logic 1900 is further configured to compare corresponding portions of reference output sequence 1842 and alternative output sequence 1872 by position, and then generate incremental subsequences having position-wise subsequence differences of positions in the corresponding portions.

[0187] In one implementation, these corresponding portions span the right flanking positions and the left flanking positions around the position under analysis. In some implementations, the pathogenicity prediction logic component 1900 is further configured to generate a pathogenicity prediction for the alternative base 1722 based on the incremental subsequence. In one implementation, the pathogenicity prediction can be a score between 0 and 1, where 0 represents absolute benign and 1 represents absolute pathogenicity. In other implementations, a cutoff value can be used, such as a pathogenicity score above 5 can be considered pathogenic and below 5 can be considered benign.

[0188] In some implementations, the pathogenicity prediction logic 1900 is further configured to accumulate the subsequence differences by position into a cumulative subsequence value, and then generate a pathogenicity prediction for the alternative base 1722 based on the cumulative subsequence value. In one implementation, the cumulative subsequence value is the average of the subsequence differences by position. In another implementation, the cumulative subsequence value is the sum or maximum of the subsequence differences by position.

[0189] Computer Systems

[0190] Fig. 202000 is an example computer system that can be used to implement various aspects of the technology disclosed in the present invention. The computer system 2000 includes at least one central processing unit (CPU) 2024 that communicates with multiple peripheral devices via a bus subsystem 2022. These peripheral devices may include a storage subsystem 2010 (including, for example, a memory device and a file storage subsystem 2018), a user interface input device 2020, a user interface output device 2028, and a network interface subsystem 2026. The input devices and output devices allow a user to interact with the computer system 2000. The network interface subsystem 2026 provides an interface to an external network, including an interface to corresponding interface devices in other computer systems.

[0191] In one specific implementation, the chromatin model 802 is communicatively linked to the storage subsystem 2010 and the user interface input device 2020 .

[0192] User interface input devices 2020 may include: keyboards; pointing devices, such as mice, trackballs, touch pads, or graphics tablets; scanners; touch screens incorporated into displays; audio input devices, such as voice recognition systems and microphones; and other types of input devices. In general, the use of the term "input device" is intended to include all possible types of devices and ways of inputting information into the computer system 2000.

[0193] The user interface output device 2028 may include a display subsystem, a printer, a fax machine, or a non-visual display (such as an audio output device). The display subsystem may include an LED display, a cathode ray tube (CRT), a flat panel device such as a liquid crystal display (LCD), a projection device, or some other mechanism for producing a visible image. The display subsystem may also provide a non-visual display, such as an audio output device. Generally speaking, the use of the term "output device" is intended to include all possible types of devices and methods for outputting information from the computer system 2000 to a user or to another machine or computer system.

[0194] The storage subsystem 2010 stores programming structures 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 2030.

[0195] Processor 2030 may 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 2030 may be provided by a deep learning cloud platform (such as Google Cloud Platform TM 、Xilinx TM and Cirrascale TM) is hosted. Examples of processors 2078 include Google's TensorProcessing Unit (TPU) TM , rack-mount solutions (such as GX4 Rackmount Series TM 、GX20RackmountSeries TM ), NVIDIA DGX-1 TM 、Microsoft's Stratix V FPGA TM , Graphcore's IntelligentProcessor 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 V100s TM Lambda GPU servers, and more.

[0196] The memory subsystem 2012 used in the storage subsystem 2010 may include multiple memories, including a main random access memory (RAM) 2014 for storing instructions and data during program execution and a read-only memory (ROM) 2016 in which fixed instructions are stored. The file storage subsystem 2018 may provide persistent storage for program files and data files, and may include a hard drive, a floppy disk drive, and associated removable media, a CD-ROM drive, an optical drive, or a removable media disk cartridge. Modules that implement certain specific implementation functions may be stored in the storage subsystem 2010 by the file storage subsystem 2018, or in other machines accessible to the processor.

[0197] The bus subsystem 2022 provides a mechanism for the various components and subsystems of the computer system 2000 to communicate with each other as intended. Although the bus subsystem 2022 is shown schematically as a single bus, alternative implementations of the bus subsystem may use multiple busses.

[0198] Computer system 2000 itself may be of various types, including a personal computer, a portable computer, a workstation, a computer terminal, a network computer, a television, a mainframe, a server farm, a widely distributed group of loosely networked computers, or any other data processing system or user device. Fig. 20 The description of the computer system 2000 depicted in the figure is intended only as a specific example for demonstrating a preferred embodiment of the present invention. The computer system 2000 may also have many other configurations, where the components are different than those of the computer system 2000. Fig. 20 There may be more or fewer computer systems depicted.

[0199] Terms

[0200] The technology disclosed in the present invention can be practiced as a system, method or product. One or more features of a specific implementation can be combined with a basic specific implementation. Specific implementations that are not mutually exclusive are taught to be combinable. One or more features of a specific implementation can be combined with other specific implementations. The present disclosure periodically reminds users of these options. The omission of the expression of repeating these options from some specific implementations should not be regarded as limiting the combination taught in the preceding section, and these expressions will be incorporated into each specific implementation in the following specific implementations by reference.

[0201] One or more specific implementations and clauses of the disclosed technology, or elements thereof, may be implemented in the form of a computer product including a non-transitory computer-readable storage medium having a computer-usable program code for performing the indicated method steps. In addition, one or more specific implementations and clauses of the disclosed technology, or elements thereof, may be implemented in the form of a device including a memory and at least one processor coupled to the memory, and operated to perform the exemplary method steps. Further, in another aspect, one or more specific implementations and clauses of the disclosed technology, or elements thereof, may be implemented in the form of a device for performing one or more of the method steps described herein; the device may include (i) a hardware module, (ii) a software module executed on one or more hardware processors, or (iii) a combination of hardware and software modules; any of (i) to (iii) implements the specific technology set forth herein, and the software module is stored in a computer-readable storage medium (or multiple such media).

[0202] The clauses described in this section may be combined into features. For the sake of brevity, combinations of features are not listed separately and are not repeated with each basic feature group. The reader will understand how the features identified in the clauses described in this section can be easily combined with basic feature sets identified as specific implementations in other parts of this application. These clauses are not meant 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 their equivalents.

