Artificial intelligence-based gene conservative and expression retention detection at base resolution
By processing genomic data using a deep learning biomass model, multiple biomass output sequences are generated, solving the problem of local dependence of complex data in genomics analysis and achieving more accurate gene expression and pathogenicity prediction.
Patent Information
- Application Number
- CN202380030454.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-08-05
- Filing Date
- 2023-08-04
- Publication Date
- 2024-11-12
AI Technical Summary
Existing technologies struggle to effectively utilize the complex spatial and longitudinal data local dependencies in genomics data, resulting in insufficient accuracy and efficiency in genomics analysis, particularly in predicting gene expression and the pathogenicity of genetic variants.
A deep learning-based biomass model is employed, which processes genomic data through convolutional neural networks and recurrent neural networks to generate multiple biomass output sequences to capture the evolutionary conservation and epigenetic characteristics of genomic sequences. Combined with a pathogenicity prediction logic component, gene expression and pathogenicity prediction at base resolution are achieved.
It improves the accuracy and efficiency of genomics analysis, particularly in predicting gene expression and pathogenicity of genetic variants, and provides a more comprehensive understanding of the evolutionary and epigenetic characteristics of genome sequences.
Smart Images

Figure CN118946932A_ABST
Abstract
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 disclosed technology relates to artificial intelligence-based gene conservation and expression retention detection 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 EPIGENETICSAT BASE RESOLUTION" (Attorney Docket No. ILLM 1040-1 / 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 / IP-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 (which are instances 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 can 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 in both spatial and longitudinal data must be accounted for. For example, shuffling the pixels of a DNA sequence or an image can severely disrupt informative 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 particular 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 instances in the sequence or position weight matrix (PWM) matching, can be used for this task. Since the k-mer or PWM instance frequencies are robust to shifting the motif within the 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 TAL1. 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 sequence alone. Applications include classifying transcription factor binding sites, and predicting molecular phenotypes such as chromatin signatures, DNA contact maps, DNA methylation, gene expression, translation efficiency, RBP binding, and microRNA (miRNA) targets. In addition to predicting molecular phenotypes from sequence, CNNs can be applied to more technical tasks that are traditionally solved by manual bioinformatics pipelines. For example, CNNs can predict guide RNA specificity, denoise ChIP–seq, improve Hi-C data resolution, predict laboratory of origin from DNA sequence, and call genetic variants. CNNs have also been used to model long-range dependencies in the genome. Although interacting regulatory elements may be located far from each other on an unfolded linear DNA sequence, these elements are often close to each other in the actual 3D chromatin conformation. Therefore, while modeling molecular phenotypes from 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 based on 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 is a flow chart illustrating the process of a system for determining evolutionary and epigenetic characteristics of a genetic sequence.
[0051] Figure 2 An exemplary input base sequence comprising nucleotide bases extracted from a sequence database is schematically shown, wherein the target base sequence is flanked by a left-end sequence containing upstream background bases and a right-end sequence containing downstream background bases.
[0052] Figure 3 Examples of alternative sequences from an exemplary reference genetic sequence with two exemplary alternative sequences are shown, and wherein the alternative sequences have corresponding single nucleotide variants at a single base position, but otherwise have the same composition as the reference sequence.
[0053] Figure 4 The genetic composition of sequences belonging to a training data set developed for one specific implementation of the disclosed technology is shown.
[0054] Figure 5 Schematically shows the application Figure 1 A specific implementation of the training procedure of the system, in which the model is used Figure 4 The first training set described in , and then with Figure 4 Retrain on the second training set described in .
[0055] Figure 6 Schematically shows the application Figure 1 Another specific implementation of the training procedure of the system of Figure 4 The first training set described in Figure 4 The model is retrained on a subset of the second training set described in , and then the model is validated on the remaining subset of samples from the second training set.
[0056] Figure 7 is used for variant classification Figure 1 Schematic diagram of a specific implementation of a system for comparing a reference sequence and an alternative sequence at base resolution by comparing the corresponding model outputs for each sequence.
[0057] Figure 8 is a flow chart of one specific implementation of the disclosed technology, in which the comparison of a reference sequence and an alternative sequence is quantified by average delta value.
[0058] Fig. 9 is a flow chart of one specific implementation of the disclosed technology, in which the comparison of a reference sequence and an alternative sequence is quantified by a final sum increment value.
[0059] Fig.10 is a flowchart of a specific implementation of the disclosed technology, wherein a biomass model generates two biomass output sequences from an input base sequence via a first set of weights and a second set of weights that are trained end-to-end.
[0060] Fig.11 is a flowchart of a specific implementation of the disclosed technology, wherein a biomass model generates three biomass output sequences from an input base sequence via a first set of weights and a second set of weights that are trained end-to-end.
[0061] Fig.12 is a flowchart of a specific implementation of the disclosed technology, wherein a biomass model includes a first set of weights for generating an alternative representation of an input base sequence and a second set of weights trained from scratch to generate multiple biomass output sequences from the alternative representation of the input base sequence.
[0062] Fig.13 is a flowchart of a specific implementation of the disclosed technology, wherein a biomass model generates a biomass output sequence from an input base sequence via a first set of weights and a second set of weights, wherein the first set of weights and the second set of weights are trained end-to-end and retrained to generate subsequent biomass output sequences on a single basis.
[0063] Fig.14 is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model includes a first set of weights and a second set of weights that are trained end-to-end to generate multiple biomass output sequences from an input base sequence, and a third set of weights and a fourth set of weights that are trained end-to-end to generate gene expression output sequences from the input multiple biomass output sequences.
[0064] Fig.15 is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model includes a first set of weights for generating alternative representations of an input base sequence, a second set of weights trained from scratch to generate multiple biomass output sequences from the alternative representations of the input base sequence, and a third set of weights and a fourth set of weights trained end-to-end to generate gene expression output sequences from the input multiple biomass output sequences.
[0065] Fig.16 is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model includes a first set of weights for generating alternative representations of an input base sequence, a second set of weights trained from scratch to generate multiple biomass output sequences, a third set of weights trained from scratch to generate alternative biomass representations from multiple biomass output sequences, and a fourth set of weights trained from scratch to generate gene expression output sequences from the alternative biomass representations.
[0066] Fig.17is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model includes a first set of weights and a second set of weights that are trained end-to-end to generate multiple biomass output sequences from an input base sequence, a third set of weights that are trained from scratch to generate alternative biomass representations from the multiple biomass output sequences, and a fourth set of weights that are trained from scratch to generate gene expression output sequences from the alternative biomass representations.
[0067] Fig.18 is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model includes training to generate multiple biomass output sequences from an input base sequence, then retraining a first set of weights as a substitute for a third set of weights to generate alternative biomass representations from the multiple biomass output sequences, and end-to-end training with the replaced first set of weights to generate a fourth set of weights from the alternative biomass representations.
[0068] Fig.19 is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model includes a first set of weights trained to generate alternative sequence representations from an input base sequence, then retrained as a substitute for a third set of weights to generate alternative biomass representations from multiple biomass output sequences, and a fourth set of weights trained from scratch to generate gene expression output sequences from the alternative biomass representations.
[0069] Fig. 20 is a flowchart of a specific implementation of the disclosed technology, wherein a biomass model includes a first set of weights trained to generate alternative sequence representations from an input base sequence, a second set of weights trained to generate multiple biomass output sequences from alternative representations of the input bases, the retrained first set of weights used as a substitute for the third set of weights to generate alternative biomass representations from multiple biomass output sequences, and a fourth set of weights trained from scratch to generate gene expression output sequences from alternative biomass representations.
[0070] Fig.21 is a flowchart of a specific implementation of the disclosed technology, wherein a biomass model includes a first set of weights trained to generate alternative sequence representations from an input base sequence, a second set of weights trained to generate multiple biomass output sequences from alternative representations of the input bases, a retrained first set of weights used as a substitute for a third set of weights to generate alternative biomass representations from multiple biomass output sequences, and a fourth set of weights trained end-to-end with the replaced first set of weights to generate gene expression output sequences from the alternative biomass representations.
[0071] Fig. 22is a flowchart of a specific implementation of the disclosed technology, wherein the model is further configured to include a pathogenicity prediction logic component, which compares a reference biomass output sequence and an alternative biomass output sequence at base resolution, wherein a fifth set of weights generates a prediction of pathogenicity of the alternative sequence from multiple biomass output sequences, and wherein the first set of weights and the second set of weights are each trained from scratch.
[0072] Fig.23 is a flowchart of a specific implementation of the disclosed technology, wherein the model is further configured to include a pathogenicity prediction logic component, which compares a reference biomass output sequence and an alternative biomass output sequence at base resolution, wherein a fifth set of weights generates a pathogenicity prediction of the alternative sequence from multiple biomass output sequences, and wherein the first set of weights and the second set of weights are trained end-to-end.
[0073] Fig.24 is a schematic diagram of an evolutionary conservation measure that can be generated by a biomass model from an input base sequence as a value for a first biomass output sequence.
[0074] Fig.25 is a schematic diagram of a transcription initiation measure that can be generated by a biomass model from an input base sequence as a value for a second biomass output sequence.
[0075] Fig.26 is a schematic diagram of epigenetic signals that can be generated by a biomass model from an input base sequence as a value of a third biomass output sequence.
[0076] Fig. 27 is a flow chart of one specific implementation of the disclosed technology, wherein an expression change classifier is configured to predict the effect of a variant on gene expression.
[0077] Fig.28 is a flow chart of one specific implementation of the disclosed technology, wherein the expression change classifier is further configured as an expression reduction classifier to predict whether the variant reduces gene expression or does not reduce gene expression.
[0078] Fig.29 is a flow chart of one specific implementation of the disclosed technology, wherein the expression change classifier is further configured as an expression decrease classifier to predict whether the variant increases or does not increase gene expression.
[0079] Fig.30 is a flowchart of one specific implementation of the disclosed technology, wherein the expression change classifier is further configured as a multi-class expression classifier that predicts whether a variant preserves gene expression, decreases gene expression, or increases gene expression.
[0080] Fig.31is a flowchart of one specific implementation of the disclosed technology, in which gene expression classifier training is used for comparison of ground truth causal scores with inferred causal scores.
[0081] Fig.32 An example computer system is shown that can be used to implement the disclosed techniques. DETAILED DESCRIPTION
[0082] 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.
[0083] The detailed description of various implementations will be better understood when read in conjunction with the accompanying drawings. To the extent that the 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.
[0084] 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.
[0085] Example applications of the disclosed technology
[0086] As emphasized by the numerous comparisons of various implementations of the biomass model 124, many implementations share overlap in architectural components. With respect to the phrase "biomass model," the biomass model predicts multiple categories of biomass from a genomic sequence. Examples of biomass categories include protein (transcription factor) binding, methylation, histone modification, DNA accessibility, conservation. Among these, methylation and histone modification are considered epigenetics. In contrast, chromatin refers to histone modification and DNA accessibility.
[0087] Thus, in some implementations, a biomass model may be referred to as an “epigenetic model.” In other implementations, a biomass model may be referred to as a “chromatin model.” In other implementations, a biomass model may be referred to as a “chromatin and epigenetic model” or an “epigenetic and chromatin model.”
[0088] Each element of the biomass model 124 has a variety of specific implementations, which can be combined in a variety of configurations. The various arrangements that can be implemented for the technology disclosed in the present invention not only provide a wide range of practicality, performance efficiency, but also provide a wide range of performance accuracy. The data transformation applied to the input base sequence in many specific implementations of the technology disclosed in the present invention generates a variety of additional sequence formats from the perspective of nucleic acid sequence and chromatin structure. It 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 adopted a variety of tools to classify variant pathogenicity with high performance. The biomass model 124 introduces another tool in the method, as well as an additional dimension when studying epigenetic signals that affect biological replication and transcription processes.
[0089] 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.
[0090] 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.
[0091] Biomass Model Overview
[0092] Figure 1 1 is a flow chart illustrating a process 100 of a system for determining evolutionary and epigenetic characteristics of a genetic sequence. An input base sequence 122 is extracted from a sequence database 110 and processed by a biomass model 124 that generates alternative representations of the input base sequence 126. The alternative representations of the input base sequence 126 are converted into alternative biomass representations in the form of a plurality of biomass output sequences 136.
[0093] Biomass model input data
[0094] Figure 2 An exemplary input base sequence 200 comprising nucleotide bases extracted from a sequence database 202 is schematically shown, wherein a target base sequence 226 is flanked by a left end sequence 224 containing upstream background bases and a right end sequence 228 containing downstream background bases. The upstream background base 224 is a nucleotide base sequence {x1, x2, x3, ..., x n}, which can be equal to adenine, thymine, cytosine or guanine. The target base sequence 226 is then followed by the upstream background base 224. The target base sequence 226 contains the nucleotide base sequence {y1, y2, y3, ..., y n}, which can be equal to adenine, thymine, cytosine or guanine. The downstream background base 228 then follows the target base sequence 226. The downstream background base is the nucleotide base sequence {z1, z2, z3, ..., z n}, which can be equal to adenine, thymine, cytosine or guanine. The input base sequence 200 may also contain unknown or missing gap positions in the upstream background base 224, the target base sequence 226 or the downstream background base 228.
[0095] Figure 3An example of an alternative sequence 300 from an exemplary reference genetic sequence 302 with two exemplary alternative sequences 322 and 342 is shown, wherein the alternative sequences have corresponding single nucleotide variants at a single base position, but otherwise have the same composition as the composition of the reference sequence. For example, the single nucleotide substitution is shown as variant 326 and variant 336 compared to nucleotide 306. In addition to having the same target base sequence, upstream sequences 304, 324, and 344 are identical to each other, and downstream sequences 308, 328, and 348 are identical to each other.
[0096] Figure 4 The genetic composition of sequences belonging to a training data set 400 developed for one specific implementation of the disclosed technology is shown. One training data set 422 contains alternative sequences confounded by epigenetic effects (e.g., alternative sequence A 432 with single nucleotide variant 433), and a second training data set 452 contains alternative sequences not confounded by epigenetic effects (e.g., sequence 462 with single nucleotide variant 463). Single nucleotide variants 433 and 463 differ in composition from reference base position 403; however, all other base positions within the reference sequence, alternative sequence A 432, and alternative sequence B 462 do not differ.
