Machine learning and bioinformatic methods for variant interpretation
By integrating saturation mutagenesis and single-cell RNA sequencing with machine learning, the method effectively predicts and characterizes the pathogenicity of genetic variants, overcoming the limitations of existing approaches by providing a comprehensive functional interpretation.
Patent Information
- Application Number
- PCT/US2024/028765
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-12-15
- Filing Date
- 2024-05-10
- Publication Date
- 2024-11-14
AI Technical Summary
Current methods for interpreting genetic variants are limited in accurately predicting pathogenicity, as they often rely on one-dimensional readouts and lack functional characterization, leading to a large number of variants of uncertain significance in clinical databases.
The use of saturation mutagenesis combined with single-cell RNA sequencing and machine learning to generate high-throughput data for determining the effects of genetic variants, enabling the identification of pathogenicity through expression profiles and functional interpretation.
This approach allows for state-of-the-art pathogenicity prediction and functional characterization of genetic variants, distinguishing between different modes of pathogenicity and improving the accuracy of variant classification, thereby addressing the limitations of existing methods.
Smart Images

Figure US2024028765_14112024_PF_FP_ABST
Abstract
Description
MACHINE LEARNING AND BIOINFORMATIC METHODS FOR VARIANT INTERPRETATIONCROSS REFERENCE TO RELATED APPLICATIONS
[0001] The present application claims the priority benefit of US Provisional Patent Application No. 63 / 465,372, filed May 10, 2023, US Provisional Patent Application No. 63 / 595,269, filed November 1, 2023, and US Provisional Patent Application No. 63 / 610,712, filed December 15, 2023, all incorporated by reference herein in their entirety.FIELD OF THE INVENTION
[0002] The technology disclosed relates to artificial intelligence type computers and digital data processing systems and corresponding data processing methods and products for emulation of intelligence (i.e., knowledge-based systems, reasoning systems, and knowledge acquisition systems); and including systems for reasoning with uncertainty (e.g., fuzzy logic systems), adaptive systems, machine learning systems, and artificial neural networks. In particular, the technology disclosed relates to using techniques for converting context of an artificial neural network (ANN) or another type of computing system that is trainable through machine learning.
[0003] Additionally, the technology disclosed relates to pre-processing of inputs for artificial intelligence type computers and digital data processing systems and corresponding data processing methods and products for emulation of intelligence as well as the actual pre-processed inputs themselves.BACKGROUND
[0004] The subject matter discussed in this section should not be assumed to be prior art merely as a result of its mention in this section. Similarly, a problem mentioned in this section or associated with the subject matter provided as background should not be assumed to have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, which in and of themselves can also correspond to implementations of the claimed technology.
[0005] Genomics, in the broad sense, also referred to as functional genomics, aims to characterize the function of every genomic element of an organism by using genome-scale assayssuch as genome sequencing, transcriptome profiling and proteomics. Genomics arose as a data- driven science - it operates by discovering novel properties from explorations of genome-scale data rather than by testing preconceived models and hypotheses. Applications of genomics include finding associations between genotype and phenotype, discovering biomarkers for patient stratification, predicting the function of genes, and charting biochemically active genomic regions such as transcriptional enhancers.
[0006] Genomics data are too large and too complex to be mined solely by visual investigation of pairwise correlations. Instead, analytical tools are required to support the discovery of unanticipated relationships, to derive novel hypotheses and models and to make predictions. Unlike some algorithms, in which assumptions and domain expertise are hard coded, machine learning algorithms are designed to automatically detect patterns in data. Hence, machine learning algorithms are suited to data-driven sciences and, in particular, to genomics. However, the performance of machine learning algorithms can strongly depend on how the data are represented, that is, on how each variable (also called a feature) is computed. For instance, to classify a tumor as malign or benign from a fluorescent microscopy image, a preprocessing algorithm could detect cells, identify the cell type, and generate a list of cell counts for each cell type.
[0007] A machine learning model can take the estimated cell counts, which are examples of handcrafted features, as input features to classify the tumor. A central issue is that classification performance depends heavily on the quality and the relevance of these features. For example, relevant visual features such as cell morphology, distances between cells or localization within an organ are not captured in cell counts, and this incomplete representation of the data may reduce classification accuracy.
[0008] Deep learning, a subdiscipline of machine learning, addresses this issue by embedding the computation of features into the machine learning model itself to yield end-to-end models. This outcome has been realized through the development of deep neural networks, machine learning models that comprise successive elementary operations, which compute increasingly more complex features by taking the results of preceding operations as input. Deep neural networks are able to improve prediction accuracy by discovering relevant features of high complexity, such as the cell morphology and spatial organization of cells in the above example. The construction and training of deep neural networks have been enabled by the explosion of data, algorithmic advances,and substantial increases in computational capacity, particularly through the use of graphical processing units (GPUs).
[0009] The goal of supervised learning is to obtain a model that takes features as input and returns a prediction for a so-called target variable. An example of a supervised learning problem is one that predicts whether an intron is spliced out or not (the target) given features on the RNA such as the presence or absence of the canonical splice site sequence, the location of the splicing branchpoint or intron length. Training a machine learning model refers to learning its parameters, which commonly involves minimizing a loss function on training data with the aim of making accurate predictions on unseen data.
[0010] For many supervised learning problems in computational biology, the input data can be represented as a table with multiple columns, or features, each of which contains numerical or categorical data that are potentially useful for making predictions. Some input data are naturally represented as features in a table (such as temperature or time), whereas other input data need to be first transformed (such as deoxyribonucleic acid (DNA) sequence 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 the canonical splice site sequence, the location of the splicing branchpoint and the intron length can be preprocessed features collected in a tabular format. Tabular data are standard for a wide range of supervised machine learning models, ranging from simple linear models, such as logistic regression, to more flexible nonlinear models, such as neural networks and many others.
[0011] Logistic regression is a binary classifier, that is, a supervised learning model that predicts a binary target variable. Specifically, logistic regression predicts the probability of the positive class by computing a weighted sum of the input features mapped to the [0,1] interval using the sigmoid function, a type of activation function. The parameters of logistic regression, or other linear classifiers that use different activation functions, are the weights in the weighted sum. Linear classifiers fail when the classes, for instance, that of an intron spliced out or not, cannot be well discriminated with a weighted sum of input features. To improve predictive performance, new input features can be manually added by transforming or combining existing features in new ways, for example, by taking powers or pairwise products.