[0203] Other specific implementations of the clauses described in this section may include a non-transitory computer-readable storage medium storing instructions that are executable by a processor to perform any of the clauses described in this section. Yet another specific implementation of the clauses described in this section may include a system that includes a memory and one or more processors that are operable to execute instructions stored in the memory to perform any of the clauses described in this section.

[0204] The present invention discloses the following clauses:

[0205] Article Set 1

[0206] 1. A computer-implemented method for identifying rare variants that cause extreme gene expression levels, comprising:

[0207] Obtain gene expression levels for a group of individuals;

[0208] normalizing the gene expression levels, and identifying those outlier individuals from the group of individuals having extreme gene expression levels, wherein the extreme gene expression levels are determined from tail quantiles of the normalized gene expression levels;

[0209] selecting rare variants from a gene sequence of the outlier individual, wherein the rare variants are selected based on an allele frequency cutoff;

[0210] fitting a causal relationship model to determine the causal relationship between the rare variant in the outlier individual and the extreme gene expression level while controlling for multiple confounding factors; and

[0211] A causality score for the rare variant is generated based on the determined causality, wherein a specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant.

[0212] 2. A computer-implemented method according to clause 1, wherein the causality score is a probability value (p-value).

[0213] 3. A computer-implemented method according to clause 1, wherein the p-value is determined by a Pearson correlation coefficient.

[0214] 4. A computer-implemented method according to clause 1, wherein the causal relationship model is a logistic regression model, a linear regression model, an analysis of covariance (ANCOVA) model and / or a multivariate analysis of covariance (MANCOVA) model.

[0215] 5. A computer-implemented method according to claim 4, wherein the fitted causal relationship model determines the causal relationship by predicting the specific gene expression level of a specific gene in a specific chromosome based on the gene expression level driven by the variant caused by the specific rare variant.

[0216] 6. A computer-implemented method according to clause 5, wherein the fitted causal model measures the contribution of the variant-driven gene expression level as a variant effect covariate.

[0217] 7. A computer-implemented method according to clause 1, wherein the plurality of confounding factors comprises distal trans-expression quantitative trait loci (eQTL) effects.

[0218] 8. A computer-implemented method according to clause 7, wherein the fitted causal model controls the distal trans-eQTL effect by predicting the specific gene expression level based on the trans gene expression levels caused by other genes in other chromosomes.

[0219] 9. A computer-implemented method according to clause 8, wherein the fitted causal model measures the contribution of the trans gene expression level as a trans effect covariate.

[0220] 10. The computer-implemented method of clause 1, wherein the plurality of confounding factors comprises local cis-eQTL effects.

[0221] 11. A computer-implemented method according to clause 10, wherein the fitted causal model controls the local cis-eQTL effect by predicting the specific gene expression level based on the cis-gene expression level caused by the presence of multiple common variants in the neighborhood of the specific gene.

[0222] 12. A computer-implemented method according to clause 11, wherein the neighborhood is defined by an offset relative to a transcription start site (TSS) in the specific gene.

[0223] 13. A computer-implemented method according to clause 11, wherein the fitted causal model measures the contribution of the cis gene expression level as a cis-effect covariate.

[0224] 14. A computer-implemented method according to clause 1, wherein the multiple confounding factors include population structure and ancestral effects.

[0225] 15. A computer-implemented method according to clause 14, wherein the population structure and ancestry effects are represented by one or more genotype-based principal components (gPCs).

[0226] 16. A computer-implemented method according to clause 15, wherein the fitted causal model controls for the population structure and ancestral effects by predicting the specific gene expression level based on the gPC gene expression level caused by the gPC.

[0227] 17. A computer-implemented method according to clause 16, wherein the fitted causal model measures the contribution of the gPC gene expression levels as covariates of population structure and ancestral effects.

[0228] 18. The computer-implemented method of clause 1, wherein the plurality of confounding factors comprises a probability estimate of expression residual (PEER) effect.

[0229] 19. The computer-implemented method of clause 18, wherein the fitted causal model controls the PEER effect by predicting the specific gene expression level based on the PEER gene expression level caused by the PEER.

[0230] 20. The computer-implemented method of clause 19, wherein the fitted causal model measures the contribution of the PEER gene expression level as a PEER effect covariate.

[0231] 21. The computer-implemented method of clause 1, wherein the plurality of confounding factors includes environmental effects.

[0232] 22. A computer-implemented method according to clause 21, wherein the fitted causal model controls for the environmental effect by predicting the specific gene expression level based on the environmental gene expression level caused by the environmental effect.

[0233] 23. A computer-implemented method according to clause 22, wherein the fitted causal model measures the contribution of the environmental gene expression level as an environmental effect covariate.

[0234] 23. A computer-implemented method according to clause 1, wherein the multiple confounding factors include gender effect, batch effect, genotyping platform effect and library construction protocol effect.

[0235] 24. A computer-implemented method according to clause 1, wherein the extreme gene expression levels include gene overexpression and gene underexpression.

[0236] 25. The computer-implemented method of clause 24, further comprising determining the causal relationship between the rare variant and the overexpression of the gene while controlling for the plurality of confounding factors.

[0237] 26. The computer-implemented method according to clause 25 further includes generating an excess causality score for the rare variant, wherein the specific excess causality score for the specific rare variant indicates the likelihood that the specific rare variant causes overexpression of a gene in those outlier individuals whose gene sequences contain the specific rare variant.

[0238] 27. A computer-implemented method according to clause 26, wherein the excess causality score is an excess probability value (excess p-value).

[0239] 28. A computer-implemented method according to clause 27, wherein the excess p-value is determined by a Pearson correlation coefficient.

[0240] 29. A computer-implemented method according to clause 27, wherein the excess p-value specifies a statistically unconfounded likelihood that the rare variant increases gene expression in a gene that otherwise has lower gene expression relative to other genes in the gene set.

[0241] 30. The computer-implemented method of clause 24, further comprising determining the causal relationship between the rare variant and the underexpression of the gene while controlling for the plurality of confounding factors.