[0097] Biomass model structure
[0098] Figure 5 Schematically shows the application Figure 1 A specific implementation 500 of the training procedure of the system 100, wherein the biomass model 124 is used Figure 4 The first training set described in , and then with Figure 4 Retrain on the second training set described in .
[0099] The sequences obtained from the sequence database 502 are first obtained from a first training data set for a first set of training iterations 566 to generate a plurality of biomass output sequences 548 from an input base sequence 524, wherein the biomass model 124 is configured to detect gene expression changes at base resolution, including a biomass model 124 that processes the input base sequence 524 and generates an alternative representation (e.g., a convolutional representation) of the input base sequence 524, and a biomass output sequence generator 528. The biomass model 124 then performs a second set of training iterations 586 on the sequences obtained from the second training data set without making any changes to the model configuration of the biomass model 124.
[0100] In one embodiment, the biomass model 124 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 non-exponentially from a lower residual block group to a higher residual block group. In other embodiments, the dilated convolution rate progresses exponentially. The convolution window size varies between multiple groups of residual blocks, each residual block including 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.
[0101] In one implementation, 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 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.
[0102] In one implementation, 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 one implementation, 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.
[0103] In one implementation, the input has one thousand upstream flanking background bases (Cu) on the left side of the input sequence and one thousand 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 one implementation, 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.
[0104] In one implementation, the input has five thousand upstream flanking background bases (Cu) on the left side of the input sequence and five thousand 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 one implementation, there are at least four groups of four residual blocks and at least four jump 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.
[0105] In general, the biomass model 124 can be a rule-based model, a tree-based model, or a machine learning model. Examples include multi-layer perceptrons (MLPs), feed-forward neural networks, fully connected neural networks, fully convolutional neural networks, ResNets, sequence-to-sequence (Seq2Seq) models (such as WaveNet), semantic segmentation neural networks, and generative adversarial networks (GANs) (e.g., CycleGAN, StyleGAN, pixelRNN, text-2-image, DiscoGAN, IsGAN).
[0106] In some implementations, the biometric model 124 can include a self-attention mechanism such as a Transformer, Visual Transformer (ViT), Bidirectional Transformer (BERT), Detection Transformer (DETR), 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.
[0107] In some embodiments, examples of the biomass model 124 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.
[0108] In some implementations, the biomass model 124 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 biomass model 124 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 biomass model 124 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 biomass model 124 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.
[0109] In some implementations, the biomass model 124 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 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 biomass model 124 can be an ensemble of multiple models.
[0110] In some implementations, the biomass model 124 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 AMSGrad.
[0111] Figure 6 Schematically shows the application Figure 1 Another embodiment 600 of the training procedure of the system 100 is to utilize Figure 4 The first training set described in Figure 4 The biomass model 124 is retrained on a subset of the second training set described in , and then the remaining sample subset from the second training set is used to perform model validation on the biomass model 124. The sequences obtained from the sequence database 602 are first obtained from the first training data set for the first set of training iterations 666 to generate a plurality of biomass output sequences 648 from the input base sequence 624, wherein the biomass model 124 is configured to detect gene expression changes at base resolution, including a biomass model 124 that processes the input base sequence 624 and generates an alternative representation (e.g., a convolutional representation) of the input base sequence 624, and a biomass output sequence generator 628. The biomass model 124 then performs a second set of training iterations 667 on the subset of sequences obtained from the second training data set without making any changes to the model configuration of the biomass model 124. After training iteration 666 and training iteration 667, the biomass model 124 undergoes a validation process 686 with a second subset of the second training data set, wherein the first subset and the second subset of the training data set do not overlap 668, which allows the biomass model 124 to be trained on more diverse and unseen training examples.
[0112] Figure 7 is used for variant classification Figure 1 Schematic diagram of a specific implementation 700 of system 100 for comparing a reference sequence 702 and an alternative sequence 704 at base resolution by comparing the corresponding model outputs of the biomass model 124 for each sequence represented in 783. The reference sequence 702 is processed solely by the biomass model 124. An alternative representation generator 722 processes the reference input base sequence 702 to generate an alternative representation (e.g., a convolutional representation sequence), and a biomass output sequence generator 742 processes the alternative representation to generate a plurality of biomass output sequences 762. The alternative sequence 704 is processed solely by the biomass model 124. An alternative representation generator 724 processes the reference input base sequence 704 to generate an alternative representation (e.g., a convolutional representation sequence), and a biomass output sequence generator 744 processes the alternative representation to generate a plurality of biomass output sequences 764. As demonstrated in the process represented by the termination flowchart 783, the plurality of biomass output sequences 762 and the plurality of biomass output sequences 764 are compared at base resolution.
[0113] Figure 8800 is a flow chart of one embodiment of the disclosed technology, wherein the comparison of a reference sequence and an alternative sequence is quantified by a final average incremental value 833. To determine the final average incremental value 833, the base resolution pathogenicity classification logic component 826 is configured to process the differences between the plurality of biomass output sequences 810 predicted for the reference sequence and the alternative sequence. The final average incremental value 833 is taken as the average of the first cumulative average incremental value 822 comparing the reference sequence and the alternative sequence and the second average incremental value 824 comparing the reference sequence and the alternative sequence. The first incremental sequence 812 is generated as a per-base difference between the first reference biomass output sequence 802 predicted from the reference sequence and the first alternative biomass output sequence 804 predicted from the alternative sequence. The second incremental sequence 814 is generated as a per-base difference between the second reference biomass output sequence 806 and the second alternative biomass output sequence 808. The first cumulative average incremental value 822 is taken as the average of the incremental values obtained for each base position within the first incremental sequence 812. The second accumulated average delta value 824 is taken as the average of the delta values obtained for each base position within the second delta sequence 814. Based on the final average delta value 833, the variant represented by the alternative sequence can be classified as a conservative state 846, where the variant can be classified as belonging to a conservative state 842 or a non-conservative state 844.
[0114] Fig. 9 9 is a flow chart of one specific implementation 900 of the disclosed technology, wherein the comparison of a reference sequence and an alternative sequence is quantified by a final sum increment value 933. To determine the final sum increment value 933, a base resolution pathogenicity classification logic component 926 is configured to process the differences between a plurality of biomass output sequences 910 predicted for the reference sequence and the alternative sequence. The final sum increment value 933 is taken as the sum of a first cumulative sum increment value 922 comparing the reference sequence and the alternative sequence and a second sum increment value 924 comparing the reference sequence and the alternative sequence. The first increment sequence 912 is generated as a per-base difference between a first reference biomass output sequence 902 predicted from the reference sequence and a first alternative biomass output sequence 904 predicted from the alternative sequence. The second increment sequence 914 is generated as a per-base difference between a second reference biomass output sequence 906 and a second alternative biomass output sequence 909. The first cumulative sum increment value 922 is taken as the sum of the increment values obtained for each base position within the first increment sequence 912. The second accumulated sum increment value 924 is taken as the sum of the increment values obtained for each base position within the second increment sequence 914. Based on the final sum increment value 933, the variant represented by the alternative sequence can be classified as a conservative state 946, where the variant can be classified as belonging to a conservative state 942 or a non-conservative state 944.
[0115] Fig.10 1 is a flow chart of an implementation 1000 of the disclosed technology, wherein a biomass model 124 generates two biomass output sequences 1084 from an input base sequence 1022 via a first set of weights 1042 and a second set of weights 1062 that are trained end-to-end. The first set of weights 1042 includes an alternative representation generator 1044 that processes the input base sequence 1022 to generate alternative representations of the input base sequence 1022. The second set of weights 1062 includes a biomass output sequence generator 1064 that processes alternative representations of the input base sequence 1022 and generates a plurality of biomass output sequences 1084. In the implementation 1000 of the disclosed technology, two output sequences 1081 and 1083 are generated by the biomass output sequence generator 1064.
[0116] Fig.11 1 is a flow chart of one specific implementation 1100 of the disclosed technology, wherein a biomass model 124 generates three biomass output sequences 1168 from an input base sequence 1102 via a first set of weights 1122 and a second set of weights 1142 that are trained end-to-end. The first set of weights 1122 includes an alternative representation generator 1124 that processes the input base sequence 1102 to generate alternative representations of the input base sequence 1102. The second set of weights 1142 includes a biomass output sequence generator 1144 that processes alternative representations of the input base sequence 1102 and generates a plurality of biomass output sequences 1168.
[0117] Fig.11 The specific implementation of 1100 and Fig.10 The specific implementation 1000 in FIG. 1 differs in that three output sequences 1162 , 1164 , and 1166 are generated by the biomass output sequence generator 1144 , compared to the two output sequences 1081 and 1083 generated by the biomass output sequence generator 1064 . Fig.10 In , the first output sequence 1081 can be a per-base measure of evolutionary conservation. The second output sequence 1083 can be a per-base measure of transcription initiation. Fig.11 In the example of FIG. 1 , the first output sequence 1162 can be a per-base measure of evolutionary conservation. The second output measure 1164 can be a per-base measure of transcription initiation.
[0118] Those skilled in the art will appreciate that multiple output sequences can be predicted simultaneously and that different permutations and combinations of information, such as measures of evolutionary conservation, measures of transcriptional initiation represented, and epigenetic signals per base, can be represented.
[0119] Fig.12is a flow chart of one specific implementation 1200 of the disclosed technology, wherein a biomass model 124 includes a first set of weights 1222 for generating an alternative representation of an input base sequence 1242 and a second set of weights 1226 trained from scratch to generate a plurality of biomass output sequences 1246 from the alternative representation of the input base sequence 1206. Fig.10 The specific implementation in 1000 and Fig.11 Compared to the specific implementation 1100 in , the first set of weights and the second set of weights are not trained end-to-end in the specific implementation 1200. Fig.10 The specific implementation in 1000 and Fig.11 1100, a first set of weights 1222 for implementation 1200 includes an alternative representation generator 1224 that processes an input base sequence 1202 to generate an alternative representation of an input base sequence 1242. However, in contrast to the implementation in which the weights are trained end-to-end, the alternative sequence representation output 1242 from the alternative representation generator 1224 is mapped to an input 1206 of a subsequent model that is configured as a biomass output sequence generator 1228 for training a second set of weights 1226 to generate a plurality of biomass output sequences 1246.
[0120] Fig.13 is a flow chart of a specific implementation 1300 of the disclosed technology, wherein a biomass model 124 generates a biomass output sequence 1362 from an input base sequence 1302 via a first set of weights 1322 and a second set of weights 1342, the first set of weights and the second set of weights being trained end-to-end and retrained to generate subsequent biomass output sequences on a single basis. Fig.13 In the given example shown, the biomass model 124 is trained on the first weight group 1322 and the second weight group 1342 to generate the first output sequence 1362 of the plurality of biomass output sequences 1366. After the first output sequence is generated, the biomass model 124 is retrained end-to-end on the first weight group 1324 and the second weight group 1344 to process the input base sequence 1304 and generate the second output sequence 1364. During the retraining process, the input base sequence 1304 is the same as the input base sequence 1302, the first weight group 1324 is the same as the first weight group 1322, and the second weight group 1344 is the same as the second weight group 1362. However, the second output sequence 1364 is a different biomass output sequence from the first biomass output sequence. For all of the plurality of biomass output sequences 1366, the first weights and the second weights are retrained end-to-end to generate each subsequent biomass output sequence.
[0121] Fig.141400 is a flowchart of a specific implementation of the disclosed technology, wherein the biomass model 124 includes a first set of weights 1422 and a second set of weights 1442 that are trained end-to-end to generate a plurality of biomass output sequences 1462 from an input base sequence, and a third set of weights 1426 and a fourth set of weights 1446 that are trained end-to-end to generate a gene expression output sequence 1466 from a plurality of input biomass output sequences 1406. Generating a biomass output sequence 1426 from an input base sequence 1402 is similar to Fig.10 The specific implementation in 1000 and Fig.11 1400, wherein an input base sequence 1402 is processed by a first set of weights 1422 configured as an alternative representation generator 1404 and a second set of weights 1442 configured as a biomass output sequence generator 1444 to produce a plurality of biomass output sequences 1462. The plurality of output sequences 1462, which are outputs of the second set of weights 1442, are then mapped to an input 1406 of a subsequent model configured to generate a gene expression output sequence 1466. The input biomass output sequence 1406 is processed by a third set of weights 1426 and a fourth set of weights 1446 that are trained end-to-end. The third set of weights 1426 is configured as a biomass alternative representation generator 1428, and the fourth set of weights is configured as a gene expression output generator 1448. The resulting gene expression output sequence 1466 is a measure of base resolution gene expression 1468.
[0122] Fig.15 15 is a flowchart of a specific implementation 1500 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 1522 for generating alternative representations of an input base sequence 1542, a second set of weights 1582 trained from scratch to generate multiple biomass output sequences 1512 from alternative representations of an input base sequence 1562, and a third set of weights 1124 and a fourth set of weights 1546 trained end-to-end to generate gene expression output sequences 1566 from multiple input biomass output sequences 1506. Generating a biomass output sequence 1512 from an input base sequence 1502 is similar to Fig.12 1200 in which the biomass model 124 includes a first set of weights 1522 for generating an alternative representation of an input base sequence 1542 and a second set of weights 1582 trained from scratch to generate multiple biomass output sequences 1512 from the alternative representation of the input base sequence 1562.
[0123] As in Fig.10 Specific implementation of 1000, Fig.11 Specific implementation of 1100, Fig.12 Specific implementation of 1200, Fig.13 The specific implementation of 1300 and Fig.14In the specific implementation 1400 of , the first set of weights 1522 is configured as an alternative representation generator 1524, and the second set of weights 1582 is configured as a biomass output sequence generator 1584. The plurality of output sequences 1512, which are outputs of the second set of weights 1582, are then mapped to the input 1506 of a subsequent model configured to generate a gene expression output sequence 1566. The input biomass output sequence 1506 is processed by the third set of weights 1124 and the fourth set of weights 1546, which are trained end-to-end. The third set of weights 1124 is configured as a biomass alternative representation generator 1528, and the fourth set of weights is configured as a gene expression output generator 1548. The resulting gene expression output sequence 1566 is a measure of base resolution gene expression 1568.