[0012] Neural networks use hidden layers to learn these nonlinear feature transformations automatically. Each hidden layer can be thought of as multiple linear models with their output transformed by a nonlinear activation function, such as the sigmoid function or the more popular rectified-linear unit (ReLU). Together, these layers compose the input features into relevant complex patterns, which facilitates the task of distinguishing two classes.
[0013] Deep neural networks use many hidden layers, and a layer is said to be fully-connected when each neuron receives inputs from all neurons of the preceding layer. Neural networks are commonly trained using stochastic gradient descent, an algorithm suited to training models on very large data sets. Implementation of neural networks using modern deep learning frameworks enables rapid prototyping with different architectures and data sets. Fully-connected neural networks can be used for a number of genomics applications, which include predicting the percentage of exons spliced in for a given sequence from sequence features such as the presence of binding motifs of splice factors or sequence conservation; prioritizing potential disease-causing genetic variants; and predicting cis-regulatory elements in a given genomic region using features such as chromatin marks, gene expression and evolutionary conservation.
[0014] Local dependencies in spatial and longitudinal data must be considered for effective predictions. For example, shuffling a DNA sequence or the pixels of an image severely disrupts informative patterns. These local dependencies set spatial or longitudinal data apart from tabular data, for which the ordering of the features is arbitrary. Consider the problem of classifying genomic regions as bound versus unbound by a particular transcription factor, in which bound regions are defined as high-confidence binding events in chromatin immunoprecipitation following by sequencing (ChlP-seq) data. Transcription factors bind to DNA by recognizing sequence motifs. A fully-connected layer based on sequence-derived features, such as the number of k-mer instances or the position weight matrix (PWM) matches in the sequence, can be used for this task. As k-mer or PWM instance frequencies are robust to shifting motifs within the sequence, such models could generalize well to sequences with the same motifs located at different positions. However, they would fail to recognize patterns in which transcription factor binding depends on a combination of multiple motifs with well-defined spacing. Furthermore, the number of possible k- mers increases exponentially with k-mer length, which poses both storage and overfitting challenges.
[0015] A convolutional layer is a special form of fully-connected layer in which the same fully- connected layer is applied locally, for example, in a 6 bp window, to all sequence positions. This approach can also be viewed as scanning the sequence using multiple PWMs, for example, for transcription factors GATA1 and TALI. By using the same model parameters across positions, the total number of parameters is drastically reduced, and the network is able to detect a motif at positions not seen during training. Each convolutional layer scans the sequence with several filters by producing a scalar value at every position, which quantifies the match between the filter and the sequence. As in fully-connected neural networks, a nonlinear activation function (commonly ReLU) is applied at each layer. Next, a pooling operation is applied, which aggregates the activations in contiguous bins across the positional axis, commonly taking the maximal or average activation for each channel. Pooling reduces the effective sequence length and coarsens the signal. The subsequent convolutional layer composes the output of the previous layer and is able to detect whether a GATA1 motif and TALI motif were present at some distance range. Finally, the output of the convolutional layers can be used as input to a fully-connected neural network to perform the final prediction task. Hence, different types of neural network layers (e.g., fully-connected layers and convolutional layers) can be combined within a single neural network.
[0016] Convolutional neural networks (CNNs) can predict various molecular phenotypes on the basis of DNA sequence alone. Applications include classifying transcription factor binding sites and predicting molecular phenotypes such as chromatin features, DNA contact maps, DNA methylation, gene expression, translation efficiency, RBP binding, and microRNA (miRNA) targets. In addition to predicting molecular phenotypes from the sequence, convolutional neural networks can be applied to more technical tasks traditionally addressed by handcrafted bioinformatics pipelines. For example, convolutional neural networks can predict the specificity of guide RNA, denoise ChlP-seq, enhance Hi-C data resolution, predict the laboratory of origin from DNA sequences and call genetic variants. Convolutional neural networks have also been employed to model long-range dependencies in the genome. Although interacting regulatory elements may be distantly located on the unfolded linear DNA sequence, these elements are often proximal in the actual 3D chromatin conformation. Hence, modelling molecular phenotypes from the linear DNA sequence, albeit a crude approximation of the chromatin, can be improved by allowing for long-range dependencies and allowing the model to implicitly learn aspects of the 3D organization, such as promoter-enhancer looping. This is achieved by using dilated convolutions,which have a receptive field of up to 32 kb. Dilated convolutions also allow splice sites to be predicted from sequence using a receptive field of 10 kb, thereby enabling the integration of genetic sequence 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)).
[0017] Different types of neural network can be characterized by their parameter-sharing schemes. For example, fully-connected layers have no parameter sharing, whereas convolutional layers impose translational invariance by applying the same filters at every position of their input. Recurrent neural networks (RNNs) are an alternative to convolutional neural networks for processing sequential data, such as DNA sequences or time series, that implement a different parameter-sharing scheme. Recurrent neural networks apply the same operation to each sequence element. The operation takes as input the memory of the previous sequence element and the new input. It updates the memory and optionally emits an output, which is either passed on to subsequent layers or is directly used as model predictions. By applying the same model at each sequence element, recurrent neural networks are invariant to the position index in the processed sequence. 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 the recognition of a certain series of inputs, such as the start codon followed by an in-frame stop codon.
[0018] The main advantage of recurrent neural networks over convolutional neural networks is that they are, in theory, able to carry over information through infinitely long sequences via memory. Furthermore, recurrent neural networks can naturally process sequences of widely varying length, such as mRNA sequences. However, convolutional neural networks combined with various tricks (such as dilated convolutions) can reach comparable or even better performances than recurrent neural networks on sequence-modelling tasks, such as audio synthesis and machine translation. Recurrent neural networks can aggregate the outputs of convolutional neural networks for predicting single-cell DNA methylation states, RBP binding, transcription factor binding, and DNA accessibility. Moreover, because recurrent neural networks apply a sequential operation, they cannot be easily parallelized and are hence much slower to compute than convolutional neural networks.
[0019] Each human has a unique genetic code, though a large portion of the human genetic code is common for all humans. In some cases, a human genetic code may include an outlier, called agenetic variant, that may be common among individuals of a relatively small group of the human population. For example, a particular human protein may comprise a specific sequence of amino acids, whereas a variant of that protein may differ by one amino acid in the otherwise same specific sequence.