[0242] 31. The computer-implemented method according to clause 30 further includes generating a deficiency causality score for the rare variant, wherein the specific deficiency causality score for the specific rare variant indicates the likelihood that the specific rare variant causes insufficient gene expression in those outlier individuals whose gene sequences contain the specific rare variant.

[0243] 32. The computer-implemented method of clause 31, wherein the deficiency causality score is a deficiency probability value (deficiency p-value).

[0244] 33. A computer-implemented method according to clause 32, wherein the deficiency p-value is determined by a Pearson correlation coefficient.

[0245] 34. A computer-implemented method according to clause 32, wherein the insufficient p-value specifies a statistically unconfounded likelihood that the rare variant reduces gene expression in a gene that otherwise has higher gene expression relative to other genes in the gene set.

[0246] 35. A computer-implemented method according to clause 1, wherein the rare variant is a non-coding variant.

[0247] 36. A computer-implemented method according to clause 17, wherein the non-coding variants include 5' untranslated region (UTR) variants, 3' UTR variants, enhancer variants and promoter variants.

[0248] 37. A computer-implemented method according to clause 1, wherein the gene expression levels are further stratified into tissue-specific gene expression levels of multiple tissues.

[0249] 38. A computer-implemented method according to clause 37, wherein quantile normalization is used to normalize the gene expression level of each gene in each tissue.

[0250] 39. The computer-implemented method of clause 38, wherein the causal model is fit separately for each tissue.

[0251] 40. A computer-implemented method as described in clause 1, wherein stratification is used to fit the causal model.

[0252] 41. The computer-implemented method of clause 1 , further comprising generating a ranking of the rare variants based on the causality score.

[0253] 42. The computer-implemented method of clause 26, further comprising generating a ranking of the rare variants based on the excess causality score.

[0254] 43. The computer-implemented method of clause 31 , further comprising generating a ranking of the rare variants based on the insufficient causality score.

[0255] 44. A computer-implemented method as described in clause 1, wherein the rare variant is a single-occurring variant.

[0256] 45. A computer-implemented method according to clause 44, wherein the single-occurring variant occurs in only one outlier individual from among the outlier individuals.

[0257] 46. ​​A system comprising one or more processors coupled to a memory loaded with computer instructions to identify rare variants that cause extreme gene expression levels, the instructions when executed on the processors performing actions comprising:

[0258] Obtain gene expression levels for a group of individuals;

[0259] normalizing the gene expression levels, and identifying those outlier individuals from the group of individuals having extreme gene expression levels, wherein the extreme gene expression levels are determined from tail quantiles of the normalized gene expression levels;

[0260] selecting rare variants from a gene sequence of the outlier individual, wherein the rare variants are selected based on an allele frequency cutoff;

[0261] fitting a causal relationship model to determine the causal relationship between the rare variant in the outlier individual and the extreme gene expression level while controlling for multiple confounding factors; and

[0262] A causality score for the rare variant is generated based on the determined causality, wherein a specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant.

[0263] 47. The system according to clause 46, further implementing the actions including executing clauses 2 to 45.

[0264] 48. A non-transitory computer readable storage medium having printed thereon computer program instructions for identifying rare variants that cause extreme gene expression levels, said instructions when executed on a processor implementing a method comprising:

[0265] Obtain gene expression levels for a group of individuals;

[0266] normalizing the gene expression levels, and identifying those outlier individuals from the group of individuals having extreme gene expression levels, wherein the extreme gene expression levels are determined from tail quantiles of the normalized gene expression levels;

[0267] selecting rare variants from a gene sequence of the outlier individual, wherein the rare variants are selected based on an allele frequency cutoff;

[0268] fitting a causal relationship model to determine the causal relationship between the rare variant in the outlier individual and the extreme gene expression level while controlling for multiple confounding factors; and

[0269] A causality score for the rare variant is generated based on the determined causality, wherein a specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant.

[0270] 49. The non-transitory computer-readable storage medium of clause 46, wherein implementing the method further comprises executing clauses 2 to 45.

[0271] Article Set 2

[0272] 1. An artificial intelligence-based system for detecting changes in gene expression at per-base resolution, comprising:

[0273] an input generation logic component that accesses a sequence database and generates an input base sequence, wherein the input base sequence includes a target base sequence, and wherein the target base sequence is flanked by a right-end base sequence having downstream background bases, and a left-end base sequence having upstream background bases;

[0274] a chromatin model that processes the input base sequence and generates an alternative representation of the input base sequence; and

[0275] a chromatin output generation logic component that processes the alternative representations of the input base sequence and generates an output sequence of corresponding per-base chromatin outputs for corresponding target bases in the target base sequence,

[0276] Wherein in the output sequence, for a given target base at a given position in the target base sequence, a given per-base chromatin output specifies a measure of evolutionary conservation of the given target base across multiple species.

[0277] 2. An artificial intelligence-based system according to claim 1, wherein the given per-base chromatin output further specifies a transcription start measure for the given target base at the given position.

[0278] 3. An artificial intelligence-based system according to claim 1, wherein the measure of evolutionary conservation is a phylogenetic P-value (phyloP) score that specifies a deviation from a neural substitution null model so that a decrease in the substitution rate of the given target base at the given position is detected as conservation and an increase in the substitution rate of the given target base at the given position is detected as acceleration.

[0279] 4. An artificial intelligence-based system according to claim 1, wherein the measure of evolutionary conservation is a phastCons score that specifies the posterior probability that the given target base at the given position has a conserved state or a non-conserved state.

[0280] 5. An artificial intelligence-based system according to claim 1, wherein the measure of evolutionary conservation is a genome evolutionary rate analysis (GERP) score that specifies a reduction in the number of substitutions for the given target base at the given position across multiple species.

[0281] 6. An artificial intelligence-based system according to claim 2, wherein the measure of transcription initiation is a Cap Analysis Gene Expression (CAGE) score that specifies the frequency of transcription initiation of the given target base at the given position.

[0282] 7. An artificial intelligence-based system according to claim 1, wherein the given per-base chromatin output further specifies the confounding factor signal level of the given target base at the given position.

[0283] 8. An artificial intelligence-based system according to claim 7, wherein the confounding factor signal level specifies a DNase I-hypersensitive site (DHS).