[0124] Fig.16 is a flow chart of one specific implementation 1600 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 1622 for generating alternative representations of an input base sequence 1642, a second set of weights 1682 trained from scratch to generate a plurality of biomass output sequences 1612, a third set of weights 1124 trained from scratch to generate alternative biomass representations 1646 from a plurality of biomass output sequences 1606, and a fourth set of weights 1686 trained from scratch to generate gene expression output sequences 1616 from alternative biomass representations 1666. Generating a biomass output sequence 1612 from an input base sequence 1602 is similar to Fig.16 1600, wherein the biomass model 124 includes a first set of weights 1622 that generates an alternative representation of an input base sequence 1642 and a second set of weights 1682 that is trained from scratch to generate multiple biomass output sequences 1612 from the alternative representation of the input base sequence 1662.
[0125] As in Fig.10 Specific implementation of 1000, Fig.11 Specific implementation of 1100, Fig.12 Specific implementation of 1200, Fig.13 The specific implementation of 1300 and Fig.14 The specific implementation of 1400 and Fig.15In the specific implementation 1500 of , the first set of weights 1622 is configured as an alternative representation generator 1624, and the second set of weights 1682 is configured as a biomass output sequence generator 1684. The plurality of biomass output sequences 1612, which are outputs of the second set of weights 1682, are then mapped to inputs 1606 of a subsequent model configured to generate alternative biomass representations 1646. The subsequent model is configured as a biomass alternative representation generator 1628, which includes a third set of weights 1124 trained from scratch to generate alternative biomass representations 1646. The alternative biomass representation output 1646, which is output of the third set of weights 1124, is then mapped to inputs 1666 of a second subsequent model to generate a gene expression output sequence 1616. The second subsequent model is configured as a gene expression output generator 1688, which includes a fourth set of weights 1686 trained from scratch to generate a gene expression output sequence 1616. The resulting gene expression output sequence 1616 is a measure of base-resolution gene expression 1618 .
[0126] Fig.17 17 is a flow chart of a specific implementation 1700 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 1722 and a second set of weights 1742, which can be trained end-to-end to generate multiple biomass output sequences 1762 from an input base sequence 1702. A third set of weights 1726 can be trained from scratch to generate an alternative biomass representation 1746 from multiple biomass output sequences 1706. A fourth set of weights 1786 can be trained from scratch to generate a gene expression output sequence 1716 from an alternative biomass representation 1766. For a given target base at a given position in the gene expression output sequence 1716, a given per-base gene expression output specifies a measure of the gene expression level of the given target base at a given position. In one specific implementation, the gene expression level is measured per base metric (such as CAGE transcription start site (CTSS)). In another specific 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 specific implementation, gene expression levels are measured in a per-gene metric such as fragments per kilobase million (FPKM).
[0127] The biomass output sequence 1762 generated from the input base sequence 1702 is similar to Fig.10 The specific implementation in 1000 and Fig.11 In the specific implementation 1100, an input base sequence 1702 can be processed by a first set of weights 1722 applied by an alternative representation generator 1724 and a second set of weights 1742 applied by a biomass output sequence generator 1744 to generate a plurality of biomass output sequences 1762.
[0128] The plurality of biomass output sequences 1762 as outputs of the second set of weights 1742 may then be mapped to inputs 1706 of a subsequent model configured to generate alternative biomass representations 1746. The subsequent model may be configured as a biomass alternative representation generator 1728, which includes a third set of weights 1726 trained from scratch to generate alternative biomass representations 1746. The alternative biomass representation outputs 1746 as outputs of the third set of weights 1726 may then be mapped to inputs 1766 of a second subsequent model to generate gene expression output sequences 1716. The second subsequent model may be configured as a gene expression output generator 1788, which includes a fourth set of weights 1786 trained from scratch to generate gene expression output sequences 1716. The resulting gene expression output sequences 1716 are measures of base resolution gene expression 1718.
[0129] Application of transfer learning in biomass modeling
[0130] Fig.18 18 is a flowchart of a specific implementation 1800 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 1812 trained to generate alternative sequence representations 1822 from an input base sequence 1802, then retrained as a replacement for a third set of weights 1842 to generate alternative biomass representations from a plurality of biomass output sequences 1832, and a fourth set of weights 1852 trained end-to-end with the replaced first set of weights to generate gene expression output sequences 1862 from the biomass output sequences 1832. The optimized weight scalar values of the weight set 1812 learned from the alternative representation generator 1813 can be transmitted as replacement scalar values for each weight within the third set of weights 1842. As shown in Fig.14 The specific implementation in 1400 and Fig.15 In the specific implementation 1500, the biomass alternative representation generator 1843 includes a third set of weights 1842 that are trained end-to-end with a fourth set of weights 1852 that includes a gene expression output generator 1863. The resulting gene expression output sequence 1862 is a measure of base resolution gene expression 1863.
[0131] Fig.1919 is a flowchart of a specific implementation 1900 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 1912 trained to generate an alternative sequence representation 1922 from an input base sequence 1902, then retrained as a substitute for a third set of weights 1932 to generate an alternative biomass representation 1942 from the alternative sequence representation 1922, and a fourth set of weights 1962 trained from scratch to generate a gene expression output sequence 1972 from the alternative biomass representation 1952. The optimized weight scalar value of the weight set 1912 learned from the alternative representation generator 1913 can be transmitted as a substitute scalar value for each weight within the third set of weights 1932.
[0132] The alternative biomass representation output 1942, which is the output of the third set of weights 1932, is then mapped to the input 1952 of the subsequent model to generate the gene expression output sequence 1716. Fig.16 The specific implementation of 1600 and Fig.17 In the specific implementation 1700, the subsequent model is configured as a gene expression output generator 1963, which includes a fourth weight group 1962 trained from scratch to generate a gene expression output sequence 1972. The resulting gene expression output sequence 1972 is a measure of base resolution gene expression 1973.
[0133] Fig. 20 is a flow chart of one specific implementation 2000 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 2052 trained to generate alternative sequence representations 2062 from an input base sequence 2042, a second set of weights 2004 trained to generate a plurality of biomass output sequences 2014, a retrained first set of weights 2052 used as a substitute for the third set of weights 2034 to generate alternative biomass representations 2044 from the plurality of biomass output sequences, and a fourth set of weights 2064 trained from scratch to generate gene expression output sequences 2074 from the alternative biomass representations 2054. As shown in Fig.19 In the specific implementation 1900, the optimized weight scalar value of the weight group 1 2052 learned from the alternative representation generator 2053 can be transmitted as the alternative scalar value of each weight within the third set of weights 2034.
[0134] The second set of weights 2004 is configured to generate a plurality of biomass output sequences 2014. Fig.14 Specific implementation of 1400, Fig.15 Specific implementation of 1500, Fig.16 The specific implementation of 1600 and Fig.17In the implementation 1700 of , the biomass output sequence 2014 is mapped to the input of a subsequent model of a biomass alternative representation generator 2035 configured to include a third set of weights 2034. In the implementation 2000, the scalar values of the third set of weights 2034 are replaced from the optimized scalar values from the trained first set of weights 2052. Fig.16 The specific implementation of 1600 and Fig.17 In the specific implementation 1700, the alternative biomass representation output 2044 as the output of the third set of weights 2034 is then mapped to the input 2044 of the subsequent model to generate a gene expression output sequence 2074. The subsequent model is configured as a gene expression output generator 2065, which includes a fourth set of weights 2064 trained from scratch to generate a gene expression output sequence 2074. The resulting gene expression output sequence 2074 is a measure of base resolution gene expression 2075.
[0135] Fig.21 is a flowchart of one specific implementation 2100 of the disclosed technology, wherein the biomass model 124 includes a first set of weights 2142 trained to generate alternative sequence representations 2152 from an input base sequence 2132, a second set of weights 2104 trained to generate multiple biomass output sequences 2114 from an alternative representation of the input bases, a retrained first set of weights 2142 used as a substitute for the third set of weights 2134 to generate alternative biomass representations from multiple biomass output sequences 2124, and a fourth set of weights 2144 trained end-to-end with the replaced first set of weights 2142 to generate gene expression output sequences 2154 from the alternative biomass representations. As shown in Fig.19 The specific implementation of 1900 and Fig. 20 In the specific implementation 2000, the optimized weight scalar value of the weight group 1 2142 learned from the alternative representation generator 2152 can be transmitted as the alternative scalar value of each weight within the third set of weights 2134.
[0136] As in Fig. 20 In the specific implementation 2000, the second set of weights 2104 is configured to generate a plurality of biomass output sequences 2114. Fig.14 Specific implementation of 1400, Fig.15 Specific implementation of 1500, Fig.16 The specific implementation of 1600 and Fig.17 In the specific implementation 1700, the biomass output sequence 2114 is mapped to the input of the subsequent model of the biomass alternative representation generator 2135 which is configured to include the third weight group 2134. Fig. 20 In the specific implementation 2000, the scalar values of the third set of weights 2134 are replaced by the optimized scalar values from the trained first set of weights 2142. Fig.14 Specific implementation in 1400, Fig.15 The specific implementation in 1500 and Fig.18 In the specific implementation 1800, the biomass alternative representation generator 2135 includes a third set of weights 2134 that are trained end-to-end with a fourth set of weights 2144 that includes a gene expression output generator 2145. The resulting gene expression output sequence 2154 is a measure of base resolution gene expression 2155.
[0137] Pathogenicity prediction at base resolution
[0138] Fig. 22 is a flow chart of one specific implementation 2200 of the disclosed technology, wherein the biomass model 124 is further configured to include a pathogenicity prediction logic component that compares a reference biomass output sequence 2253 and an alternative biomass output sequence 2256 at base resolution, wherein a fifth set of weights 2223 generates an alternative sequence pathogenicity prediction 2233 from the plurality of biomass output sequences 2253 and 2256, wherein the first set of weights 2212 and 2214 and the second set of weights 2242 and 2244 are each trained from scratch. As in Fig.12 Specific implementation in 1200, Fig.15 The specific implementation in 1500 and Fig.16 In the implementation 1600 of , the first set of weights 2212 for the implementation 2200 includes an alternative representation generator 2213 that processes a reference input base sequence 2202 to generate an alternative representation 2222 of the reference input base sequence 2202. The alternative sequence representation output 2222 from the alternative representation generator 2213 is mapped to an input 2232 of a subsequent model configured as a biomass output sequence generator 2243 for training a second set of weights 2242 to generate a plurality of biomass output sequences 2253.
[0139] In parallel, a first set of weights 2214 for implementation 2200 includes an alternative representation generator 2216 that processes an alternative input base sequence 2204 to generate an alternative representation 2224 of the alternative input base sequence 2204. The alternative sequence representation output 2224 from the alternative representation generator 2216 is mapped to an input 2234 of a subsequent model configured as a biomass output sequence generator 2246 for training a second set of weights 2244 to generate a plurality of biomass output sequences 2256. The optimized scalar values of the weights for each corresponding biomass model 124 of the reference sequence 2202 and the alternative sequence 2204 can be transmitted to a fifth set of weights 2223 that generates a base-resolution alternative sequence pathogenicity prediction 2233.
[0140] In one implementation, the pathogenicity prediction may be a score between 0 and 1, where 0 represents absolute benignity and 1 represents absolute pathogenicity. In other implementations, a cutoff value may be used, such as a pathogenicity score above 5 may be considered pathogenic and below 5 may be considered benign.
[0141] Fig.23 is a flow chart of one specific implementation 2300 of the disclosed technology, wherein the biomass model 124 is further configured to include a pathogenicity prediction logic component that compares a reference biomass output sequence 2362 and an alternative biomass output sequence 2364 at base resolution, wherein a fifth set of weights 2353 generates an alternative sequence pathogenicity prediction 2363 from the plurality of biomass output sequences 2362 and 2364, wherein the first set of weights 2322 and 2324 and the second set of weights 2342 and 2344 are trained end-to-end. As in Fig.11 Specific implementation in 1100, Fig.12 Specific implementation in 1200, Fig.13 Specific implementation in 1300, Fig.14 The specific implementation in 1400 and Fig.17 In the specific implementation 1700, the reference input base sequence 2302 is processed by a first set of weights 2322 configured as an alternative representation generator 2323 and a second set of weights 2342 configured as a biomass output sequence generator 2343 to generate a plurality of biomass output sequences 2362. The first set of weights 2322 and the second set of weights 2342 are trained end-to-end.
[0142] In parallel, a first set of weights 2324 processes the alternative input base sequence 2304, including an alternative representation generator 2326, and a second set of weights 2344 includes a biomass output sequence generator 2346 to generate a plurality of biomass output sequences 2364 in the same manner as the plurality of biomass output sequences 2362 were generated from the reference input base sequence 2302. The optimized scalar values of the weights for each respective biomass model 124 for the reference sequence 2302 and the alternative sequence 2304 can be transmitted to a fifth set of weights 2353 that generates a base-resolution alternative sequence pathogenicity prediction 2363.
[0143] Biomass Output Sequence Data
[0144] Fig.24is a schematic diagram of an evolutionary conservation measure 2400 as a value for a first biomass output sequence that can be generated by the biomass model 124 from an input base sequence. After the input base sequence is processed by the alternative representation generator 2401, the biomass output sequence generator 2402 can generate the first biomass output sequence in the form of a phyloP score 2403. After the input base sequence is processed by the alternative representation generator 2404, the biomass output sequence generator 2405 can generate the first biomass output sequence in the form of a phastCons value 2406. After the input base sequence is processed by the alternative representation generator 2407, the biomass output sequence generator 2408 can generate the first biomass output sequence in the form of a phastCons value 2409.
[0145] Fig.25 2504 can generate a second biomass output sequence in the form of a Cap Analysis Gene Expression (CAGE) value 2506.
[0146] Fig.26 is a schematic diagram of an epigenetic signal 2600 as a value of a third biomass output sequence that can be generated by the biomass model 124 from an input base sequence. After the input base sequence is processed by the alternative representation generator 2601, the biomass output sequence generator 2602 can generate a first biomass output sequence in the form of a DNase I-hypersensitive site prediction output 2603. After the input base sequence is processed by the alternative representation generator 2604, the biomass output sequence generator 2605 can generate a first biomass output sequence in the form of a transcription factor binding site prediction output 2606. After the input base sequence is processed by the alternative representation generator 2607, the biomass output sequence generator 2608 can generate a first biomass output sequence in the form of a histone modification mark prediction output 2609.