[0020] Genetic variants may be pathogenetic, leading to diseases. Though most of such genetic variants have been depleted from genomes by natural selection, an 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 diagnostics, treatments, or cures. The clinical interpretation of millions of human genetic variants remains unclear. Some of the most frequent pathogenic variants are single nucleotide missense mutations that change the amino acid of a protein. However, not all missense mutations are pathogenic.
[0021] Models that can predict molecular phenotypes directly from biological sequences can be used as in silico perturbation tools to probe the associations between genetic variation and phenotypic variation and have emerged as new methods for quantitative trait loci identification and variant prioritization. These approaches are of major importance given that the majority of variants identified by genome-wide association studies of complex phenotypes are non-coding, which makes it challenging to estimate their effects and contribution to phenotypes. Moreover, linkage disequilibrium results in blocks of variants being co-inherited, which creates difficulties in pinpointing individual causal variants. Thus, sequence-based deep learning models that can be used as interrogation tools for assessing the impact of such variants offer a promising approach to find potential drivers of complex phenotypes. One example includes predicting the effect of noncoding single-nucleotide variants and short insertions or deletions (indels) indirectly from the difference between two variants in terms of transcription factor binding, chromatin accessibility or gene expression predictions. Another example includes predicting novel splice site creation from sequence or quantitative effects of genetic variants on splicing.
[0022] The vast majority of all potential ~70 million protein-altering variants in the human genome remain of uncertain significance. Only ~0.1% of variants are annotated in clinical variant databases. Closing this gap is essential to identify clinically relevant variants and understand their mechanism of action. Scalable experimental methods are needed to enable assaying all possible variants in a gene. While Al models are becoming highly accurate in predicting variantpathogenicity, experimental measurements remain necessary to understand the functional consequences of variants.
[0023] The recent success of deep learning algorithms provide an opportunity to develop tools for automatically extracting relationships between thousands of genes. Therefore, an opportunity arises to predict variant pathogenicity of millions of variants along the entire length of a gene as input to deep neural networks.FIGURES
[0024] Fig. 1. Saturation mutagenesis with single cell RNA-seq readout. Schematic representation of the experimental workflow, consisting of 1) lentiviral transduction of a saturation mutagenesis library with random variant barcodes, followed by single cell RNA sequencing (scRNA-seq); 2) linking of variant barcodes to their corresponding cell barcode directly from the scRNA-seq data; and 3) targeted amplicon sequencing of the single cell cDNA library with a tiling set of primers, linking each random variant barcode to a specific mutation. Bottom panels: read pileups from four amplicons of a variant transcript in the TP53 library, with corresponding long read data for validation.
[0025] Fig. 2 Characterization of pl6 variants with saturation mutagenesis in single cells. (A) UMAP of expression profiles from 141,973 single A549 cells expressing a pl6 variant, labeled by cell cycle phase (Methods). (B) Same UMAP as in (A), labeled by PrimateAI-3D score (missense variants), showing a correlation between variant pathogenicity and cycling cells. (C) PCA of average expression profiles for amino acid variants detected in >30 cells, with missense variants labeled by PrimateAI-3D score. (D) Same PCA as in (C), with ClinVar variants and clusters from unsupervised clustering (Methods) labeled. (E) Top: heatmap showing the average expression profile (z-scores) at each residue position, computed as the average of amino acid-level profiles weighted by predicted pathogenicity (Methods). Bottom panels: proliferation score for missense and nonsense variants, with significantly enriched variants labeled (q-value < 0.05; Methods); predicted pathogenicity from expression (expression score; Methods); PrimateAI-3D scores for all missense variants, with ClinVar annotations. Light gray: regions excluded due to dropout in the saturation mutagenesis library. (F) Expression programs corresponding to gene modules up- and down-regulated upon expression of pathogenic variants. (G) Structure of pl 6 bound to CDK6 (PDB: 1BI7), with pl6 residues labeled by average predicted pathogenicity. (H) Distribution ofthe average expression of module 1 genes, proliferation score, and predicted pathogenicity at residue positions within the helices comprising the ankyrin repeats of pl6, showing that variants in helix 1 of the first three repeats are on average more pathogenic. (I) ROC curves and area under the curve (AUC) for single cell gene expression and proliferation-derived scores and computational predictions for distinguishing ClinVar pathogenic from benign variants. (J) Classification results for additional computational predictions on the subset of variants from (I) scored by all methods, with methods trained using clinical variant annotations or human allele frequencies indicated by *.
[0026] Fig. 3. Domain-specific variant effects in TP53 are associated with multiple axes of pathogenicity. (A) UMAP of expression profiles from 206,394 single A549 cells expressing a TP53 variant, labeled by cell cycle phase (Methods). (B) Same UMAP as in (A), showing missense variants labeled by PrimateAI-3D. (C) UMAP of expression profiles averaged by amino acid for missense variants labeled with PrimateAI-3D scores, with the distribution of synonymous variants shown as density contours. (D) Top: heatmap showing the expression profile (z-score) averaged across neighboring residue positions (Methods). Bottom panels: proliferation score for missense and nonsense variants, with significantly enriched or depleted variants indicated (q-value < 0.05; Methods); expression-derived pathogenicity prediction (expression score; Methods) for variants with an expression signature matching that of variants in the DNA binding domain (DBD-like) or in Transactivation Domain I (TAD I— like); allele frequency of germline TP53 variants across -416,800 individuals, showing an absence of variants in a subregion of TAD I (highlighted in red across panels); PrimateAI-3D scores for all missense variants, with annotations from ClinVar. (E) Transcriptional programs corresponding to the gene modules identified from unsupervised clustering of gene expression (Methods). (F) Experimental scores and computational pathogenicity predictions in TAD I. (G) UMAP of expression profiles averaged across neighboring residue positions (Methods), labeled by unsupervised clusters (top) and domain annotations (bottom). (H) Correlation between the average expression profiles for pathogenic variants at each residue position, juxtaposed with the contact map from the AlphaFold prediction. (I) Proliferation scores for variants classified as having TAD I-, DBD-, or synonymous-like expression signatures across domains. (J) Mean expression of genes in modules 3 and 6 for variants classified as having TAD I-, DBD-, or synonymous-like expression signatures across domains. (K) Expression of core p53 pathway genes for variants with TAD I-, DBD- and synonymous-like signatures. (L-N) Structuresof MDM2 bound to TAD I (L; PDB: 1YCR), DBD (M; PDB: 1TUP) and 4D tetramer (N; PDB: 1C26), labeled by average TAD I- and DBD-like pathogenicity scores.