[0284] 9. An artificial intelligence-based system according to claim 7, wherein the confounding factor signal level specifies transcription factor (TF) binding.

[0285] 10. An artificial intelligence-based system according to claim 7, wherein the confounding factor signal level specifies a histone modification (HM) mark.

[0286] 11. An artificial intelligence-based system according to claim 7, wherein the confounding factor signal level specifies a DNA methylation marker.

[0287] 12. The artificial intelligence-based system according to clause 1, further configured to include:

[0288] a gene expression model that processes the output sequence and generates an alternative representation of the output sequence; and

[0289] a gene expression model output generation logic component that processes the alternative representations of the output sequence and generates a gene expression output sequence of corresponding per-base gene expression outputs for the corresponding target bases in the target base sequence,

[0290] Wherein for the given target base at the given position in the gene expression output sequence, the given per-base gene expression output specifies a measure of the gene expression level of the given target base at the given position.

[0291] 13. An artificial intelligence-based system according to clause 12, wherein the gene expression level is measured as a per-base metric such as CAGE transcription start site (CTSS).

[0292] 14. An artificial intelligence-based system according to clause 12, wherein the gene expression level is measured in a per-gene metric such as transcripts per million (TPM) or reads per kilobase of transcript (RPKM).

[0293] 15. An artificial intelligence-based system according to clause 12, wherein the gene expression level is measured in a per-gene metric such as fragments per kilobase million (FPKM).

[0294] 16. The artificial intelligence based system of clause 1, further configured to include a variant classification logic component.

[0295] 17. An artificial intelligence-based system according to claim 16, wherein the variant classification logic component is further configured to include a reference input generation logic component, which accesses the sequence database and generates a reference base sequence, wherein the reference base sequence includes a reference target base sequence, wherein the reference target base sequence includes a reference base at a position under analysis, and wherein the reference base is flanked by a right-end base sequence having downstream background bases, and a left-end base sequence having upstream background bases.

[0296] 18. An artificial intelligence-based system according to claim 17, wherein the variant classification logic component is further configured to include an alternative input generation logic component, which accesses the sequence database and generates an alternative base sequence, wherein the alternative base sequence includes an alternative target base sequence, wherein the alternative target base sequence includes an alternative base at a position in the analysis, and wherein the alternative base is flanked by the right-end base sequence having the downstream background base, and the left-end base sequence having the upstream background base.

[0297] 19. An artificial intelligence-based system according to claim 17, wherein the variant classification logic component is further configured to include a reference processing logic component, which enables the chromatin model to process the reference base sequence and generate an alternative representation of the reference base sequence, and also enables the chromatin output generation logic component to process the alternative representation of the reference base sequence and generate a reference output sequence of corresponding per-base reference chromatin outputs for the corresponding reference target bases in the reference target base sequence.

[0298] wherein in the reference output sequence, for a given reference target base at a given position in the reference target base sequence, a given per-base reference chromatin output specifies a measure of evolutionary conservation of the given reference target base across the plurality of species.

[0299] 20. An artificial intelligence-based system according to clause 18, wherein the variant classification logic component is further configured to include an alternative processing logic component, which causes the chromatin model to process the alternative base sequence and generate an alternative representation of the alternative base sequence, and also causes the chromatin output generation logic component to process the alternative representation of the alternative base sequence and generate an alternative output sequence of corresponding per-base alternative chromatin outputs for corresponding alternative target bases in the alternative target base sequence,

[0300] Wherein in the alternative output sequence, for a given alternative target base at a given position in the alternative target base sequence, a given per-base alternative chromatin output specifies a measure of evolutionary conservation of the given alternative target base across the plurality of species.

[0301] 21. An artificial intelligence-based system according to claim 20, wherein the variant classification logic component is further configured to include a pathogenicity prediction logic component, which compares the reference output sequence and the alternative output sequence by position and then generates an incremental sequence having position-by-position sequence differences between the positions in the reference output sequence and the alternative output sequence.

[0302] 22. An artificial intelligence-based system according to clause 21, wherein the pathogenicity prediction logic component is further configured to generate a pathogenicity prediction for the alternative base based on the incremental sequence.

[0303] 23. An artificial intelligence-based system according to clause 21, wherein the pathogenicity prediction logic component is further configured to accumulate the position-based sequence differences into cumulative sequence values, and then generate the pathogenicity prediction of the alternative base based on the cumulative sequence values.

[0304] 24. An artificial intelligence-based system according to clause 23, wherein the accumulated sequence value is an average or maximum value of the position-by-position sequence differences.

[0305] 25. An artificial intelligence-based system according to clause 23, wherein the accumulated sequence value is the sum of the position-by-position sequence differences.

[0306] 26. An artificial intelligence-based system according to claim 21, wherein the pathogenicity prediction logic component is further configured to compare the reference output sequence and the corresponding portions of the alternative output sequence by position, and then generate incremental subsequences having position-by-position subsequence differences of the positions in the corresponding portions.

[0307] 27. An artificial intelligence-based system according to clause 26, wherein the corresponding portion spans a right-end flanking position and a left-end flanking position around the position under analysis.

[0308] 28. The artificial intelligence-based system of clause 26, wherein the pathogenicity prediction logic component is further configured to generate a pathogenicity prediction for the alternative base based on the incremental subsequence.

[0309] 29. An artificial intelligence-based system according to clause 26, wherein the pathogenicity prediction logic component is further configured to accumulate the subsequence differences by position into cumulative subsequence values, and then generate the pathogenicity prediction of the alternative base based on the cumulative subsequence values.

[0310] 30. An artificial intelligence-based system according to clause 29, wherein the accumulated subsequence value is an average of the position-wise subsequence differences.

[0311] 31. An artificial intelligence-based system according to clause 29, wherein the accumulated subsequence value is the sum or maximum of the position-wise subsequence differences.

[0312] 32. An artificial intelligence-based system according to clause 1, wherein the target base sequence is a coding region of a gene.

[0313] 33. An artificial intelligence-based system according to clause 1, wherein the target base sequence is a non-coding region of a gene.

[0314] 34. An artificial intelligence-based system according to clause 33, wherein the non-coding region spans the transcription start site, 5' untranslated region (UTR), 3' UTR, enhancer and promoter.