[0147] Gene expression classification model
[0148] Fig. 2727 is a flowchart of a specific implementation of the disclosed technology, wherein an expression change classifier 2700 is configured to predict the effect of a variant on gene expression. A variant expression retention causal score validation dataset 2702 is used to generate ground truth forks of a set of variants for a binary classification model 2704 with a specified decision threshold (e.g., a p-value cutoff less than 0.01, 0.0001, or 1e-14), which can be used to classify variants into a gene expression change class 2724 (i.e., a variant that changes gene expression) or a gene expression retention class 2744 (i.e., a variant that does not change gene expression). The classification of variants is learned from a specified causal score that specifies the statistically unconfounded likelihood of changing gene expression.
[0149] Fig.28 Flowchart of a specific implementation of the disclosed technology, wherein the expression change classifier is further configured as an expression reduction classifier 2800 to predict whether a variant reduces gene expression or does not reduce gene expression. The variant underexpression causal score validation set is used to generate an underexpression ground truth fork of a variant set for a binary classification model 2804 with a specified decision threshold (e.g., a p-value cutoff value is less than 0.01, 0.0001, or 1e-14), which can be used to classify variants into a gene expression reduction class 2824 (i.e., a variant that reduces gene expression) or a gene expression non-reduction class 2844 (i.e., a variant that does not reduce gene expression). The classification of variants is learned from a specified underexpression causal score that specifies the statistically unconfounded possibility of reducing gene expression.
[0150] Fig.29 2900 is a flowchart of a specific implementation of the disclosed technology, wherein the expression change classifier is further configured as an expression increase classifier 2900 to predict whether a variant increases gene expression or does not increase gene expression. The variant overexpression causal score validation dataset 2902 is used to generate overexpression ground truth forks for a set of variants for a binary classification model 2904 with a specified decision threshold (e.g., a p-value cutoff less than 0.01, 0.0001, or 1e-14), which can be used to classify the variants into a gene expression reduction class 2924 or a gene expression non-reduction class 2944. The classification of the variant is learned from the specified overexpression causal score of the statistically unconfounded probability of increasing gene expression.
[0151] Fig.30is a flowchart of one specific implementation of the disclosed technology, wherein the expression change classifier is further configured as a multi-class expression classifier 3000 that predicts whether a variant preserves gene expression, decreases gene expression, or increases gene expression. The variant dataset is processed by system 3000, wherein a first performance metric is generated for comparison of the inferred fork with the ground truth fork by a binary classifier 3026 with a decision threshold, which generates an output corresponding to a gene expression change class 3028 or a gene expression preservation class 3038 from a gene expression change causal score 3024, a second performance metric is generated for comparison of the inferred fork with the ground truth fork by a binary classifier 3046 with a decision threshold, which generates an output corresponding to a gene expression decrease class 3048 or a gene expression non-decrease class 3058 from a gene expression decrease causal score 3044, and a third performance metric is generated for comparison of the inferred fork with the ground truth fork by a binary classifier 3066 with a decision threshold, which generates an output corresponding to a gene expression increase class 3068 or a gene expression non-increase class 3078 from a gene expression increase causal score 3064.
[0152] The system 3000 is configured to require the first inferred fork, the second inferred fork, and the third inferred fork to classify the same number of variants in the genetic alteration class, so that the first performance metric, the second performance metric, and the third performance metric are comparable to each other. The system 3000 is further configured to compare the respective performances of the first binary classifier 3026, the second binary classifier 3046, and the third binary classifier 3066 with decision thresholds on the validation data 3002 based on the comparison of the first performance metric, the second performance metric, and the third performance metric. The decision thresholds of the first binary classifier 3026, the second binary classifier 3046, and the third binary classifier 3066 with decision thresholds can be the same. The decision thresholds of the first binary classifier 3026, the second binary classifier 3046, and the third binary classifier 3066 with decision thresholds can be different.
[0153] The system 3000 is further configured to generate the ground truth triads and inferred triads of the variant set 3002 into a gene expression preservation class 3082, a gene expression reduction class 3084, and a gene expression increase class 3086 by a multi-class classifier 3080, which processes the one-hot encoded vectors containing the ground truth triads and inferred triads from the first binary classifier 3026 with a decision threshold, the second binary classifier 3046, and the third binary classifier 3066.
[0154] Fig.31is a flow chart of one specific implementation of the disclosed technology, wherein gene expression classifier training 3100 is used for comparison of ground truth causal scores with inferred causal scores 3144. A set of ground truth causal scores 3122 is generated for a variant training dataset 3102. A binary classifier 3142 with a decision threshold processes the variant training dataset 3102 to generate inferred causal scores classified into an inferred first class 3161 and an inferred second class 3163. The gene expression classifier training protocol 3100 performs multiple iterations of backpropagation on the weights of the binary classifier 3142 to optimize the loss function.
[0155] Computer Systems
[0156] Fig.32 An example computer system 3200 is shown that can be used to implement the disclosed technology. The computer system 3200 includes at least one central processing unit (CPU) 3272 that communicates with multiple peripheral devices via a bus subsystem 3255. These peripheral devices may include a storage subsystem 3210 (including, for example, a memory device and a file storage subsystem 3232), a user interface input device 3238, a user interface output device 3276, and a network interface subsystem 3274. The input devices and output devices allow a user to interact with the computer system 3200. The network interface subsystem 3274 provides an interface to an external network, including interfaces to corresponding interface devices in other computer systems.
[0157] In one implementation, the biomass model 124 is communicatively linked to the storage subsystem 3210 and the user interface input device 3238 .
[0158] User interface input devices 3238 may include: a keyboard; a pointing device, such as a mouse, trackball, touch pad, or graphics tablet; a scanner; a touch screen incorporated into a display; audio input devices, such as a voice recognition system and a microphone; and other types of input devices. In general, use of the term "input device" is intended to include all possible types of devices and ways of inputting information into the computer system 3200.
[0159] The user interface output device 3276 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 3200 to a user or to another machine or computer system.
[0160] The storage subsystem 3210 stores programming and data structures that provide some or all of the modules and methods described herein. These software modules are typically executed by the processor 3278.
[0161] Processor 3278 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 3278 may be provided by a deep learning cloud platform such as Google Cloud Platform TM 、Xilinx TM and Cirrascale TM ) hosted. Examples of processors 3278 include Google's Tensor Processing Unit (TPU) TM , Rackmount solutions (such as GX4 Rackmount Series TM 、GX32RackmountSeries TM ), NVIDIA DGX-1 TM 、Microsoft's Stratix V FPGA TM 、Graphcore's Intelligent Processor 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.
[0162] The memory subsystem 3222 used in the storage subsystem 3210 may include multiple memories, including a main random access memory (RAM) 3232 for storing instructions and data during program execution and a read-only memory (ROM) 3234 in which fixed instructions are stored. The file storage subsystem 3232 may provide persistent storage for program files and data files, and may include a hard disk drive, a floppy disk drive and associated removable media, a CD-ROM drive, an optical drive, or a removable media tape disk. Modules that implement certain specific implementation functions may be stored in the storage subsystem 3210 by the file storage subsystem 3232, or in other machines accessible to the processor.
[0163] The bus subsystem 3255 provides a mechanism for the various components and subsystems of the computer system 3200 to communicate with each other as intended. Although the bus subsystem 3255 is shown schematically as a single bus, alternative implementations of the bus subsystem may use multiple busses.
[0164] Computer system 3200 itself can be of different 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. As the nature of computers and networks continues to change, the Fig.32 The description of the computer system 3200 depicted in the figure is intended only as a specific example for illustrating a preferred embodiment of the present invention. The computer system 3200 may also have many other configurations in which the components are different from each other. Fig.32 There may be more or fewer computer systems depicted.
[0165] Terms
[0166] 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.
[0167] 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).
[0168] 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.
[0169] 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.
[0170] The present invention discloses the following clauses:
[0171] Article Set 1
[0172] 1. An artificial intelligence-based system for detecting changes in gene expression at base resolution, comprising:
[0173] 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;
[0174] a biomass model that processes the input base sequence and generates alternative representations of the input base sequence; and
[0175] a biomass output generation logic component that processes the alternative representations of the input base sequence and generates a plurality of biomass output sequences,
[0176] wherein a first biomass output sequence in the plurality of biomass output sequences comprises a first corresponding per-base biomass output for a corresponding target base in the target base sequence,
[0177] wherein the first corresponding per-base biomass output specifies a corresponding measure of evolutionary conservation of the corresponding target base across a plurality of species, and wherein a second biomass output sequence in the plurality of biomass output sequences comprises a second corresponding per-base biomass output for the corresponding target base in the target base sequence,
[0178] Wherein the second corresponding biomass per base output specifies a corresponding measurement of transcription initiation of the corresponding target base at a corresponding position in the target base sequence.
[0179] 2. An artificial intelligence-based system according to claim 1, wherein the corresponding 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 a given target base at a given position in the target base sequence is detected as conservation, and an increase in the substitution rate of the given target base at the given position is detected as acceleration.
[0180] 3. An artificial intelligence-based system according to claim 2, wherein the corresponding 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.
[0181] 4. An artificial intelligence-based system according to claim 2, wherein the corresponding measure of evolutionary conservation is a genome evolutionary rate analysis (GERP) score that specifies a decrease in the number of substitutions for the given target base at the given position across the multiple species.
[0182] 5. An artificial intelligence-based system according to claim 1, wherein the corresponding 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.
[0183] 6. The artificial intelligence-based system of clause 1, wherein a third biomass output sequence in the plurality of biomass output sequences comprises a third corresponding per-base biomass output for the corresponding target base in the target base sequence,
[0184] Wherein the third corresponding biomass per base output specifies a corresponding measurement of the epigenetic signal level of the corresponding target base at the corresponding position in the target base sequence.
[0185] 7. An artificial intelligence-based system according to claim 6, wherein the epigenetic signal level specifies DNase I-hypersensitive sites (DHS) or assay for transposase accessible chromatin sequencing (ATAC-Seq).
[0186] 8. An artificial intelligence-based system according to claim 6, wherein the epigenetic signal level specifies transcription factor (TF) binding.
[0187] 9. An artificial intelligence-based system according to claim 6, wherein the epigenetic signal level specifies a histone modification (HM) mark.
[0188] 10. The artificial intelligence based system according to clause 1, further configured to include:
[0189] a gene expression model that processes the plurality of biomass output sequences and generates alternative representations of the plurality of biomass output sequences; and
[0190] a gene expression output generation logic component that processes the alternative representations of the plurality of biomass output sequences and generates a gene expression output sequence of corresponding per-base gene expression outputs for the corresponding target bases in the target base sequence,
[0191] 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.
[0192] 11. An artificial intelligence-based system according to clause 10, wherein the gene expression level is measured as a per-base metric such as CAGE transcription start site (CTSS).
[0193] 12. An artificial intelligence-based system according to clause 10, 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).
[0194] 13. An artificial intelligence-based system according to clause 10, wherein the gene expression level is measured in a per-gene metric such as fragments per kilobase million (FPKM).
[0195] 14. The artificial intelligence based system of clause 1, further configured to include a variant classification logic component.
[0196] 15. An artificial intelligence-based system according to claim 14, 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.
[0197] 16. An artificial intelligence-based system according to claim 15, 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.
[0198] 17. The artificial intelligence-based system of clause 15, wherein the variant classification logic is further configured to include a reference processing logic that causes the biomass model to process the reference base sequence and generate an alternative representation of the reference base sequence, and further causes the biomass output generation logic to process the alternative representation of the reference base sequence and generate a plurality of reference biomass output sequences,
[0199] Each reference biomass output sequence in the plurality of reference biomass output sequences comprises a corresponding per-base reference biomass output for a corresponding reference target base in the reference target base sequence.
[0200] 18. The artificial intelligence-based system of clause 17, wherein a first reference biomass output sequence in the plurality of reference biomass output sequences comprises a first corresponding per-base reference biomass output for the corresponding reference target base in the reference target base sequence,
[0201] wherein the first corresponding per-base reference biomass output specifies a corresponding measure of evolutionary conservation of the corresponding reference target base across the plurality of species.
[0202] 19. The artificial intelligence-based system of clause 17, wherein a second reference biomass output sequence in the plurality of reference biomass output sequences comprises a second corresponding per-base reference biomass output for the corresponding reference target base in the reference target base sequence,
[0203] Wherein the second corresponding per base reference biomass output specifies a corresponding measurement of transcription initiation of the corresponding reference target base at a corresponding position in the reference target base sequence.
[0204] 20. The artificial intelligence-based system of clause 16, wherein the variant classification logic is further configured to include an alternative processing logic that causes the biomass model to process the alternative base sequence and generate an alternative representation of the alternative base sequence, and further causes the biomass output generation logic to process the alternative representation of the alternative base sequence and generate a plurality of alternative biomass output sequences,
[0205] Each alternative biomass output sequence in the plurality of alternative biomass output sequences comprises a corresponding per-base alternative biomass output for a corresponding alternative target base in the alternative target base sequence.
[0206] 21. An artificial intelligence-based system according to clause 20, wherein a first alternative biomass output sequence in the plurality of alternative biomass output sequences comprises a first corresponding per-base alternative biomass output for the corresponding alternative target base in the alternative target base sequence,
[0207] wherein the first corresponding per-base alternative biomass output specifies a corresponding measure of evolutionary conservation of the corresponding alternative target base across the plurality of species.
[0208] 22. An artificial intelligence-based system according to clause 20, wherein a second alternative biomass output sequence in the plurality of alternative biomass output sequences comprises a second corresponding per-base alternative biomass output for the corresponding alternative target base in the alternative target base sequence,
[0209] wherein the second corresponding alternative biomass per base output specifies a corresponding measurement of transcription initiation of the corresponding alternative target base at a corresponding position in the alternative target base sequence.
[0210] 23. 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 first reference biomass output sequence and the first alternative biomass output sequence by position, and generates a first incremental sequence having first position-by-position sequence differences between the positions in the first reference biomass output sequence and the first alternative biomass output sequence.
[0211] 24. An artificial intelligence-based system according to claim 23, wherein the pathogenicity prediction logic component is further configured to compare the second reference biomass output sequence and the second alternative biomass output sequence by position, and generate a second incremental sequence having a second position-by-position sequence difference of positions in the second reference biomass output sequence and the second alternative biomass output sequence.