[0027] Fig. 4. Variant classification performance for TP53. Benchmarking of single cell derived scores (expression alone (Expression), proliferation alone (Proliferation), and ensemble (Expression & Proliferation)) against computational predictors of pathogenicity (Methods) on the task of distinguishing pathogenic from benign variants annotated in ClinVar. (A) ROC curves and area under the curve (AUC) for single cell scores, PrimateAI-3D, AlphaMissense and the bulk assay dominant-negative score from (Giacomelli et al., 2018). (B) Performance of experimentally derived scores and computational models on the task of classifying pathogenic vs. benign ClinVar variants, with models that were trained on ClinVar or other clinical variant annotations or human allele frequencies indicated by *, and scores from this work labeled in blue. (C) Comparison of the ensemble score and PrimateAI-3D from (A), showing that the ensemble score better separates pathogenic from benign variants. (D) Performance of the same methods as in (C) on the task of distinguishing between TAD I variants that were observed in the population vs. variants that were never observed as a proxy for pathogenicity in TAD I. Error bars indicate standard deviations of the scores from bootstrapping (Methods).
[0028] Fig. 5. Identification of distinct modes of SOD1 pathogenicity improves classification of ALS variants. (A) Top: heatmap of average expression profiles at each amino acid position for 594 genes that were correlated with proliferation, with domain annotations as in (B) and gene modules from unsupervised clustering (Methods). Bottom panels: proliferation score for missense and nonsense variants, with significantly depleted variants labeled (q-value < 0.05; Methods); expression score; PrimateAI-3D scores for all missense variants, with pathogenic variants from ClinVar; frequency of SOD1 variants in a cohort of 1,003 ALS cases (Opie-Martin et al., 2022), with the most common pathogenic variants labeled. (B) Structure of the WT SOD1 homodimer (PDB: 2C9V), with zinc-binding and electrostatic loops and beta barrel labeled. (C) Expression programs corresponding to genes anti -correlated (module 1) and correlated (module 2) with proliferation. (D) Proliferation score for 18 variants across categories of effects on enzymatic activity. (E) Scatter plot showing the nonlinear relationship between PrimateAI-3D predictions and proliferation score, with ClinVar pathogenic variants indicated and marginal distributions of proliferation scores for each the variant category. (F-I) SOD1 monomer (same structure andorientation as in (B)) colored by proliferation score, expression score, PrimateAI-3D score, and change in PrimateAI-3D score after regressing out proliferation and expression scores (Methods). (J) Classification performance for distinguishing ClinVar pathogenic variants from variants not observed in any annotation of likely pathogenic variants (Methods).
[0029] Fig. 6 Saturation mutagenesis libraries and validation of amplicon sequencing with long reads. (A) Coverage of each library from amplicon sequencing, shown as number of variant barcodes per alternative base at each position. Shaded areas for CDKN2A indicate positions with poor coverage that were excluded from analyses. (B) Number of variant barcode-variant combinations in each library as detected by amplicon sequencing. Each library contained variant barcodes associated with the wild type (WT) sequence or multiple variants in addition to single variants. (C) Examples of amplicon sequencing data for two TP53 variants, with pileups from each amplicon shown in the top panels and the corresponding HiFi read in the bottom panel. (D) Concordance of variant calls from amplicon sequencing with HiFi long reads (PacBio), for variant barcodes with the WT sequence or 1 variant.
[0030] Fig. 7. Properties of pl6 pathogenicity scores from saturation mutagenesis in single cells. (A) Distribution of the number of cells obtained per cDNA variant for synonymous (synon.), missense, and nonsense variants. (B) Distribution of PrimateAI-3D scores stratified by cell cycle phase, with cell numbers indicated at the top. (C) Concordance between the expression and proliferation scores, with dashed lines indicating the 0.05 FDR threshold for each score for classifying pathogenic variants. (D) Correlation of the proliferation and expression scores with pathogenicity predictions for missense variants from PrimateAI-3D. (E) Correlation of the proliferation score from A549 lung carcinoma epithelial cells (this work) with proliferation scores from two pancreatic ductal adenocarcinoma cell lines for 43 variants assayed in (Kimura et al., 2022). (F) Performance of experimental scores and computational predictions on the task of distinguishing 31 pathogenic variants in ClinVar from 4 benign variants supplemented with variants observed in human populations up to a selected minor allele frequency (MAF) threshold. (G) Precision-recall curves and area under the curve (AUC) for single cell gene expression and proliferation-derived scores and computational predictions for classifying ClinVar pathogenic from benign variants.
[0031] Fig.8. Properties of domain-specific modes of pathogenicity in TP53. (A) Distribution of the number of cells obtained per cDNA variant for synonymous (synon.), missense, and nonsense variants. (B) Distribution of PrimateAI-3D scores stratified by cell cycle phase, with cell numbers indicated at the top. (C) UMAP of expression profiles averaged by amino acid, labeled by ClinVar annotation (left) and by the predicted pathogenicity from the ensemble classifier (right), with the distribution of synonymous variants shown as density contours and the densities along UMAP dimension 1 shown at the top. (D) UMAP of expression profiles averaged by residue position, labeled by unsupervised clusters (left) and domain annotations (right). (E) Expression scores for variants with an expression signature matching that of variants in the DBD, in TAD I, or in 4D, showing that no variants had a signature specific to the tetramerization domain. (F) Top: heatmap showing the expression profile (z-score) of nonsense variants averaged across neighboring residue positions (Methods). Bottom panels: proliferation score for nonsense variants, with significantly enriched or depleted variants indicated (q-value < 0.05); expression score for nonsense variants, with variants predicted to be pathogenic indicted (q-value < 0.05); loss-of-function bulk assay score from (Giacomelli et al., 2018). (G) Precision-recall curves and area under the curve (AUC) for the benchmark from Figure 4A.