[0315] 35. An artificial intelligence-based system according to clause 18, wherein the alternative base is a rare variant outlier individual among an outlier individual cohort,

[0316] The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.

[0317] 36. An artificial intelligence-based system according to clause 35, wherein the extreme gene expression levels are determined from the tail quantiles of the normalized gene expression levels.

[0318] 37. An artificial intelligence-based system according to clause 36, wherein the extreme gene expression levels include gene overexpression and gene underexpression.

[0319] 38. An artificial intelligence-based system according to clause 35, wherein the rare variant is a coding variant.

[0320] 39. An artificial intelligence-based system according to clause 35, wherein the rare variant is a non-coding variant.

[0321] 40. An artificial intelligence-based system according to clause 39, wherein the non-coding variant is a promoter variant.

[0322] 41. An artificial intelligence-based system according to clause 39, wherein the non-coding variant is an enhancer variant.

[0323] 42. An artificial intelligence-based system according to claim 1, wherein the chromatin model has a first set of weights and wherein the chromatin output generation logic component has a second set of weights.

[0324] 43. The artificial intelligence-based system of clause 42, wherein during training, the first set of weights of the chromatin model is trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence,

[0325] The second set of weights of the chromatin output generation logic component are trained end-to-end from scratch together with the first set of weights of the chromatin model to process the alternative representation of the input base sequence and generate the output sequence.

[0326] 44. An artificial intelligence-based system according to clause 63, wherein, during inference, the chromatin model uses a trained first set of weights,

[0327] Wherein, during the inference, the chromatin output generation logic component uses the trained second set of weights.

[0328] 45. An artificial intelligence-based system according to clause 1, wherein the gene expression model has a third set of weights,

[0329] The gene expression output generation logic component has a fourth set of weights.

[0330] 46. ​​An artificial intelligence-based system according to clause 45, wherein the third set of weights of the gene expression model is trained from scratch to process the output sequence and generate the alternative representation of the output sequence,

[0331] The fourth set of weights of the gene expression output generation logic component is trained end-to-end from scratch together with the third set of weights of the gene expression model to process the alternative representation of the output sequence and generate the gene expression output sequence.

[0332] 47. An artificial intelligence-based system according to clause 46, wherein, during inference, the gene expression model uses a trained third set of weights,

[0333] Wherein, during the inference, the gene expression output generation logic component uses a trained fourth set of weights.

[0334] 48. An artificial intelligence-based system according to clause 45, wherein, during training, the first set of weights of the chromatin model is first trained to process the input base sequence and generate the alternative representation of the input base sequence, and then the first set of weights is retrained as a substitute for the third set of weights of the gene expression model to process the output sequence and generate the alternative representation of the output sequence, wherein the fourth set of weights of the gene expression output generation logic component are trained end-to-end from scratch together with the replaced trained first set of weights in the gene expression model to process the alternative representation of the output sequence generated by the replaced trained first set of weights in the gene expression model and generate the gene expression output sequence.

[0335] 49. An artificial intelligence-based system according to clause 48, wherein, during inference, the chromatin model uses the retrained first set of weights,

[0336] wherein, during the inference, the chromatin output generation logic uses a trained second set of weights,

[0337] wherein, during the inference, the gene expression model uses the retrained first set of weights, and

[0338] Wherein, during the inference, the gene expression output generation logic component uses a trained fourth set of weights.

[0339] 50. The artificial intelligence-based system of clause 19, wherein during training, the first set of weights of the chromatin model is first trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence,

[0340] wherein during the training, the second set of weights of the chromatin output generation logic component is first trained end-to-end from scratch together with the first set of weights of the chromatin model to process the alternative representation of the input base sequence and generate the output sequence,

[0341] wherein, during the training, the trained first set of weights of the chromatin model is then retrained to process the reference base sequence and generate the alternative representation of the reference base sequence, and the alternative base sequence and generate the alternative representation of the alternative base sequence, and

[0342] Wherein, during the training, the trained second set of weights of the chromatin output generation logic component is then re-trained end-to-end together with the trained first set of weights of the chromatin model to process the alternative representation of the reference base sequence and generate the reference output sequence, and the alternative representation of the alternative base sequence and generate the alternative output sequence.

[0343] 51. An artificial intelligence-based system according to clause 50, wherein, during inference, the chromatin model uses the retrained first set of weights,

[0344] wherein, during the inference, the chromatin output generation logic component uses the retrained second set of weights.

[0345] 52. The artificial intelligence-based system of clause 21, wherein the pathogenicity prediction logic component has a fifth set of weights.

[0346] 53. The artificial intelligence-based system of clause 52, wherein during training, the first set of weights of the chromatin model is first trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence,

[0347] wherein, during the training, the second set of weights of the chromatin output generation logic component is first trained end-to-end from scratch together with the first set of weights of the chromatin model to process the alternative representation of the input base sequence and generate the output sequence, and

[0348] Wherein, during the training, the trained first set of weights of the chromatin model and the trained second set of weights of the chromatin output generation logic component are then retrained end-to-end to generate the pathogenicity prediction of the alternative base.

[0349] 54. An artificial intelligence-based system according to clause 53, wherein, during inference, the chromatin model uses the retrained first set of weights,

[0350] wherein, during the inference, the chromatin output generation logic uses the retrained second set of weights, and

[0351] Wherein, during the inference, the pathogenicity prediction logic component uses a trained fifth set of weights.

[0352] 55. An artificial intelligence-based system according to clause 19, wherein in the reference output sequence, for the given reference target base at the given position in the reference target base sequence, the given per-base reference chromatin output specifies the confounding factor signal level of the given reference target base at the given position.

[0353] 56. An artificial intelligence-based system according to clause 20, wherein in the alternative output sequence, for the given alternative target base at the given position in the alternative target base sequence, the given per-base alternative chromatin output specifies the confounding factor signal level of the given alternative target base at the given position.

[0354] 57. An artificial intelligence-based system according to claim 1, wherein, during training, the chromatin model and the chromatin output generation logic component are first trained end-to-end from scratch to convert an analysis of an input base sequence into a base-by-base evolutionarily conserved chromatin sequence, and then retrained end-to-end to convert an analysis of an input base sequence into a base-by-base transcription start frequency chromatin sequence.