[0212] 25. The artificial intelligence-based system of clause 24, wherein the pathogenicity prediction logic component is further configured to generate a pathogenicity prediction for the alternative base based on the first incremental sequence and the second incremental sequence.
[0213] 26. An artificial intelligence-based system according to clause 24, wherein the pathogenicity prediction logic component is further configured to accumulate the first position-based sequence difference values into a first cumulative sequence value, and accumulate the second position-based sequence difference values into a second cumulative sequence value.
[0214] 27. An artificial intelligence-based system according to clause 26, wherein the first accumulated sequence value is an average of the first position-wise sequence difference values, and the second accumulated sequence value is an average of the second position-wise sequence difference values.
[0215] 28. An artificial intelligence-based system according to clause 26, wherein the first accumulated sequence value is a sum of the first position-wise sequence difference values and the second accumulated sequence value is a sum of the second position-wise sequence difference values.
[0216] 29. The artificial intelligence-based system of clause 26, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the first accumulated sequence value and the second accumulated sequence value.
[0217] 30. The artificial intelligence-based system of clause 29, wherein the pathogenicity prediction logic is further configured to generate the pathogenicity prediction for the alternative base based on an average of the first cumulative sequence value and the second cumulative sequence value.
[0218] 31. The artificial intelligence-based system of clause 29, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the sum of the first cumulative sequence value and the second cumulative sequence value.
[0219] 32. An artificial intelligence-based system according to claim 23, wherein the pathogenicity prediction logic component is further configured to classify the positions in the first incremental sequence as belonging to a conservative state or a non-conservative state based on the first positional sequence difference.
[0220] 33. An artificial intelligence-based system according to claim 32, wherein the pathogenicity prediction logic component is further configured to classify those positions in the second incremental sequence as belonging to a signal state that overlap with those positions in the first incremental sequence classified as belonging to the conservative state, and to classify those positions in the second incremental sequence as belonging to a noise state that overlap with those positions in the first incremental sequence classified as belonging to the non-conservative state.
[0221] 34. The artificial intelligence-based system of clause 33, wherein the pathogenicity prediction logic is further configured to accumulate a subset of the second position-wise sequence difference values into a modulated cumulative sequence value,
[0222] The second position-sequence difference values in the subset of the second position-sequence difference values are located at those positions in the second increment sequence that are classified as belonging to the signal state.
[0223] 35. An artificial intelligence-based system according to clause 34, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the modulated cumulative sequence value.
[0224] 36. An artificial intelligence-based system as described in clause 34, wherein the modulated accumulated sequence value is an average of the second position-wise sequence difference values in the subset of the second position-wise sequence difference values.
[0225] 37. An artificial intelligence-based system as described in clause 34, wherein the modulated accumulated sequence value is a sum of the second position-wise sequence difference values in the subset of the second position-wise sequence difference values.
[0226] 38. An artificial intelligence-based system according to claim 24, wherein the pathogenicity prediction logic component is further configured to compare the first reference biomass output sequence and the corresponding portions of the first alternative biomass output sequence by position, and generate a first incremental subsequence having a first position-by-position subsequence difference of the positions in the corresponding portions.
[0227] 39. An artificial intelligence-based system according to claim 38, wherein the pathogenicity prediction logic component is further configured to compare the corresponding portions of the second reference biomass output sequence and the second alternative biomass output sequence by position, and generate a second incremental subsequence having a second position-by-position subsequence difference of the positions in the corresponding portions.
[0228] 40. An artificial intelligence-based system according to clause 39, wherein the corresponding portion spans a right-end flanking position and a left-end flanking position around the position under analysis.
[0229] 41. The artificial intelligence-based system of clause 40, wherein the pathogenicity prediction logic component is further configured to generate a pathogenicity prediction for the alternative base based on the first incremental subsequence and the second incremental subsequence.
[0230] 42. An artificial intelligence-based system according to clause 40, wherein the pathogenicity prediction logic component is further configured to accumulate the first position-based subsequence difference values into a first cumulative subsequence value, and accumulate the second position-based subsequence difference values into a second cumulative subsequence value.
[0231] 43. An artificial intelligence-based system according to clause 42, wherein the first accumulated subsequence value is an average of the first position-wise subsequence difference values, and the second accumulated subsequence value is an average of the second position-wise subsequence difference values.
[0232] 44. An artificial intelligence-based system according to clause 42, wherein the first accumulated subsequence value is the sum of the first position-wise subsequence difference values, and the second accumulated subsequence value is the sum of the second position-wise subsequence difference values.
[0233] 45. An artificial intelligence-based system according to clause 42, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the first accumulated subsequence value and the second accumulated subsequence value.
[0234] 46. An artificial intelligence-based system according to clause 45, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on an average of the first cumulative subsequence value and the second cumulative subsequence value.
[0235] 47. An artificial intelligence-based system according to clause 45, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the sum of the first cumulative subsequence value and the second cumulative subsequence value.
[0236] 48. An artificial intelligence-based system according to clause 38, wherein the pathogenicity prediction logic component is further configured to classify positions in the first incremental subsequence as belonging to a conservative state or a non-conservative state based on the first positional subsequence difference.
[0237] 49. An artificial intelligence-based system according to claim 48, wherein the pathogenicity prediction logic component is further configured to classify those positions in the second incremental subsequence as belonging to a signal state that overlaps with those positions in the first incremental subsequence that are classified as belonging to the conservative state, and to classify those positions in the second incremental subsequence as belonging to a noise state that overlaps with those positions in the first incremental subsequence that are classified as belonging to the non-conservative state.
[0238] 50. The artificial intelligence-based system of clause 49, wherein the pathogenicity prediction logic is further configured to accumulate a subset of the second position-wise subsequence difference values into a modulated cumulative subsequence value,
[0239] The second position-wise subsequence difference values in the subset of the second position-wise subsequence difference values are located at those positions in the second increment subsequence that are classified as belonging to the signal state.
[0240] 51. An artificial intelligence-based system according to clause 50, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the modulated cumulative subsequence value.
[0241] 52. An artificial intelligence-based system as described in clause 50, wherein the modulated accumulated subsequence value is an average of the second position-wise subsequence difference values in the subset of the second position-wise subsequence difference values.
[0242] 53. An artificial intelligence-based system as described in clause 50, wherein the modulated accumulated subsequence value is a sum of the second position-wise subsequence difference values in the subset of the second position-wise subsequence difference values.
[0243] 54. An artificial intelligence-based system according to clause 1, wherein the target base sequence is a coding region of a gene.
[0244] 55. An artificial intelligence-based system according to clause 1, wherein the target base sequence is a non-coding region of a gene.
[0245] 56. An artificial intelligence-based system according to clause 55, wherein the non-coding region spans the transcription start site, 5' untranslated region (UTR), 3' UTR, enhancer and promoter.
[0246] 56. An artificial intelligence-based system according to clause 16, wherein the alternative base is a singleton variant that occurs in only one outlier individual among a cohort of outlier individuals,
[0247] The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.
[0248] 57. An artificial intelligence-based system according to clause 56, wherein the extreme gene expression levels are determined from the tail quantiles of the normalized gene expression levels.
[0249] 58. An artificial intelligence-based system according to clause 57, wherein the extreme gene expression levels include gene overexpression and gene underexpression.
[0250] 59. An artificial intelligence-based system according to clause 56, wherein the singleton variant is a coded variant.
[0251] 60. An artificial intelligence-based system according to clause 56, wherein the singleton variant is a non-coding variant.
[0252] 61. An artificial intelligence-based system according to clause 60, wherein the non-coding variant is a promoter variant.
[0253] 62. An artificial intelligence-based system according to clause 60, wherein the non-coding variant is an enhancer variant.
[0254] 62. The artificial intelligence-based system of clause 1, wherein the biomass model has a first set of weights,
[0255] Wherein the biomass output generation logic component has a second set of weights.
[0256] 63. The artificial intelligence-based system of clause 62, wherein during training, the first set of weights of the biomass model is trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence,
[0257] The second set of weights of the biomass output generation logic are trained end-to-end from scratch together with the first set of weights of the biomass model to process the alternative representations of the input base sequence and generate the plurality of biomass output sequences.
[0258] 64. The artificial intelligence-based system of clause 63, wherein during inference, the biomass model uses a trained first set of weights,
[0259] wherein, during said inferring, said biomass output generation logic uses a trained second set of weights.
[0260] 65. An artificial intelligence-based system according to clause 1, wherein the gene expression model has a third set of weights,
[0261] The gene expression output generation logic component has a fourth set of weights.
[0262] 66. The artificial intelligence-based system of clause 65, wherein the third set of weights of the gene expression model is trained from scratch to process the plurality of biomass output sequences and generate the alternative representations of the plurality of biomass output sequences,
[0263] The fourth set of weights of the gene expression output generation logic are trained end-to-end from scratch together with the third set of weights of the gene expression model to process the alternative representations of the plurality of biomass output sequences and generate the gene expression output sequences.
[0264] 67. An artificial intelligence-based system according to clause 66, wherein, during inference, the gene expression model uses a trained third set of weights,
[0265] Wherein, during the inference, the gene expression output generation logic component uses a trained fourth set of weights.
[0266] 68. The artificial intelligence-based system of clause 65, wherein during training, the first set of weights of the biomass model is first trained from scratch 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 plurality of biomass output sequences and generate the alternative representation of the plurality of biomass output sequences,
[0267] wherein the fourth set of weights of the gene expression output generation logic component is end-to-end trained from scratch together with the trained first set of weights replaced in the gene expression model to process the alternative representations of the multiple biomass output sequences generated by the trained first set of weights replaced in the gene expression model and generate the gene expression output sequences.
[0268] 69. The artificial intelligence-based system of clause 68, wherein during inference, the biomass model uses the retrained first set of weights,
[0269] wherein, during said inference, said biomass output generation logic uses said trained second set of weights,
[0270] wherein, during the inference, the gene expression model uses the retrained first set of weights, and
[0271] Wherein, during the inference, the gene expression output generation logic component uses a trained fourth set of weights.
[0272] 70. The artificial intelligence-based system of clause 17, wherein during training, the first set of weights of the biomass model is first trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence,
[0273] wherein during the training, the second set of weights of the biomass output generation logic are first trained end-to-end from scratch together with the first set of weights of the biomass model to process the alternative representations of the input base sequence and generate the plurality of biomass output sequences,
[0274] wherein, during the training, the trained first set of weights of the biomass model is then retrained to process the reference base sequence and generate the alternative representation of the reference base sequence, and to process the alternative base sequence and generate the alternative representation of the alternative base sequence, and
[0275] wherein, during said training, said trained second set of weights of said biomass output generation logic are then retrained end-to-end with said trained first set of weights of said biomass model to process said alternative representations of said reference base sequence and generate said plurality of reference biomass output sequences, and
[0276] The alternative representations of the alternative base sequences are processed and the plurality of alternative biomass output sequences are generated.
[0277] 71. The artificial intelligence-based system of clause 70, wherein during inference, the biomass model uses the retrained first set of weights,
[0278] wherein, during the inference, the biomass output generation logic uses the retrained second set of weights.
[0279] 72. The artificial intelligence-based system of clause 23, wherein the pathogenicity prediction logic component has a fifth set of weights.
[0280] 73. The artificial intelligence-based system of clause 72, wherein during training, the first set of weights of the biomass model is first trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence,
[0281] wherein, during the training, the second set of weights of the biomass output generation logic are first trained end-to-end from scratch together with the first set of weights of the biomass model to process the alternative representations of the input base sequence and generate the plurality of biomass output sequences, and
[0282] wherein, during the training, the trained first set of weights of the biomass model and the trained second set of weights of the biomass output generation logic are then retrained end-to-end to generate the pathogenicity predictions for the alternative bases.
[0283] 74. The artificial intelligence-based system of clause 73, wherein during inference, the biomass model uses the retrained first set of weights,
[0284] wherein, during said inference, said biomass output generation logic uses said retrained second set of weights, and
[0285] Wherein, during the inference, the pathogenicity prediction logic component uses a trained fifth set of weights.
[0286] 75. An artificial intelligence-based system according to claim 18, wherein the first corresponding per-base reference biomass output specifies a corresponding measurement of a first reference epigenetic signal level of the corresponding reference target base at the corresponding position in the reference target base sequence.
[0287] 76. An artificial intelligence-based system according to claim 29, wherein the second corresponding per-base reference biomass output specifies a corresponding measurement of a second reference epigenetic signal level of the corresponding reference target base at the corresponding position in the reference target base sequence.
[0288] 77. An artificial intelligence-based system according to claim 21, wherein the first corresponding alternative biomass per base output specifies a corresponding measurement of a first alternative epigenetic signal level of the corresponding alternative target base at the corresponding position in the alternative target base sequence.
[0289] 78. An artificial intelligence-based system according to claim 22, wherein the second corresponding alternative biomass per base output specifies a corresponding measurement of a second alternative epigenetic signal level of the corresponding alternative target base at the corresponding position in the alternative target base sequence.
[0290] 79. An artificial intelligence-based system according to claim 1, wherein during training, the biomass model and the biomass 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 evolutionarily conserved 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.
[0291] 80. An artificial intelligence-based system according to claim 1, wherein, during training, the biomass model and the biomass 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 epigenetic 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.
[0292] 81. An artificial intelligence-based system according to claim 1, wherein during training, the biomass model and the biomass 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 epigenetic 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.
[0293] 82. An artificial intelligence-based system according to claim 1, wherein, during training, the biomass model and the biomass 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 epigenetic 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.
[0294] 83. The artificial intelligence-based system of 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 multiple epigenetic effects.
[0295] 84. An artificial intelligence-based system according to claim 83, wherein the epigenetic effects among the multiple epigenetic 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.
[0296] 85. The artificial intelligence-based system of clause 83, 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 multiple epigenetic effects.
[0297] 86. An artificial intelligence-based system according to clause 85, wherein the variants in the second training set are reliably determined to alter gene expression and cause extreme gene expression levels.
[0298] 87. An artificial intelligence-based system according to clause 86, wherein the variants in the second training set include overexpression-causing variants that increase gene expression levels.
[0299] 88. An artificial intelligence-based system according to clause 86, wherein the variants in the second training set include under-expression-causing variants that reduce gene expression levels.