[0032] Fig. 9. Variant classification performance as a function of sequencing depth and number of cells. (A) Classification performance (AUROC) on the task of distinguishing pathogenic from benign ClinVar variants as a function of total cell number and mean sequencing depth per cell, for the subset of variants detected across all downsampling conditions. (B) Spearman correlation between expression score (i.e., the probability of a variant having DBD- or TAD I— like pathogenicity based on its expression profile) and predicted pathogenicity from PrimateAI-3D. (C) UMAP of amino acid-level average expression profiles computed from a total of 50k (top) and 200k cells (bottom) sampled to a mean of 50k UMIs, showing that larger cell numbers improve the detection of DBD- and TAD I— like pathogenic effects. UMI: unique molecular identifier.
[0033] Fig. 10. Benchmarking of catalysis- and ALS-associated pathogenicity in SOD1. (A) Distribution of the number of cells obtained per cDNA variant for synonymous (synon.), missense, and nonsense variants. (B) UMAP of expression profiles averaged by amino acid for missense variants labeled with PrimateAI-3D scores, with the distribution of synonymous variants shown as density contours. (C) Performance of experimentally derived scores and computational modelson the task of classifying variants that have been shown to reduce the enzymatic activity of SOD1 vs. variants with no effect on enzymatic activity, with models that were trained on ClinVar or other clinical variant annotations or human allele frequencies indicated by *, and scores from this work labeled in blue. (D) Allele frequency of SOD1 missense observed across -416,800 individuals in UK Biobank, gnomAD, and TOPMed, containing 42 of the 56 pathogenic variants annotated in ClinVar. (E) Precision-recall curves and area under the curve (AUC) for the benchmark from Figure 5J. (F) Performance of experimentally derived scores and computational models on classifying pathogenic vs. benign ClinVar variants, for raw scores and scores residualized by the expression score.BRIEF SUMMARY
[0034] The present disclosure encompasses the discovery that the effects of variants can be mapped using variant perturb-seq with high throughput sequencing and machine learning to determine the pathogenicity of each variant.
[0035] Some embodiments presented herein include a method of simultaneously determining the effects of a plurality of variants in a gene comprising a) generating a mutant expression vector for each of the plurality of variants; b) expressing each variant in a cell in a plurality of single cells; and c) sequencing the RNA expressed in each single cell and evaluating expression patterns that result in each single cell to generate single cell RNA sequencing data.
[0036] Some embodiments presented herein include a method of simultaneously determining the effects of a gene comprising: a) perturbing the function of a gene in a cell in a plurality of single cells; and b) sequencing the RNA expressed in each single cell and evaluating expression patterns that result in each single cell to generate single cell RNA sequencing data.
[0037] Some embodiments presented herein include a method of improving prediction of pathogenicity of one or more variants in a gene comprising: a) a) generating a mutant expression vector for each of the plurality of variants; b) expressing each variant in a cell in a plurality of single cells; c) sequencing the RNA expressed in each single cell and evaluating expression patterns that result in each single cell to generate single cell RNA sequencing data; d) obtaining a quantitative value of enzymatic function of the gene for each variant to yield an enzyme function score; and e) training a machine learning model with the enzyme function scores, thereby refining the predictive power of a pathogenicity prediction model.
[0038] In certain aspects of the above embodiments, each mutant expression vector comprises arandom barcode for each variant. In certain aspects of the above embodiments, the method further comprises evaluating the single cell RNA sequencing data to determine the pathogenicity of each variant. In certain aspects of the above embodiments, the method further comprises generating a cell barcode during reverse transcription, resulting in said barcode being coupled to the cDNA.
[0039] In certain aspects of the above embodiments, the method comprises generating the single cell RNA sequencing data comprises generating a plurality of amplicons that comprise both the variant and the random barcode.
[0040] In certain aspects of the above embodiments, generating the single cell RNA sequencing data comprises generating a plurality of amplicons that comprise both the variant and the cell barcode.
[0041] In certain aspects of the above embodiments, evaluating the single cell RNA sequencing data comprises use of machine learning.
[0042] In certain aspects of the above embodiments, variants are classified as either pathogenic, likely pathogenic, likely benign, or benign.
[0043] In certain aspects of the above embodiments, each cell of the plurality of single cells is configured to express a single variant of the plurality of variants.
[0044] In certain aspects of the above embodiments, each single variant of the plurality of variants is expressed in more than one cell, thereby creating redundancy of the single cell RNA sequencing data.
[0045] In certain aspects of the above embodiments, the plurality of single cells comprises at least 30,000, at least 40,000, at least 50,000, at least 60,000, at least 70,000, at least 80,000, at least 90,000 or at least 100,000, 150,000, 200,000 or 250,000 single cells.
[0046] In certain aspects of the above embodiments, the plurality of variants comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1100, 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900, 2000, 2100, 2200, 2300, 2400, 2500, 2600, 2700, 2800, 2900 or at least 3000 variants.
[0047] In certain aspects of the above embodiments, the gene is selected from genes associated with rare disease.
[0048] In certain aspects of the above embodiments, the gene is selected from genes associated with oncology.
[0049] In certain aspects of the above embodiments, the mutant expression vector is generatedusing CRISPR.
[0050] In certain aspects of the above embodiments, the mutant expression vector is generated using CRISPRi.
[0051] In certain aspects of the above embodiments, the mutant expression vector is generated using CRISPR.
[0052] In certain aspects of the above embodiments, the RNA from each single cell is identified using barcoding and sequencing.
[0053] In certain aspects of the above embodiments, the method further comprises repeating steps a) - c) in more than one set of primary cells or cell lines.
[0054] In certain aspects of the above embodiments, the method further comprises evaluating the single cell RNA sequencing data to determine the pathogenicity of each perturbed gene.
[0055] In certain aspects of the above embodiments, perturbing the function of a gene comprises using one or more drug compounds.
[0056] In certain aspects of the above embodiments, a different drug compound is administered to each single cell.
[0057] In certain aspects of the above embodiments, the plurality of drug compounds comprises at least 100, at least 200, at least 300, at least 500, at least 1,000, at least 2,000, at least 3,000 at least 4,000, at least 5,000, or at least 10,000 compounds.
[0058] In certain aspects of the above embodiments, evaluating the single cell RNA sequencing data comprises use of machine learning.
[0059] The details of one or more embodiments are set forth in the claims and the description below. Other features, objects, and advantages will be apparent from the description, and from the claims.DETAILED DESCRIPTION
[0060] The present disclosure provides improved techniques for detecting and characterizing the pathogenicity of any of a variety of variants.