[0355] 58. An artificial intelligence-based system according to claim 1, wherein, during training, the chromatin model and the chromatin output generation logic component are first trained end-to-end from scratch to convert the analysis of the input base sequence into a base-by-base confounding factor signal level chromatin sequence, and then re-trained end-to-end to convert the analysis of the input base sequence into a base-by-base evolutionarily conserved chromatin sequence.

[0356] 59. An artificial intelligence-based system according to claim 1, wherein, during training, the chromatin model and the chromatin output generation logic component are first trained end-to-end from scratch to convert the analysis of the input base sequence into a base-by-base confounding factor signal level chromatin sequence, and then re-trained end-to-end to convert the analysis of the input base sequence into a base-by-base transcription start frequency chromatin sequence.

[0357] 60. An artificial intelligence-based system according to claim 1, wherein, during training, the chromatin model and the chromatin output generation logic component are first trained end-to-end from scratch to convert the analysis of the input base sequence into a base-by-base confounding factor signal level chromatin sequence, and then re-trained end-to-end to convert the analysis of the input base sequence into a base-by-base evolutionary conserved chromatin sequence and a base-by-base transcription start frequency chromatin sequence.

[0358] 61. The artificial intelligence-based system according to clause 1 is further configured to include a first training set of training input base sequences, wherein the first training set includes variants confounded by the effects of multiple confounding factors.

[0359] 62. An artificial intelligence-based system according to claim 61, wherein the multiple confounder effects include inter-chromosomal effects, intragenic effects, population structure and ancestral effects, probability estimates of expression residual (PEER) effects, environmental effects, sex effects, batch effects, genotyping platform effects and / or library construction protocol effects.

[0360] 63. The artificial intelligence-based system according to clause 61 is further configured to include a second training set of training input base sequences, wherein the second training set includes variants that are not confounded by the effects of the multiple confounding factors.

[0361] 64. An artificial intelligence-based system according to clause 63, wherein the variants in the second training set are reliably determined to alter gene expression and cause extreme gene expression levels.

[0362] 65. An artificial intelligence-based system according to clause 64, wherein the variants in the second training set include overexpression-causing variants that increase gene expression levels.

[0363] 66. An artificial intelligence-based system according to clause 64, wherein the variants in the second training set include under-expression-causing variants that reduce gene expression levels.

[0364] 67. An artificial intelligence-based system according to clause 65, wherein the second training set specifies an overexpression probability of the variant, wherein the overexpression probability specifies the likelihood that the variant causes overexpression of a gene.

[0365] 68. An artificial intelligence-based system according to clause 66, wherein the second training set specifies a probability of underexpression of the variant, wherein the probability of underexpression specifies the likelihood that the variant causes underexpression of a gene.

[0366] 69. An artificial intelligence-based system according to clause 64, wherein each variant in the second training set is a rare variant that occurs in an outlier individual among a cohort of outlier individuals,

[0367] The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.

[0368] 70. An artificial intelligence-based system according to clause 69, wherein the extreme gene expression levels are determined from the tail quantiles of the normalized gene expression levels.

[0369] 71. An artificial intelligence-based system according to clause 70, wherein the extreme gene expression levels include gene overexpression and gene underexpression.

[0370] 72. An artificial intelligence-based system according to clause 69, wherein the rare variant is a coding variant.

[0371] 73. An artificial intelligence-based system according to clause 69, wherein the rare variant is a non-coding variant.

[0372] 74. An artificial intelligence-based system according to clause 73, wherein the non-coding variant is a 5' untranslated region (UTR) variant, a 3'UTR variant, an enhancer variant or a promoter variant.

[0373] 75. An artificial intelligence-based system according to clause 64, wherein the variants in the second training set span multiple tissue types.

[0374] 76. An artificial intelligence-based system according to clause 64, wherein the variants in the second training set span multiple cell types.

[0375] 77. An artificial intelligence-based system according to clause 1, wherein the input base sequence and the multiple output sequences span the multiple tissue types.

[0376] 78. An artificial intelligence-based system according to claim 1, wherein the input base sequence and the multiple output sequences span the multiple cell types.

[0377] 79. An artificial intelligence-based system according to clause 12, wherein the gene expression output sequences span the multiple tissue types.

[0378] 80. An artificial intelligence-based system according to clause 12, wherein the gene expression output sequences span the multiple cell types.

[0379] 81. An artificial intelligence-based system according to claim 1, wherein the chromatin model and the chromatin output generation logic component are first trained end-to-end on the first training set and then retrained on the second training set.

[0380] 82. An artificial intelligence-based system according to claim 1, wherein the variants in the second training set are a pathogenic set labeled with a first ground truth label indicating that gene expression is changed, and the common variants are a benign set labeled with a second ground truth label indicating that gene expression is not changed.

[0381] 83. An artificial intelligence-based system according to clause 102, wherein the benign set is balanced in terms of trinucleotide background, homopolymers, k-mers, genomic GC frequency, and sequencing depth.

[0382] 84. An artificial intelligence-based system according to claim 102, wherein, based on a cutoff probability applied to the overexpression probability and the underexpression probability, the variants in the second training set are divided into an overexpression variant training set having a first ground truth label indicating an increase in gene expression, an underexpression variant training set having a second ground truth label indicating a decrease in gene expression, and a neural expression variant training set indicating maintenance of gene expression.

[0383] 85. An artificial intelligence-based system according to clause 12, wherein the gene expression model and the gene expression output generation logic component are first trained end-to-end on the first training set and then retrained on the second training set.

[0384] 86. An artificial intelligence-based system according to claim 1, wherein the chromatin model and the chromatin output generation logic component are first trained end-to-end on the first training set, and then retrained on those variants that appear on odd chromosomes in the second training set.

[0385] 87. An artificial intelligence-based system according to claim 12, wherein the gene expression model and the gene expression output generation logic component are first trained end-to-end on the first training set, and then retrained on those variants that appear on odd chromosomes in the second training set.