[0300] 89. An artificial intelligence-based system according to clause 87, 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.
[0301] 90. An artificial intelligence-based system according to clause 88, 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.
[0302] 91. An artificial intelligence-based system according to clause 86, wherein each variant in the second training set is a singleton variant that occurs in only one outlier individual among the outlier individual cohort,
[0303] The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.
[0304] 92. An artificial intelligence-based system according to clause 91, wherein the extreme gene expression levels are determined from the tail quantiles of the normalized gene expression levels.
[0305] 93. An artificial intelligence-based system according to clause 91, wherein the extreme gene expression levels include gene overexpression and gene underexpression.
[0306] 94. An artificial intelligence-based system according to clause 91, wherein the singleton variant is a coded variant.
[0307] 95. An artificial intelligence-based system according to clause 91, wherein the singleton variant is a non-coding variant.
[0308] 96. An artificial intelligence-based system according to clause 95, wherein the non-coding variant is a 5' untranslated region (UTR) variant, a 3'UTR variant, an enhancer variant or a promoter variant.
[0309] 97. An artificial intelligence-based system according to clause 85, wherein the variants in the second training set span multiple tissue types.
[0310] 98. An artificial intelligence-based system according to clause 85, wherein the variants in the second training set span multiple cell types.
[0311] 99. An artificial intelligence-based system according to claim 1, wherein the input base sequence and the multiple biomass output sequences span the multiple tissue types.
[0312] 100. An artificial intelligence-based system according to claim 1, wherein the input base sequence and the multiple biomass output sequences span the multiple cell types.
[0313] 101. An artificial intelligence-based system according to clause 10, wherein the gene expression output sequences span the multiple tissue types.
[0314] 102. An artificial intelligence-based system according to clause 10, wherein the gene expression output sequences span the multiple cell types.
[0315] 103. The artificial intelligence based system of clause 1, wherein the biomass model and the
[0316] The biomass output generation logic component is first trained end-to-end on the first training set and then retrained on the second training set.
[0317] 104. 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.
[0318] 105. The artificial intelligence-based system of clause 104, wherein the benign set is in three cores
[0319] The sequences were balanced with respect to nucleotide background, homopolymers, k-mers, GC frequency, and sequencing depth.
[0320] 106. An artificial intelligence based system according to clause 104, wherein based on the application of
[0321] The overexpression probability and the underexpression probability are cutoff probabilities, and the variants in the second training set are divided into an overexpression variant training set with a first ground truth label indicating an increase in gene expression, an underexpression variant training set with a second ground truth label indicating a decrease in gene expression, and a neural expression variant training set indicating maintenance of gene expression.
[0322] 107. An artificial intelligence-based system according to claim 10, 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.
[0323] 108. An artificial intelligence-based system according to claim 1, wherein the biomass model and the biomass 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.
[0324] 109. An artificial intelligence-based system according to claim 10, 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.
[0325] 110. An artificial intelligence-based system according to claim 85, 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 biomass model 124, the trained biomass output generation logic component, the trained gene expression model and the trained gene expression output generation logic component.
[0326] 111. An artificial intelligence-based system according to clause 110, wherein those variants in the second training set that appear on even chromosomes are used as the validation set.
[0327] 112. An artificial intelligence-based system according to clause 1, wherein the size of the target base sequence is
[0328] The size varies during training to account for changes in the offset position of the transcription start site (TSS).
[0329] 113. An artificial intelligence-based system for detecting changes in gene expression at base resolution, comprising:
[0330] 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,
[0331] and a left-end base sequence with upstream background bases;
[0332] a biomass model that processes the input base sequence and generates alternative representations of the input base sequence; and
[0333] a biomass output generation logic component that processes the alternative representations of the input base sequence and generates a plurality of biomass output sequences,
[0334] Each of the plurality of biomass output sequences comprises a corresponding per-base biomass output for a corresponding target base in the target base sequence.
[0335] 114. The artificial intelligence-based system of clause 113, wherein the plurality of biomass inputs
[0336] The first biomass output sequence in the output sequence includes a first corresponding per-base biomass output for the corresponding target base in the target base sequence, and
[0337] Wherein the first corresponding per-base biomass output specifies a corresponding measure of evolutionary conservation of the corresponding target base across multiple species.
[0338] 115. The artificial intelligence-based system of clause 113, wherein the plurality of biomass inputs
[0339] The second biomass output sequence in the output sequence includes a second corresponding per-base biomass output for the corresponding target base in the target base sequence, and
[0340] Wherein the second corresponding biomass per base output specifies a corresponding measurement of transcription initiation of the corresponding target base at a corresponding position in the target base sequence.
[0341] Article Set 2
[0342] 1. A system, comprising:
[0343] validation data with sets of variants and sets of causal scores, where the causal score specifies the statistically unconfounded likelihood of altering gene expression;
[0344] a validation set discretization logic component configured to classify each causal score in the set of causal scores into a gene expression altered class or a gene expression retained class based on applying a cutoff value to the set of causal scores, thereby generating a ground truth bifurcation of the set of variants into the gene expression altered class and the gene expression retained class;
[0345] an inference logic component configured to cause a model to generate a set of prediction scores for the set of variants, wherein a prediction score in the set of prediction scores specifies an inferred likelihood of altering gene expression, and wherein the model is trained to determine gene expression variability of variants;
[0346] a model score discretization logic component configured to classify each prediction score in the set of prediction scores into the gene expression altered class or the gene expression retained class based on applying a threshold to the set of prediction scores, thereby generating an inferred bifurcation of the set of variants into the gene expression altered class and the gene expression retained class; and
[0347] A validation logic component is configured to determine a performance measure of the model based on a comparison of the inferred bifurcation to the ground truth bifurcation.
[0348] 2. A system according to claim 1, wherein the ground truth fork assigns a first label (e.g., 0) to those variants classified as the gene expression alteration class and assigns a second label (e.g., 1) to those variants classified as the gene expression preservation class.
[0349] 3. A system according to claim 2, wherein the ground truth fork assigns a first label (e.g., 0) to those variants classified as the gene expression alteration class and assigns a second label (e.g., 1) to those variants classified as the gene expression preservation class.
[0350] In other implementations, the ground truth fork forks the variant into three categories: -1, 0, 1, corresponding to reduced gene expression, no change in gene expression, and increased gene expression. That is, in such implementations, a single classifier can perform 3-way classification.
[0351] 4. The system of clause 3, further configured to encode the ground truth bifurcation in a first vector and encode the inferred bifurcation in a second vector.
[0352] 5. The system of clause 4, further configured to determine the performance metric of the model based on a component-by-component comparison of the first vector and the second vector.
[0353] 6. The system according to clause 1 is further configured to determine the performance measure of the model based on the number of variants classified as the gene expression change class by the inferred bifurcation, the number of variants classified as the gene expression retention class by the inferred bifurcation, the number of variants classified as the gene expression change class by the ground truth bifurcation, and the odds ratio of the number of variants classified as the gene expression retention class by the ground truth bifurcation.
[0354] 7. A system according to clause 1, wherein the set of variants has a set of underexpression causal scores, wherein the underexpression causal score specifies a statistically unconfounded likelihood of reducing gene expression.
[0355] 8. A system according to clause 7, wherein the validation set discretization logic component is further configured to classify each underexpression causal score in the underexpression causal score set into a gene expression reduced class or a gene expression non-reduced class based on applying an underexpression cutoff value to the underexpression causal score set, thereby generating an underexpression ground truth bifurcation of the variant set into the gene expression reduced class and the gene expression non-reduced class.
[0356] 9. The system of clause 8, wherein the inference logic component is further configured to cause the model to generate a set of underexpression prediction scores for the set of variants, wherein the underexpression prediction scores in the set of underexpression prediction scores specify an inferred likelihood of reducing gene expression, and wherein the model is trained to determine the gene expression reducibility of variants,
[0357] wherein the model score discretization logic component is further configured to classify each underexpression prediction score in the underexpression prediction score set into the gene expression reduction class or the gene expression non-reduction class based on applying an underexpression threshold to the underexpression prediction score set, thereby generating a bifurcation of underexpression inference of the variant set into the gene expression reduction class and the gene expression non-reduction class, and
[0358] Wherein the validation logic component is further configured to determine an underexpression performance measure of the model based on a comparison of the underexpression inferred fork with the underexpression ground truth fork.
[0359] 10. A system according to clause 9, wherein the model score discretization logic component is further configured to sort the set of underexpression prediction scores in descending order to classify a subset of the N lowest underexpression prediction scores in the sorted set of underexpression prediction scores as the gene expression reduced class, and to classify a subset of the remaining underexpression prediction scores in the sorted set of underexpression prediction scores as the gene expression not reduced class.
[0360] 11. A system according to claim 8, wherein the underexpression ground truth fork assigns a first label (e.g., 0) to those variants classified as the reduced gene expression class and assigns a second label (e.g., 1) to those variants classified as the non-reduced gene expression class.
[0361] 12. A system according to claim 11, wherein the underexpression inference bifurcation assigns the first label (e.g., 0) to those variants classified as the gene expression reduction class and assigns the second label (e.g., 1) to those variants classified as the gene expression non-reduced class.
[0362] 13. The system of clause 12, further configured to encode the under-expressed ground truth bifurcations in a first vector and to encode the inferred bifurcations in a second vector.
[0363] 14. The system of clause 13, further configured to determine the under-expression performance metric of the model based on a component-by-component comparison of the first vector and the second vector.
[0364] 15. The system according to clause 9 is further configured to determine the underexpression performance measure of the model based on the number of variants classified as the reduced gene expression class by the fork inferred by the underexpression, the number of variants classified as the non-reduced gene expression class by the fork inferred by the underexpression, the number of variants classified as the reduced gene expression class by the fork of the underexpression ground truth, and the odds ratio of the number of variants classified as the non-reduced gene expression class by the fork of the underexpression ground truth.
[0365] 16. A system according to clause 1, wherein the set of variants has a set of overexpression causal scores, wherein the overexpression causal score specifies a statistically unconfounded likelihood of increasing gene expression.
[0366] 17. A system according to clause 16, wherein the validation set discretization logic component is further configured to classify each overexpression causal score in the overexpression causal score set into a gene expression increase class or a gene expression non-increase class based on applying an overexpression cutoff value to the overexpression causal score set, thereby generating an overexpression ground truth bifurcation of the variant set into the gene expression increase class and the gene expression non-increase class.
[0367] 18. The system of clause 17, wherein the inference logic component is further configured to cause the model to generate a set of overexpression prediction scores for the set of variants, wherein the overexpression prediction scores in the set of overexpression prediction scores specify an inferred likelihood of increasing gene expression, and wherein the model is trained to determine the gene expression increaseability of variants,
[0368] wherein the model score discretization logic component is further configured to classify each overexpression prediction score in the set of overexpression prediction scores into the gene expression increased class or the gene expression non-increased class based on applying an overexpression threshold to the set of overexpression prediction scores, thereby generating a bifurcation of overexpression inference of the variant set into the gene expression increased class and the gene expression non-increased class, and
[0369] Wherein the validation logic component is further configured to determine an overexpression performance measure of the model based on a comparison of the overexpression inferred fork with the overexpression ground truth fork.
[0370] 19. A system according to clause 18, wherein the model score discretization logic component is further configured to sort the set of overexpression prediction scores in descending order to classify a subset of N highest overexpression prediction scores in the sorted set of overexpression prediction scores as the gene expression increased class, and to classify a subset of the remaining overexpression prediction scores in the sorted set of overexpression prediction scores as the gene expression not increased class.
[0371] 20. A system according to claim 17, wherein the overexpression ground truth fork assigns a first label (e.g., 0) to those variants classified as the gene expression increase class and assigns a second label (e.g., 1) to those variants classified as the gene expression non-increase class.
[0372] 21. A system according to claim 20, wherein the overexpression inferred bifurcation assigns the first label (e.g., 0) to those variants classified as the gene expression increase class and assigns the second label (e.g., 1) to those variants classified as the gene expression non-increase class.
[0373] 22. The system of clause 21, further configured to encode the over-expressed ground truth bifurcation in a first vector and to encode the inferred bifurcation in a second vector.
[0374] 23. The system of clause 22, further configured to determine the over-expression performance metric of the model based on an element-by-element comparison of the first vector and the second vector.
[0375] 24. The system according to clause 18 is further configured to determine the overexpression performance measure of the model based on the odds ratio of the number of variants classified as the gene expression increase class by the fork inferred by the overexpression, the number of variants classified as the gene expression non-increase class by the fork inferred by the overexpression, the number of variants classified as the gene expression increase class by the fork inferred by the overexpression, and the number of variants classified as the gene expression non-increase class by the fork of the overexpression ground truth.
[0376] 25. The system of clause 1, wherein the inference logic component is further configured to cause a first model to generate a first set of prediction scores for the set of variants, wherein the prediction scores in the first set of prediction scores specify an inferred likelihood of altered gene expression, and wherein the first model is trained to determine gene expression variability of variants,
[0377] wherein the model score discretization logic component is further configured to classify each prediction score in the first set of prediction scores into the gene expression altered class or the gene expression preserved class based on applying a first threshold to the first set of prediction scores, thereby generating a first inferred bifurcation of the variant set into the gene expression altered class and the gene expression preserved class, and
[0378] Wherein the validation logic component is further configured to determine a first performance metric of the first model based on a comparison of the first inferred bifurcation with the ground truth bifurcation.
[0379] 26. The system of clause 25, wherein the inference logic component is further configured to cause a second model to generate a second set of prediction scores for the set of variants, wherein the prediction scores in the second set of prediction scores specify an inferred likelihood of altered gene expression, and wherein the second model is trained to determine gene expression variability of variants,
[0380] wherein the model score discretization logic component is further configured to classify each prediction score in the second set of prediction scores into the gene expression altered class or the gene expression retained class based on applying a second threshold to the second set of prediction scores, thereby generating a second inferred bifurcation of the variant set into the gene expression altered class and the gene expression retained class, and wherein the validation logic component is further configured to determine a second performance measure of the second model based on a comparison of the second inferred bifurcation with the ground truth bifurcation.