[0061] As described in Adamson, et al. (2016). "A Multiplexed Single-Cell CRISPR Screening Platform Enables Systematic Dissection of the Unfolded Protein Response" (Cell. 167 (7): 1867— 1882. e21) and Dixit, et al. (2016). "Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens" (Cell. 167 (7): 1853-1866. el 7), the content of each is hereby incorporated by reference in its entirety, a high-throughput methodologycan be used to identify the functions and relationships of a number of genes. The embodiments presented herein expand upon those initial discoveries, and show that the effects of variants can be mapped using high throughput sequencing and machine learning to determine the pathogenicity of each variant.EXAMPLE 1
[0062] Single cell sequencing as a general variant interpretation assay
[0063] The human genome contains around 70 million possible protein-altering genetic variants, the vast majority of which are of uncertain clinical significance. Closing this gap is essential for accurate diagnosis of disease-causing variants in human genome and exome sequencing and understanding their mechanisms of action. Towards this goal, we developed a pooled perturbation approach combining saturation mutagenesis with single cell RNA sequencing to map the functional effects of every single nucleotide variant in a gene of interest. We sequenced -440,000 cells expressing variants in the CDKN2A (pl6INK4a), TP53, and SOD1 genes, observing almost all possible protein-coding variants in at least 20 cells each. Using single cell gene expression signatures, we show that each gene may contain multiple types of pathogenic variants that affect distinct downstream pathways, corresponding to domain-specific structure-function relationships. We demonstrate that single cell expression signatures outperform existing experimental assays and computational models for predicting pathogenicity and summarize both the utility and potential limitations of single cell sequencing as a general variant interpretation assay.
[0064] Understanding the functional consequences of genetic variation is key to identifying pathogenic variants and their role in disease. Only a small fraction of all possible -70 million protein-altering variants in the human genome have been categorized and annotated in clinical variant databases such as ClinVar (Landrum et al., 2016), with the vast majority remaining variants of uncertain significance (VUS). Significant progress in variant characterization has been made using high-throughput saturation mutagenesis experiments also known as deep mutational scanning (Fowler and Fields, 2014), which typically rely on cellular survival or proliferation as a readout of pathogenicity (Findlay et al., 2018; Giacomelli et al., 2018; Matreyek et al., 2018). However, such one-dimensional readouts preclude identifying the functional mechanisms associated with individual variants or distinguishing between different modes of pathogenicity, or the contexts wherein they may arise. Moreover, while similar progress to predict variantpathogenicity has been achieved with deep learning models (Brandes et al., 2023; Cheng et al., 2023; Frazer et al., 2021; Gao et al., 2023), these models also lack a functional readout.
[0065] In combination with saturation mutagenesis assays, where each cell receives a different perturbing variant, single cell sequencing has the potential to provide a rich expression signature for each individual perturbation (Ursu et al., 2022). Although RNA-seq from individual single cells is sparse, averaging expression across cells that received the same variant can provide robust gene expression signatures of the perturbed downstream pathways, while also minimizing batch effects between perturbations (Replogle et al., 2022). Recent Perturb-seq experiments with CRISPRi and CRISPRa on the scale of thousands of genes have also shown that single cell expression signatures can serve as a general phenotypic readout (Replogle et al., 2022; Schmidt et al., 2022), reducing the necessity of developing bespoke assays for individual genes, which may only measure a subset of their functions. To date, single cell expression readouts have only been performed for a limited number of variants per gene (Ursu et al., 2022), nevertheless showing that the effects of benign and pathogenic variants can be distinguished by their transcriptomic signatures. This existing method relied on fixed molecular barcodes to encode the identity of 100 variants, which only represented a small fraction (-3%) of the total number of possible exonic single nucleotide variants (SNVs). To construct saturation mutagenesis libraries containing the thousands of possible variants in most genes, we instead use random barcodes to encode each variant and read out the variant corresponding to each barcode in a separate sequencing step.
[0066] We apply this approach to characterize almost all amino acid variants encoded by SNVs in three genes: CDKN2A (cyclin-dependent kinase inhibitor 2A), specifically the p!6INK4aisoform, a tumor suppressor and cell cycle regulator that is inactivated by somatic mutation in many cancers and for which germline mutations underlie familial melanoma (Hussussian et al., 1994) and pancreatic cancer (Goldstein et al., 1995); TP53 (the tumor suppressor p53), the most frequently mutated gene across human cancers wherein germline mutations result in the cancer-predisposing Li-Fraumeni syndrome (McBride et al., 2014); and SOD1 (superoxide dismutase 1), the first gene causally linked to amyotrophic lateral sclerosis (ALS) (Rosen et al., 1993), with SOD1 mutations accounting for -12% of familial and -1-2% of sporadic ALS cases (Goutman et al., 2022). Using gene expression profiles from hundreds of thousands of cells expressing mutant alleles of these genes, we apply machine learning approaches to identify distinct modes of pathogenicity linked to specific structure-function relationships. We benchmark our pathogenicity scores againstcomputational predictions and prior deep mutational scanning experiments to demonstrate that our single cell assay enables state-of-the-art pathogenicity prediction coupled to a functional interpretation of variant effects.
[0067] Methods used in this example are set forth below in the section titled “Methods”.RESULTSSaturation mutagenesis with single cell RNA sequencing readout
[0068] We generated saturation mutagenesis libraries for the three genes (Methods), obtaining close to uniform coverage for all possible SNVs in TP53 and SOD1, but with some regions of low or missing coverage in CDKN2A due to synthesis limitations related to high GC content (Figure 6A). The libraries were cloned into lentiviral vectors containing in the 3’ UTR a 19 bp semirandom variant barcode designed to minimize homopolymer and polyadenylation motif formation (Methods), resulting in over -750,000 different variant barcode-variant combinations per gene (Figure 6B). For each gene, these barcoded lentiviral vectors were then transduced into cells at low multiplicity of infection (MOI; Methods and Figure 1), selected to target a single variant per cell. All experiments were conducted in A549 lung carcinoma epithelial cells, which have a homozygous deletion of the CDKN2A locus (Ikediobi et al., 2006). Following lentiviral transduction, we prepared 3’ scRNA-seq libraries, from which the variant barcode(s) in each cell were read out directly (Figure 1 and Methods).