[0386] 88. An artificial intelligence-based system according to clause 64, wherein the variants in the second training set are not used for training, but are used as a validation set to evaluate the performance of the trained chromatin model, the trained chromatin output generation logic component, the trained gene expression model and the trained gene expression output generation logic component.

[0387] 89. An artificial intelligence-based system according to clause 88, wherein those variants in the second training set that appear on even chromosomes are used as the validation set.

[0388] 90. An artificial intelligence-based system according to claim 1, wherein the size of the target base sequence varies during training to account for changes in the offset position of the transcription start site (TSS).

[0389] 91. An artificial intelligence-based system for detecting changes in gene expression at per-base resolution, comprising:

[0390] an input generation logic component that accesses a sequence database and generates an input base sequence, wherein the input base sequence includes a target base sequence, and wherein the target base sequence is flanked by a right-end base sequence having downstream background bases, and a left-end base sequence having upstream background bases;

[0391] a chromatin model that processes the input base sequence and generates an alternative representation of the input base sequence; and

[0392] an output generation logic component that processes the alternative representations of the input base sequence and generates an output sequence of corresponding per-base chromatin outputs for corresponding target bases in the target base sequence,

[0393] Wherein in the output sequence, for a given target base at a given position in the target base sequence, a given per-base chromatin output specifies a measure of transcription initiation of the given target base at the given position.

[0394] 92. An artificial intelligence-based system for detecting changes in gene expression at per-base resolution, comprising:

[0395] an input generation logic component that accesses a sequence database and generates an input base sequence, wherein the input base sequence includes a target base sequence, and wherein the target base sequence is flanked by a right-end base sequence having downstream background bases, and a left-end base sequence having upstream background bases;

[0396] a gene expression model that processes the input base sequence and generates an alternative representation of the input base sequence; and

[0397] a gene expression model output generation logic component that processes the alternative representations of the input base sequence and generates a gene expression output sequence of corresponding per-base gene expression outputs for corresponding target bases in the target base sequence,

[0398] Wherein for the given target base at the given position in the gene expression output sequence, the given per-base gene expression output specifies a measure of the gene expression level of the given target base at the given position.

[0399] 93. An artificial intelligence-based system for detecting changes in gene expression at per-base resolution, comprising:

[0400] an input generation logic component that accesses a sequence database and generates an input base sequence, wherein the input base sequence includes a target base sequence, and wherein the target base sequence is flanked by a right-end base sequence having downstream background bases, and a left-end base sequence having upstream background bases;

[0401] a chromatin model that processes the input base sequence and generates an alternative representation of the input base sequence; and

[0402] A chromatin output generation logic component processes the alternative representations of the input base sequence and generates an output sequence of corresponding per-base chromatin outputs for corresponding target bases in the target base sequence.

[0403] 94. An artificial intelligence-based system according to clause 93, wherein in the output sequence, for a given target base at a given position in the target base sequence, a given per-base chromatin output specifies a measure of the evolutionary conservation of the given target base across multiple species.

[0404] 95. An artificial intelligence-based system according to clause 93, wherein the given per-base chromatin output further specifies a transcription start measure for the given target base at the given position.

[0405] Although the present invention is disclosed by referring to the preferred embodiments and examples described in detail above, it should be understood that these examples are intended to be illustrative rather than restrictive. It is expected that modifications and combinations will be easily conceived by those skilled in the art, and these modifications and combinations will be within the spirit of the present invention and the scope of the following claims.

Claims

1. A computer-implemented method for identifying rare variants that cause extreme gene expression levels, comprising: Obtain gene expression levels for a group of individuals; normalizing the gene expression levels, and identifying those outlier individuals from the group of individuals having extreme gene expression levels, wherein the extreme gene expression levels are determined from tail quantiles of the normalized gene expression levels; selecting rare variants from a gene sequence of the outlier individual, wherein the rare variants are selected based on an allele frequency cutoff; fitting a causal relationship model to determine the causal relationship between the rare variant in the outlier individual and the extreme gene expression level while controlling for multiple confounding factors; as well as A causality score for the rare variant is generated based on the determined causality, wherein a specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant.

2. The computer-implemented method of claim 1, wherein the causality score is a probability value (p-value).

3. The computer-implemented method of claim 1, wherein the p-value is determined by a Pearson correlation coefficient.

4. The computer-implemented method of claim 1, wherein the causal relationship model is a logistic regression model, a linear regression model, an analysis of covariance (ANCOVA) model, and / or a multivariate analysis of covariance (MANCOVA) model.

5. The computer-implemented method of claim 4, wherein fitting the causal relationship model determines the causal relationship by predicting a specific gene expression level of a specific gene in a specific chromosome based on a variant-driven gene expression level caused by a specific rare variant.

6. A computer-implemented method according to claim 5, wherein the fitted causal model measures the contribution of the variant-driven gene expression level as a variant effect covariate.

7. The computer-implemented method of claim 1, wherein the plurality of confounding factors comprises distal trans-expression quantitative trait loci (eQTL) effects.

8. The computer-implemented method of claim 7, wherein the fitted causal model controls the distal trans-eQTL effect by predicting the specific gene expression level from trans-gene expression levels caused by other genes in other chromosomes.

9. The computer-implemented method of claim 8, wherein the fitted causal model measures the contribution of the trans-acting gene expression level as a trans-effect covariate.

10. The computer-implemented method of claim 1, wherein the plurality of confounding factors comprises local cis-eQTL effects.

11. The computer-implemented method of claim 10, wherein the fitted causal model controls for the local cis-eQTL effect by predicting the specific gene expression level from the cis-gene expression levels caused by the presence of multiple common variants in the specific gene neighborhood.

12. The computer-implemented method of claim 11, wherein the neighborhood is defined by an offset relative to a transcription start site (TSS) in the specific gene.

13. The computer-implemented method of claim 11, wherein the fitted causal model measures the contribution of the cis gene expression level as a cis-effect covariate.

14. The computer-implemented method of claim 1, wherein the plurality of confounding factors comprises population structure and ancestry effects.

15. The computer-implemented method of claim 14, wherein the population structure and ancestry effects are represented by one or more genotype-based principal components (gPCs).