[0381] 27. The system of clause 26, wherein the inference logic component is further configured to cause a third model to generate a third set of prediction scores for the set of variants, wherein the prediction scores in the third set of prediction scores specify an inferred likelihood of altered gene expression, and wherein the third model is trained to determine gene expression variability of variants,
[0382] wherein the model score discretization logic component is further configured to classify each prediction score in the third set of prediction scores into the gene expression altered class or the gene expression retained class based on applying a third threshold to the third set of prediction scores, thereby generating a third inferred bifurcation of the variant set into the gene expression altered class and the gene expression retained class,
[0383] And wherein the validation logic component is further configured to determine a third performance metric of the third model based on a comparison of the third inferred bifurcation with the ground truth bifurcation.
[0384] 28. The system according to clause 27 is further configured to require that the first inferred fork, the second inferred fork, and the third inferred fork classify the same number of variants in the gene expression change class, so that the first performance metric, the second performance metric, and the third performance metric are comparable to each other.
[0385] 29. The system of clause 28, further configured to compare the respective performances of the first model, the second model, and the third model on the validation data based on a comparison of the first performance metric, the second performance metric, and the third performance metric.
[0386] 30. The system of clause 27, wherein the first threshold, the second threshold, and the third threshold are different.
[0387] 31. The system of clause 27, wherein at least some of the first threshold, the second threshold, and the third threshold are the same.
[0388] 32. The system of clause 1, further configured to generate corresponding ground truth bifurcations of the set of variants into the gene expression altered class and the gene expression preserved class based on corresponding application of different cutoff values to the set of causal scores.
[0389] 33. The system according to clause 1 is further configured to generate a ground truth trinity of the variant set into a gene expression reduction class, a gene expression increase class, and a gene expression preservation class.
[0390] 34. The system of clause 33, further configured to generate an inferred trinity of the set of variants into the gene expression decreased class, the gene expression increased class, and the gene expression preserved class.
[0391] 35. The system of clause 1, wherein each variant in the set of variants is a singleton variant that occurs in only one outlier individual among a cohort of outlier individuals,
[0392] The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.
[0393] 36. A system according to clause 35, wherein the extreme gene expression levels are determined from the tail quantiles of the normalized gene expression levels.
[0394] 37. A system according to clause 35, wherein the extreme gene expression levels include gene overexpression and gene underexpression.
[0395] 38. A system according to clause 35, wherein the singleton variant is a coding variant.
[0396] 39. A system according to clause 35, wherein the singleton variant is a non-coding variant.
[0397] 40. The system according to clause 39, wherein the non-coding variant is a 5' untranslated region (UTR) variant, a 3' UTR variant, an enhancer variant or a promoter variant.
[0398] 41. A system according to clause 1, wherein the set of variants spans multiple tissue types.
[0399] 42. A system according to claim 1, wherein the set of variants spans multiple cell types.
[0400] 43. A system, comprising:
[0401] Validation data, which has a set of observations and a set of ground truth scores;
[0402] a validation set discretization logic component configured to classify each ground truth score in the set of ground truth scores into a first class or a second class based on applying a cutoff value to the set of ground truth scores, thereby generating a ground truth bifurcation of the observation set into the first class and the second class;
[0403] an inference logic component configured to cause the trained model to generate a set of prediction scores for the set of observations;
[0404] a model score discretization logic component configured to classify each prediction score in the set of prediction scores into the first class or the second class based on applying a threshold to the set of prediction scores, thereby generating a bifurcation of inference of the set of observations into the first class and the second class; and
[0405] A validation logic component is configured to determine a performance measure of the trained model based on a comparison of the inferred bifurcation to the ground truth bifurcation.
[0406] 44. The system according to clause 43 is further configured to generate corresponding inferred forks for corresponding models, so that each corresponding inferred fork in the corresponding inferred forks is required to classify the same number of observations into the first category, so that the corresponding performance measures of the corresponding models are comparable to each other.
[0407] 45. The system of clause 44, further configured to compare the respective performances of the respective models on the validation data based on the comparison of the respective performance metrics.
[0408] 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. An artificial intelligence-based system for detecting changes in gene expression at base resolution, comprising: 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; a biomass model that processes the input base sequence and generates alternative representations of the input base sequence; as well as a biomass output generation logic component that processes the alternative representations of the input base sequence and generates a plurality of biomass output sequences, wherein a first biomass output sequence in the plurality of biomass output sequences comprises a first corresponding per-base biomass output for a corresponding target base in the target base sequence, wherein the first corresponding per-base biomass output specifies a corresponding measure of evolutionary conservation of the corresponding target base across multiple species, and wherein a second biomass output sequence in the plurality of biomass output sequences comprises a second corresponding per-base biomass output for the corresponding target base in the target base sequence, Wherein the second corresponding biomass per base output specifies a corresponding measurement of transcription initiation of the corresponding target base at a corresponding position in the target base sequence.
2. An artificial intelligence-based system according to claim 1, wherein the corresponding measure of evolutionary conservation is a phylogenetic P-value (phyloP) score, which specifies the deviation from the neural substitution null model so that a decrease in the substitution rate of a given target base at a given position in the target base sequence is detected as conservation, and an increase in the substitution rate of the given target base at the given position is detected as acceleration.
3. An artificial intelligence-based system according to claim 2, wherein the corresponding 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.
4. An artificial intelligence-based system according to claim 2, wherein the corresponding measure of evolutionary conservation is a genome evolutionary rate analysis (GERP) score, which specifies the reduction in the number of substitutions of the given target base at the given position across the multiple species.
5. The artificial intelligence-based system of claim 1 , wherein the corresponding 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.
6. The artificial intelligence-based system of claim 1 , wherein a third biomass output sequence in the plurality of biomass output sequences comprises a third corresponding per-base biomass output for the corresponding target base in the target base sequence, Wherein the third corresponding biomass per base output specifies a corresponding measurement of the epigenetic signal level of the corresponding target base at the corresponding position in the target base sequence.
7. The artificial intelligence-based system of claim 6, wherein the epigenetic signal level specifies DNase I-hypersensitive sites (DHS) or assay for transposase-accessible chromatin sequencing (ATAC-Seq).
8. The artificial intelligence-based system of claim 6, wherein the epigenetic signal level specifies transcription factor (TF) binding.
9. The artificial intelligence-based system of claim 6, wherein the epigenetic signal level specifies a histone modification (HM) mark.
10. The artificial intelligence based system according to claim 1, further configured to include: a gene expression model that processes the plurality of biomass output sequences and generates alternative representations of the plurality of biomass output sequences; as well as a gene expression output generation logic component that processes the alternative representations of the plurality of biomass output sequences and generates a gene expression output sequence of corresponding per-base gene expression outputs for the corresponding target bases in the target base sequence, 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.
11. An artificial intelligence based system according to claim 10, wherein the gene expression level is measured in a per base metric such as CAGE transcription start site (CTSS).
12. The artificial intelligence-based system of claim 10, wherein the gene expression levels are measured in a per-gene metric such as transcripts per million (TPM) or reads per kilobase of transcript (RPKM).
13. The artificial intelligence based system of claim 10, wherein the gene expression level is measured in a per gene metric such as fragments per kilobase million (FPKM).
14. The artificial intelligence based system of claim 1, further configured to include a variant classification logic component.
15. The artificial intelligence-based system of claim 14, wherein the variant classification logic component is further configured to include a reference input generation logic component that 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.
16. An artificial intelligence-based system according to claim 15, 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 the 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.
17. The artificial intelligence-based system of claim 15, wherein the variant classification logic is further configured to include a reference processing logic that causes the biomass model to process the reference base sequence and generate an alternative representation of the reference base sequence, and further causes the biomass output generation logic to process the alternative representation of the reference base sequence and generate a plurality of reference biomass output sequences, Each reference biomass output sequence in the plurality of reference biomass output sequences comprises a corresponding per-base reference biomass output for a corresponding reference target base in the reference target base sequence.
18. The artificial intelligence-based system of claim 17, wherein a first reference biomass output sequence in the plurality of reference biomass output sequences comprises a first corresponding per-base reference biomass output for the corresponding reference target base in the reference target base sequence, wherein the first corresponding per-base reference biomass output specifies a corresponding measure of evolutionary conservation of the corresponding reference target base across the plurality of species.
19. The artificial intelligence-based system of claim 17, wherein a second reference biomass output sequence among the plurality of reference biomass output sequences comprises a second corresponding per-base reference biomass output for the corresponding reference target base in the reference target base sequence, wherein the second corresponding per-base reference biomass output specifies a corresponding measurement of transcription start of the corresponding reference target base at a corresponding position in the reference target base sequence.
20. The artificial intelligence-based system of claim 16, wherein the variant classification logic is further configured to include an alternative processing logic that causes the biomass model to process the alternative base sequence and generate an alternative representation of the alternative base sequence, and further causes the biomass output generation logic to process the alternative representation of the alternative base sequence and generate a plurality of alternative biomass output sequences, Each alternative biomass output sequence in the plurality of alternative biomass output sequences comprises a corresponding per-base alternative biomass output for a corresponding alternative target base in the alternative target base sequence.
21. The artificial intelligence-based system of claim 20, wherein a first alternative biomass output sequence of the plurality of alternative biomass output sequences comprises a first corresponding per-base alternative biomass output for the corresponding alternative target base in the alternative target base sequence, wherein the first corresponding per-base alternative biomass output specifies a corresponding measure of evolutionary conservation of the corresponding alternative target base across the plurality of species.
22. The artificial intelligence-based system of claim 20, wherein a second alternative biomass output sequence of the plurality of alternative biomass output sequences comprises a second corresponding per-base alternative biomass output for the corresponding alternative target base in the alternative target base sequence, wherein the second corresponding per-base alternative biomass output specifies a corresponding measurement of transcription start of the corresponding alternative target base at a corresponding position in the alternative target base sequence.
23. The artificial intelligence-based system of claim 20, wherein the variant classification logic is further configured to include a pathogenicity prediction logic component that compares the first reference biomass output sequence and the first alternative biomass output sequence by position and generates a first incremental sequence having first position-by-position sequence differences of positions in the first reference biomass output sequence and the first alternative biomass output sequence.
24. The artificial intelligence-based system of claim 23, wherein the pathogenicity prediction logic component is further configured to compare the second reference biomass output sequence and the second alternative biomass output sequence by position, and generate a second incremental sequence having second position-by-position sequence differences of positions in the second reference biomass output sequence and the second alternative biomass output sequence.
25. The artificial intelligence-based system of claim 24, wherein the pathogenicity prediction logic component is further configured to generate a pathogenicity prediction for the alternative base based on the first incremental sequence and the second incremental sequence.
26. The artificial intelligence-based system of claim 24, wherein the pathogenicity prediction logic component is further configured to accumulate the first position-by-position sequence difference values into a first cumulative sequence value, and accumulate the second position-by-position sequence difference values into a second cumulative sequence value.
27. The artificial intelligence-based system of claim 26, wherein the first accumulated sequence value is an average of the first position-wise sequence difference values, and the second accumulated sequence value is an average of the second position-wise sequence difference values.
28. The artificial intelligence based system of claim 26, wherein the first accumulated sequence value is a sum of the first position-wise sequence difference values, and the second accumulated sequence value is a sum of the second position-wise sequence difference values.
29. The artificial intelligence-based system of claim 26, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the first accumulated sequence value and the second accumulated sequence value.
30. The artificial intelligence-based system of claim 29, wherein the pathogenicity prediction logic is further configured to generate the pathogenicity prediction for the alternative base based on an average of the first cumulative sequence value and the second cumulative sequence value.
31. The artificial intelligence-based system of claim 29, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the sum of the first cumulative sequence value and the second cumulative sequence value.
32. The artificial intelligence-based system of claim 23, wherein the pathogenicity prediction logic component is further configured to classify positions in the first incremental sequence as belonging to a conservative state or a non-conservative state based on the first positional sequence difference.
33. An artificial intelligence-based system according to claim 32, wherein the pathogenicity prediction logic component is further configured to classify those positions in the second incremental sequence as belonging to the signal state that coincide with those positions in the first incremental sequence classified as belonging to the conservative state, and classify those positions in the second incremental sequence as belonging to the noise state that coincide with those positions in the first incremental sequence classified as belonging to the non-conservative state.
34. The artificial intelligence-based system of claim 33, wherein the pathogenicity prediction logic is further configured to accumulate a subset of the second position-wise sequence difference values into a modulated cumulative sequence value, The second position-sequence difference values in the subset of the second position-sequence difference values are located at those positions in the second increment sequence that are classified as belonging to the signal state.
35. The artificial intelligence-based system of claim 34, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the modulated cumulative sequence value.
36. The artificial intelligence based system of claim 34, wherein the modulated accumulated sequence value is an average of the second position-wise sequence difference values in the subset of the second position-wise sequence difference values.
37. The artificial intelligence based system of claim 34, wherein the modulated accumulated sequence value is a sum of the second position-wise sequence difference values in the subset of the second position-wise sequence difference values.
38. The artificial intelligence-based system of claim 24, wherein the pathogenicity prediction logic component is further configured to compare corresponding portions of the first reference biomass output sequence and the first alternative biomass output sequence by position, and generate a first incremental subsequence having first positional subsequence differences of positions in the corresponding portions.
39. The artificial intelligence-based system of claim 38, wherein the pathogenicity prediction logic component is further configured to compare corresponding portions of the second reference biomass output sequence and the second alternative biomass output sequence by position, and generate a second incremental subsequence having second positional subsequence differences of positions in the corresponding portions.
40. The artificial intelligence-based system of claim 39, wherein the corresponding portion spans a right-end flanking position and a left-end flanking position around the position under analysis.
41. The artificial intelligence-based system of claim 40, wherein the pathogenicity prediction logic component is further configured to generate a pathogenicity prediction for the alternative base based on the first increment subsequence and the second increment subsequence.
42. The artificial intelligence-based system of claim 40, wherein the pathogenicity prediction logic component is further configured to accumulate the first positional subsequence difference values into a first cumulative subsequence value, and accumulate the second positional subsequence difference values into a second cumulative subsequence value.
43. The artificial intelligence-based system of claim 42, wherein the first accumulated subsequence value is an average of the first position-wise subsequence difference values, and the second accumulated subsequence value is an average of the second position-wise subsequence difference values.
44. The artificial intelligence-based system of claim 42, wherein the first accumulated subsequence value is a sum of the first position-wise subsequence difference values, and the second accumulated subsequence value is a sum of the second position-wise subsequence difference values.