[0069] To recover the variant corresponding to each variant barcode, we performed amplicon sequencing on the cDNA library from each experiment, with primers designed to amplify exogenous transcripts and tiling the coding sequence of each gene (Figures 1 and 6C and Methods). Because lentiviral recombination can rearrange variant barcode-variant pairings, we performed amplicon sequencing on each experiment to maximize the accuracy of variant assignments. We developed a custom pipeline to call variants from amplicon sequencing, using stringent quality controls to only retain barcodes that could be confidently assigned to a single variant or wild type (WT) sequence (-1-22% of each library consisted of WT transcripts; Figure 6B and Methods). To verify the accuracy of our variant calls, we sequenced the TP53 library using long reads, obtaining >99% concordance (Figures 1, 6C and 6D and Methods).
[0070] We integrated the variant calls with the scRNA-seq data by matching on variant barcodes, retaining only cells that contained a single variant and following standard quality controlprocedures (Luecken and Theis, 2019) to select cells for analysis, from which we generated the covariate-corrected and normalized gene expression matrices (as z-scores; Methods) that we used to characterize variant effects on the transcriptome.A gradient of pathogenicity of pl6 variants linked to disruption of CDK6 binding
[0071] After applying quality controls, we obtained 141,973 single cells expressing 1,389 of 1,413 (98.3%) possible CDKN2A / pl6INK4a (henceforth pl6) variants, with 1,230 SNVs (88.6%) observed in at least 20 cells, corresponding to 964 of 1,069 (90.2%) possible amino acid variants (Figure 7A). We excluded 81 amino acid variants from all analyses due to insufficient coverage in the saturation mutagensis library (Figure 6A and Methods). Assigning each cell to a cell cycle phase based on its expression profile showed that cycling cells were enriched for pathogenic variants as predicted by PrimateAI-3D (Gao et al., 2023) (Figures 2A, 2B and 7B and Methods), consistent with the role of pl6 in inhibiting cell proliferation by binding to cyclin-dependent kinases 4 and 6 (CDK4 / 6) (Russo et al., 1998). For simplicity, we used PrimateAI-3D scores to predict variant pathogenicity throughout this work given their performance in prior benchmarks (Gao et al., 2023), but we also comprehensively benchmark our results against other computational predictions (Methods). Averaging the expression profiles across cells for each variant revealed a separation into two unsupervised clusters (Methods) corresponding to likely pathogenic and synonymous-like benign variants, respectively, on the basis of PrimateAI-3D scores (Figure 2C) and ClinVar annotations (Figure 2D).
[0072] We trained a classifier on per-variant averaged gene expression profiles to distinguish between synonymous and nonsynonymous variants (Methods), defining pathogenicity as the probability of a variant being nonsynonymous; we refer to this as the ‘expression score’ for simplicity (Figure 2E). Using this score, we classified 155 / 868 missense variants and 17 / 21 nonsense variants as pathogenic (Methods). The expression score was highly concordant with cell proliferation, calculated as the ratio of cells containing a specific variant over its abundance in the original library, normalized using synonymous variants (Methods); we refer to this as the ‘proliferation score’ (97.3% of missense and all nonsense variants were assigned the same label by both scores; Figures 2E and 7C). Both scores were also correlated with PrimateAI-3D (Figure 7D; Spearman’s p = 0.54-0.58) but showed a stronger separation between putatively benign and pathogenic variants than computational predictions (Figure 2E). The proliferation score replicatedresults from a prior study that assayed 43 missense variants in pancreatic ductal adenocarcinoma cell lines (Kimura et al., 2022) (Spearman’s p = 0.74-0.78, Figure 7E).
[0073] To identify relationships between variant pathogenicity and protein structure, we computed residue position-level average expression profiles by weighting the amino acid-level profiles of missense variants by their corresponding expression score (Figure 2E). This showed that pathogenic variants were primarily concentrated in the four ankyrin repeats of pl 6, each consisting of two alpha helices, Hl and H2 (Figure 2E), as well as in the loops connecting each pair of helices, with variants in the N and C termini appearing to be benign. Most nonsense variants were pathogenic as expected given the CDKN2A null background, except for variants in the C terminus, which had no measurable effects in our assay. The E2* variant had a similar synonymous-like effect, consistent with stop codon read-through (Dabrowski et al., 2015). Pathogenic variants had stable and expected transcriptomic effects, including upregulation of E2F target and G2-M checkpoint genes (gene module 1) and downregulation of genes in pathways including TNFa signaling, epithelial to mesenchymal transition, and apoptosis (Gil and Peters, 2006) (gene modules 2 and 3; Figures 2E and 2F).
[0074] Mapping the expression score onto the structure of pl6 bound to CDK6 (Russo et al., 1998) further showed that the residue positions with the highest average pathogenicity were those closest to the binding interface between the two proteins, as well as internal to pl6 (Figure 2G). Outward-facing residue positions in the H2 helices were less pathogenic than inward-facing ones. This was consistent across measures of pathogenicity, with the average expression of module 1 genes and our expression and proliferation scores all showing higher pathogenicity in the Hl helices except for the last repeat (Figure 2H). Variants in the Hl-2 and Hl-3 helices were the most pathogenic (Figure 2H), consistent with their role in binding CDK6 (Russo et al., 1998).