16. The computer-implemented method of claim 15, wherein the fitted causal model controls for the population structure and ancestry effects by predicting the specific gene expression level based on the gPC gene expression level caused by the gPC.

17. The computer-implemented method of claim 16, wherein the fitted causal model measures the gPC gene expression levels as contributions of population structure and ancestry effect covariates.

18. The computer-implemented method of claim 1, wherein the plurality of confounders comprises a probability estimate of expression residual (PEER) effect.

19. The computer-implemented method of claim 18, wherein the fitted causal model controls the PEER effect by predicting the specific gene expression level based on the PEER gene expression level caused by the PEER.

20. The computer-implemented method of claim 19, wherein the fitted causal model measures the contribution of the PEER gene expression level as a PEER effect covariate.

21. The computer-implemented method of claim 1, wherein the plurality of confounding factors includes environmental effects.

22. The computer-implemented method of claim 21, wherein the fitted causal model controls for the environmental effect by predicting the specific gene expression level based on the environmental gene expression level caused by the environmental effect.

23. A computer-implemented method according to claim 22, wherein the fitted causal model measures the contribution of the environmental gene expression level as an environmental effect covariate.

24. The computer-implemented method of claim 1, wherein the plurality of confounding factors comprises sex effect, batch effect, genotyping platform effect, and library construction protocol effect.

25. The computer-implemented method of claim 1, wherein the extreme gene expression levels include gene overexpression and gene underexpression.

26. The computer-implemented method of claim 24, further comprising determining the causal relationship between the rare variant and the gene overexpression while controlling for the plurality of confounding factors.

27. The computer-implemented method of claim 25, further comprising generating an excess causality score for the rare variant, wherein the specific excess causality score for the specific rare variant indicates the likelihood that the specific rare variant causes overexpression of a gene in those outlier individuals whose gene sequences contain the specific rare variant.

28. The computer-implemented method of claim 26, wherein the excess causality score is an excess probability value (excess p-value).

29. The computer-implemented method of claim 27, wherein the excess p-value is determined by a Pearson correlation coefficient.

30. The computer-implemented method of claim 27, wherein the excess p-value specifies a statistically unconfounded likelihood that the rare variant increases gene expression in a gene that otherwise has lower gene expression relative to other genes in a gene set.

31. The computer-implemented method of claim 24, further comprising determining the causal relationship between the rare variant and underexpression of the gene while controlling for the plurality of confounding factors.

32. The computer-implemented method of claim 30, further comprising generating a deficiency causality score for the rare variant, wherein the specific deficiency causality score for the specific rare variant indicates the likelihood that the specific rare variant causes insufficient gene expression in those outlier individuals whose gene sequences contain the specific rare variant.

33. The computer-implemented method of claim 31, wherein the deficiency causality score is a deficiency probability value (deficiency p-value).

34. The computer-implemented method of claim 32, wherein the deficiency p-value is determined by a Pearson correlation coefficient.

35. The computer-implemented method of claim 32, wherein the insufficient p-value specifies a statistically unconfounded likelihood that the rare variant reduces gene expression in a gene that otherwise has higher gene expression relative to other genes in the gene set.

36. The computer-implemented method of claim 1, wherein the rare variant is a non-coding variant.

37. The computer-implemented method of claim 17, wherein the non-coding variants include 5' untranslated region (UTR) variants, 3' UTR variants, enhancer variants, and promoter variants.

38. The computer-implemented method of claim 1, wherein the gene expression levels are further stratified into tissue-specific gene expression levels of multiple tissues.

39. The computer-implemented method of claim 37, wherein quantile normalization is used to normalize the gene expression levels of each gene in each tissue.

40. The computer-implemented method of claim 38, wherein the causal model is fit separately for each tissue.

41. The computer-implemented method of claim 1, wherein the causal model is fit using stratification.

42. The computer-implemented method of claim 1, further comprising generating a ranking of the rare variants based on the causality score.

43. The computer-implemented method of claim 26, further comprising generating a ranking of the rare variants based on the excess causality score.

44. The computer-implemented method of claim 31 , further comprising generating a ranking of the rare variants based on the insufficient causality scores.

45. The computer-implemented method of claim 1, wherein the rare variant is a single-occurring variant.

46. ​​The computer-implemented method of claim 44, wherein a single-occurring variant occurs in only one outlier individual from among the outlier individuals.

47. A system comprising one or more processors coupled to a memory loaded with computer instructions to identify rare variants that cause extreme gene expression levels, the instructions when executed on the processors performing actions comprising: Obtain gene expression levels for a group of individuals; normalizing the gene expression levels, and identifying those outlier individuals from the group of individuals having extreme gene expression levels, wherein the extreme gene expression levels are determined from tail quantiles of the normalized gene expression levels; selecting rare variants from a gene sequence of the outlier individual, wherein the rare variants are selected based on an allele frequency cutoff; fitting a causal relationship model to determine the causal relationship between the rare variant in the outlier individual and the extreme gene expression level while controlling for multiple confounding factors; as well as A causality score for the rare variant is generated based on the determined causality, wherein a specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant.

48. The system of claim 46, further implemented to include performing the actions of claims 2 to 45.

49. A non-transitory computer readable storage medium having printed thereon computer program instructions for identifying rare variants that cause extreme gene expression levels, the instructions when executed on a processor implementing a method comprising: Obtain gene expression levels for a group of individuals; normalizing the gene expression levels, and identifying those outlier individuals from the group of individuals having extreme gene expression levels, wherein the extreme gene expression levels are determined from tail quantiles of the normalized gene expression levels; selecting rare variants from a gene sequence of the outlier individual, wherein the rare variants are selected based on an allele frequency cutoff; fitting a causal relationship model to determine the causal relationship between the rare variant in the outlier individual and the extreme gene expression level while controlling for multiple confounding factors; as well as A causality score for the rare variant is generated based on the determined causality, wherein a specific causality score for a specific rare variant indicates the likelihood that the specific rare variant causes an extreme gene expression level in those outlier individuals whose gene sequences contain the specific rare variant.

50. The non-transitory computer readable storage medium of claim 46, implementing the method further comprising executing claims 2 to 45.

Citation Information

Cited By

  • Method and device for predicting RNA variable region structure, medium and program product

    CN119541629A