45. The artificial intelligence-based system of claim 42, wherein the pathogenicity prediction logic is further configured to generate the pathogenicity prediction for the alternative base based on the first accumulated subsequence value and the second accumulated subsequence value.
46. The artificial intelligence-based system of claim 45, wherein the pathogenicity prediction logic is further configured to generate the pathogenicity prediction for the alternative base based on an average of the first cumulative subsequence value and the second cumulative subsequence value.
47. The artificial intelligence-based system of claim 45, wherein the pathogenicity prediction logic is further configured to generate the pathogenicity prediction for the alternative base based on a sum of the first cumulative subsequence value and the second cumulative subsequence value.
48. The artificial intelligence-based system of claim 38, wherein the pathogenicity prediction logic component is further configured to classify positions in the first incremental subsequence as belonging to a conserved state or a non-conserved state based on the first positional subsequence difference.
49. An artificial intelligence-based system according to claim 48, wherein the pathogenicity prediction logic component is further configured to classify those positions in the second incremental subsequence as belonging to the signal state that coincide with those positions in the first incremental subsequence classified as belonging to the conserved state, and classify those positions in the second incremental subsequence as belonging to the noise state that coincide with those positions in the first incremental subsequence classified as belonging to the non-conserved state.
50. The artificial intelligence-based system of claim 49, wherein the pathogenicity prediction logic is further configured to accumulate a subset of the second position-wise subsequence difference values into a modulated cumulative subsequence value, The second position-wise subsequence difference values in the subset of the second position-wise subsequence difference values are located at those positions in the second increment subsequence that are classified as belonging to the signal state.
51. The artificial intelligence-based system of claim 50, wherein the pathogenicity prediction logic component is further configured to generate the pathogenicity prediction for the alternative base based on the modulated cumulative subsequence value.
52. The artificial intelligence based system of claim 50, wherein the modulated accumulated subsequence value is an average of the second positional subsequence difference values in the subset of the second positional subsequence difference values.
53. The artificial intelligence based system of claim 50, wherein the modulated accumulated subsequence value is a sum of the second position-wise subsequence difference values in the subset of the second position-wise subsequence difference values.
54. The artificial intelligence-based system of claim 1, wherein the target base sequence is a coding region of a gene.
55. The artificial intelligence-based system of claim 1, wherein the target base sequence is a non-coding region of a gene.
56. The artificial intelligence-based system of claim 55, wherein the non-coding region spans the transcription start site, 5' untranslated region (UTR), 3' UTR, enhancer and promoter.
57. The artificial intelligence-based system of claim 16, wherein the alternative base is a singleton variant that occurs in only one outlier individual among a cohort of outlier individuals, The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.
58. The artificial intelligence-based system of claim 56, wherein the extreme gene expression levels are determined from tail quantiles of normalized gene expression levels.
59. The artificial intelligence-based system of claim 57, wherein the extreme gene expression levels include gene overexpression and gene underexpression.
60. The artificial intelligence-based system of claim 56, wherein the singleton variant is a coding variant.
61. The artificial intelligence-based system of claim 56, wherein the singleton variant is a non-coding variant.
62. The artificial intelligence-based system of claim 60, wherein the non-coding variant is a promoter variant.
63. The artificial intelligence-based system of claim 60, wherein the non-coding variant is an enhancer variant.
64. The artificial intelligence based system of claim 1, wherein the biomass model has a first set of weights, Wherein the biomass output generation logic component has a second set of weights.
65. The artificial intelligence based system of claim 62, wherein: During training, the first set of weights of the biomass model is trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence, The second set of weights of the biomass output generation logic are trained end-to-end from scratch together with the first set of weights of the biomass model to process the alternative representations of the input base sequence and generate the plurality of biomass output sequences.
66. The artificial intelligence based system of claim 63, wherein: During inference, the biomass model uses the first set of weights trained, wherein, during said inferring, said biomass output generation logic uses a trained second set of weights.
67. The artificial intelligence based system of claim 1, wherein the gene expression model has a third set of weights, The gene expression output generation logic component has a fourth set of weights.
68. The artificial intelligence-based system of claim 65, wherein the third set of weights of the gene expression model is trained from scratch to process the plurality of biomass output sequences and generate the alternative representations of the plurality of biomass output sequences, The fourth set of weights of the gene expression output generation logic are trained end-to-end from scratch together with the third set of weights of the gene expression model to process the alternative representations of the plurality of biomass output sequences and generate the gene expression output sequences.
69. The artificial intelligence based system of claim 66, wherein: During inference, the gene expression model uses the trained third set of weights, Wherein, during the inference, the gene expression output generation logic component uses a trained fourth set of weights.
70. The artificial intelligence based system of claim 65, wherein:During training, the first set of weights of the biomass model is first trained from scratch to process the input base sequence and generate the alternative representations 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 plurality of biomass output sequences and generate the alternative representations of the plurality of biomass output sequences, wherein the fourth set of weights of the gene expression output generation logic component is end-to-end trained from scratch together with the trained first set of weights replaced in the gene expression model to process the alternative representations of the multiple biomass output sequences generated by the trained first set of weights replaced in the gene expression model and generate the gene expression output sequences.
71. The artificial intelligence based system of claim 68, wherein: During inference, the biomass model uses the retrained first set of weights, wherein, during said inference, said biomass output generation logic uses said trained second set of weights, wherein, during inference, the gene expression model uses the retrained first set of weights, and Wherein, during the inference, the gene expression output generation logic component uses the trained fourth set of weights.
72. The artificial intelligence based system of claim 17, wherein: During training, the first set of weights of the biomass model is first trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence, wherein during the training, the second set of weights of the biomass output generation logic are first trained end-to-end from scratch together with the first set of weights of the biomass model to process the alternative representations of the input base sequence and generate the plurality of biomass output sequences, wherein, during the training, the trained first set of weights of the biomass model is then retrained to process the reference base sequence and generate the alternative representation of the reference base sequence, and to process the alternative base sequence and generate the alternative representation of the alternative base sequence, and wherein, during the training, the trained second set of weights of the biomass output generation logic component is then retrained end-to-end together with the trained first set of weights of the biomass model to process the alternative representations of the reference base sequence and generate the multiple reference biomass output sequences, and to process the alternative representations of the alternative base sequence and generate the multiple alternative biomass output sequences.
73. The artificial intelligence based system of claim 70, wherein: During inference, the biomass model uses the retrained first set of weights, wherein, during the inference, the biomass output generation logic uses the retrained second set of weights.
74. The artificial intelligence-based system of claim 23, wherein the pathogenicity prediction logic component has a fifth set of weights.
75. The artificial intelligence based system of claim 72, wherein: During training, the first set of weights of the biomass model is first trained from scratch to process the input base sequence and generate the alternative representation of the input base sequence, wherein, during the training, the second set of weights of the biomass output generation logic are first trained end-to-end from scratch together with the first set of weights of the biomass model to process the alternative representations of the input base sequence and generate the plurality of biomass output sequences, and wherein, during the training, the trained first set of weights of the biomass model and the trained second set of weights of the biomass output generation logic are then retrained end-to-end to generate the pathogenicity predictions for the alternative bases.
76. The artificial intelligence based system of claim 73, wherein: During inference, the biomass model uses the retrained first set of weights, wherein, during said inference, said biomass output generation logic uses said retrained second set of weights, and Wherein, during the inference, the pathogenicity prediction logic component uses a trained fifth set of weights.
77. An artificial intelligence-based system according to claim 18, wherein the first corresponding per-base reference biomass output specifies a corresponding measurement of a first reference epigenetic signal level of the corresponding reference target base at the corresponding position in the reference target base sequence.
78. An artificial intelligence-based system according to claim 29, wherein the second corresponding per-base reference biomass output specifies a corresponding measurement of a second reference epigenetic signal level of the corresponding reference target base at the corresponding position in the reference target base sequence.
79. An artificial intelligence-based system according to claim 21, wherein the first corresponding alternative biomass per base output specifies a corresponding measurement of a first alternative epigenetic signal level of the corresponding alternative target base at the corresponding position in the alternative target base sequence.
80. An artificial intelligence-based system according to claim 22, wherein the second corresponding alternative biomass per base output specifies a corresponding measurement of a second alternative epigenetic signal level of the corresponding alternative target base at the corresponding position in the alternative target base sequence.
81. The artificial intelligence based system of claim 1, wherein: During training, the biomass model and the biomass output generation logic components are first trained end-to-end from scratch to convert the analysis of the input base sequence into a base-by-base evolutionarily conserved 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.
82. The artificial intelligence based system of claim 1, wherein: During training, the biomass model and the biomass 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 epigenetic 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.
83. The artificial intelligence based system of claim 1, wherein: During training, the biomass model and the biomass output generation logic components are first trained end-to-end from scratch to convert the analysis of the input base sequence into a base-by-base epigenetic 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.
84. The artificial intelligence based system of claim 1, wherein: During training, the biomass model and the biomass 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 epigenetic 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.
85. The artificial intelligence-based system of claim 1, further configured to include a first training set of training input base sequences, wherein the first training set includes variants confounded by multiple epigenetic effects.
86. The artificial intelligence-based system of claim 83, wherein the epigenetic effects in the plurality of epigenetic 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. PEER stands for "Probabilistic 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.
87. The artificial intelligence-based system of claim 83, 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 multiple epigenetic effects.
88. The artificial intelligence-based system of claim 85, wherein the variants in the second training set are reliably determined to alter gene expression and cause extreme gene expression levels.
89. The artificial intelligence-based system of claim 86, wherein the variants in the second training set include overexpression-causing variants that increase gene expression levels.
90. The artificial intelligence-based system of claim 86, wherein the variants in the second training set include underexpression-causing variants that reduce gene expression levels.
91. The artificial intelligence-based system of claim 87, wherein the second training set specifies an overexpression probability for the variant, the overexpression probability specifying the likelihood that the variant causes overexpression of a gene.
92. The artificial intelligence-based system of claim 88, wherein the second training set specifies a probability of underexpression for the variant, the probability of underexpression specifying the likelihood that the variant causes underexpression of a gene.
93. The artificial intelligence-based system of claim 86, wherein each variant in the second training set is a singleton variant that occurs in only one outlier individual among the outlier individual cohort, The outlier individuals in the outlier individual cohort exhibit extreme gene expression levels.
94. An artificial intelligence-based system according to claim 91, wherein the extreme gene expression levels are determined from the tail quantiles of the normalized gene expression levels.
95. The artificial intelligence-based system of claim 91, wherein the extreme gene expression levels include gene overexpression and gene underexpression.
96. The artificial intelligence-based system of claim 91, wherein the singleton variant is a coding variant.
97. The artificial intelligence-based system of claim 91, wherein the singleton variant is a non-coding variant.
98. The artificial intelligence-based system of claim 95, wherein the non-coding variant is a 5' untranslated region (UTR) variant, a 3' UTR variant, an enhancer variant, or a promoter variant.
99. The artificial intelligence-based system of claim 85, wherein the variants in the second training set span multiple tissue types.
100. The artificial intelligence-based system of claim 85, wherein the variants in the second training set span multiple cell types.
101. The artificial intelligence-based system of claim 1, wherein the input base sequence and the plurality of biomass output sequences span the plurality of tissue types.
102. The artificial intelligence-based system of claim 1, wherein the input base sequence and the plurality of biomass output sequences span the plurality of cell types.
103. The artificial intelligence-based system of claim 10, wherein the gene expression output sequences span the multiple tissue types.
104. The artificial intelligence-based system of claim 10, wherein the gene expression output sequences span the plurality of cell types.
105. The artificial intelligence based system of claim 1, wherein the biomass model and the biomass output generation logic are first trained end-to-end on the first training set and then retrained on the second training set.
106. The artificial intelligence-based system of claim 1, wherein the variants in the second training set are a pathogenic set labeled with a first ground truth label indicating altered gene expression, and the common variants are a benign set labeled with a second ground truth label indicating unchanged gene expression.
107. The artificial intelligence-based system of claim 104, wherein the benign set is balanced in terms of trinucleotide background, homopolymers, k-mers, genomic GC frequency, and sequencing depth.
108. The artificial intelligence based system of claim 104, wherein: Based on the cutoff probabilities 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.
109. The artificial intelligence-based system of claim 10, 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.
110. An artificial intelligence based system according to claim 1, wherein the biomass model and the biomass 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.
111. An artificial intelligence-based system according to claim 10, 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.
112. An artificial intelligence-based system according to claim 85, 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 biomass model 124, the trained biomass output generation logic component, the trained gene expression model and the trained gene expression output generation logic component.
113. An artificial intelligence-based system according to claim 110, wherein those variants in the second training set that appear on even chromosomes are used as the validation set.
114. 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).
115. An artificial intelligence-based system for detecting changes in gene expression at base resolution, comprising: 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; a biomass model that processes the input base sequence and generates alternative representations of the input base sequence; as well as a biomass output generation logic component that processes the alternative representations of the input base sequence and generates a plurality of biomass output sequences, Each of the plurality of biomass output sequences comprises a corresponding per-base biomass output for a corresponding target base in the target base sequence.
116. The artificial intelligence-based system of claim 113, wherein a first biomass output sequence in the plurality of biomass output sequences comprises a first corresponding per-base biomass output for the corresponding target base in the target base sequence, and Wherein the first corresponding per-base biomass output specifies a corresponding measure of evolutionary conservation of the corresponding target base across multiple species.
117. The artificial intelligence-based system of claim 113, wherein a second biomass output sequence in the plurality of biomass output sequences comprises a second corresponding per-base biomass output for the corresponding target base in the target base sequence, and Wherein the second corresponding biomass per base output specifies a corresponding measurement of transcription initiation of the corresponding target base at a corresponding position in the target base sequence.
Citation Information
Patent Citations
Artificial intelligence-based analysis of protein three-dimensional (3D) structures
US12217829B2
Deep Learning-Based Techniques for Training Deep Convolutional Neural Networks
US20190114511A1
Semi-Supervised Learning for Training an Ensemble of Deep Convolutional Neural Networks
US20190114544A1
Deep Learning-Based Techniques for Pre-Training Deep Convolutional Neural Networks
US20190266493A1
Deep Convolutional Neural Networks for Variant Classification
US20200065675A1