[0075] Next, we sought to benchmark our pathogenicity scores on the task of classifying pathogenic from benign variants in ClinVar. Since only four pl6 variants were annotated as benign in ClinVar, we complemented these with 6 variants observed at allele frequencies >5e-5 across gnomAD, TOPMed, and UK Biobank for benchmarking purposes, reasoning that such variants were likely benign. We chose this threshold to maximize the number of putatively benign variants; the results were robust across allele frequency thresholds (Figure 7F). We benchmarked our scores on a common set of 34 pathogenic and 9 putatively benign variants against predictions that were scored by 22 computational models (Methods), including SIFT (Ng, 2003), PolyPhen-2 (Adzhubeiet al., 2013), REVEL (loannidis et al., 2016), CADD (Rentzsch et al., 2019), AlphaMissense (Cheng et al., 2023) and PrimateAI-3D; this included methods that were trained using clinical variant annotations (e.g., ClinVar) or human allele frequencies, as well as methods trained without such annotations. The classifier based on our expression score achieved the best performance (ROC AUC = 0.987), outperforming the proliferation-based classifier (ROC AUC = 0.963, Mann- Whitney P = 3.1e-85) and computational predictions, including VEST-4 (ROC AUC = 0.937, Mann-Whitney P = 5.7e-176; trained on human disease variants) (Carter et al., 2013a) and PrimateAI-3D (ROC AUC = 0.897, Mann-Whitney P = 2.7e-195; trained without human variant annotations) (Figures 21, 7G and 2J). Based on these results, the expression score enabled classifying 47 ClinVar VUS as likely pathogenic and 317 as likely benign (with 4 not scored due to insufficient coverage). The significantly lower performance of computational predictors in this benchmark is likely due to a combination of pathogenic pl6 variants disrupting binding to CDK4 / 6 and the absence of protein-protein interactions from the data used to train these predictors. Taken together, our results show that single cell gene expression profiles accurately capture perturbations induced by pathogenic variants and enable state-of-the-art prediction of variant pathogenicity in pl6.Domain-specific pathogenic effects of TP53 missense variants
[0076] We sequenced 206,394 single cells expressing TP53 variants, observing all 3,546 possible SNVs, with 3,442 (97.1%) in at least 20 cells, corresponding to 2,739 of possible 2,814 (97.3%) amino acid variants (Figure 8A). Similar to pl 6, two-dimensional projection of the single cell gene expression profiles showed that cycling cells were enriched for pathogenic variants as predicted by Primate AI-3D (Figures 3 A, 3B and 8B), consistent with the role of TP53 in regulating cell cycle progression (Hafner et al., 2019) and dominant negative effects of TP53 missense mutations (Boettcher et al., 2019; Giacomelli et al., 2018). While this association was relatively weak at the single cell level, averaging the expression profiles for each variant revealed a clear distinction of variants predicted to be pathogenic from synonymous cells, as well as a separate smaller cluster of nonsynonymous cells (Figures 3C and 8C).
[0077] To globally identify the effects of pathogenic variants on gene expression and link these effects to functional mechanisms, we first computed residue position-level average expression profiles (Figure 3D). Since the average expression changes were weaker compared with pl6 variants due to the expression of TP53 variants on a WT background, we additionally averagedthe per-position averages across neighboring residue positions in the 3D structure (Methods). The resulting expression profdes revealed distinct, domain-specific signatures of variants, notably in Transactivation Domain I (TAD I) and the DNA-binding domain (DBD) (Figure 3D). We identified six distinct gene modules with perturbed expression by unsupervised clustering of genes across residue positions (Methods), with the main effects corresponding to the upregulation of E2F target and G2-M checkpoint genes among related pathways (module 6) for pathogenic variants in DBD, an apoptosis signature (module 4), and a p53 / TNFa signaling signature (module 2), with the latter two being specific to pathogenic variants in TAD I (Figure 3E). These changes in gene expression were consistent with effects on proliferation, with the highest proliferation scores (computed as for pl 6; Methods) observed in DBD and a depletion of cells containing TAD I variants (Figures 3D and 3F), matching an apoptosis signature (module 4). Consistent with this, apoptosis-inducing TAD I variants were not observed in a population of -416,800 individuals (aggregated across gnomAD, TOPMed, and UK Biobank; Figure 3D).
[0078] Two-dimensional projection and unsupervised clustering of the per-position gene expression profiles (Methods) grouped residue positions by domain, with effects of variants in TAD I, DBD, and the tetramerization domain (4D) being distinct from any other domain (Figure 3G). This grouping was also observed using position-level average expression profiles, but with less distinct separation of positions in TAD I and DBD (Figure 8D). Variants in Transactivation Domain II (TAD II) could not be distinguished from variants in unstructured regions or from some synonymous variants (Figure 3G). Based on these results, we trained a classifier to differentiate between TAD I-, DBD-, 4D- and synonymous-like signatures. All variants were scored as TAD I- , DBD- or synonymous-like, with pathogenic variants in 4D being classified as DBD-like (Figures 3D and 8E). This also showed that variants in 4D had generally weaker DBD-like effects, and that some residues within the DBD had TAD I-like signatures, notably at S121 and N239. Variants with TAD I-like signatures had consistent depleting effects across domains, and the weaker DBD- like effects of 4D variants were also apparent on proliferation (Figure 31). Moreover, the weaker effect of 4D variants was also apparent on gene expression, with lower average down- or upregulation of module 3 and 6 genes, respectively (Figure 3 J). The classifier also predicted 90 of the 110 nonsense variants in the library as pathogenic (Figure 8F). However, since nonsense- mediated decay is likely abnormal in the absence of splicing of exogenous mRNAs, dominant negative effects of truncated p53 variants are likely unphy si ologi cal. Nonetheless, our observationthat variants occurring prior to TAD II are less pathogenic and variants in the C terminus likely benign, are consistent with loss of function scores from a bulk assay performed in a TP53 null background (Giacomelli et al., 2018) (Figure 8F).
[0079] Using the average expression profde at each residue position of variants classified as pathogenic, i.e., variants having either DBD- or TAD I-like pathogenicity, we computed the correlation of expression profiles across positions (Figure 3H). This further confirmed that our assay captured two orthogonal, domain-specific axes of pathogenicity in TP53, showing anticorrelation between residues with DBD-like pathogenicity and residues with TAD I-like pathogenicity, and with residues having correlated gene expression signatures broadly matching contacts in the 3D structure (Figure 3G).
[0080] Next, we sought to characterize the structure-function relationships underlying domainspecific pathogenic effects, starting with variants in TAD I. The hydrophobic residues Fl 9, L22, W23, L25 and L26 had notably high TAD I-like pathogenicity scores (Figures 3F and 3K) and are highly conserved and essential for binding to other proteins (Raj and Attardi, 2017), including MDM2, which negatively regulates TP53 by blocking transactivation. Indeed, mapping the average TAD I-like expression score onto the structure of TAD I bound to MDM2 showed that residues at the interface between the two proteins had the highest average pathogenicity (Figure 3K), consistent with disruption ofMDM2 binding leading to p53-induced apoptosis (Chene, 2003). In contrast, residue P27, which is also hydrophobic and highly conserved and is essential for binding by the negative regulator adenoviral oncoprotein E1B, had a low average TAD I-like score, a high DBD-like score, and increased proliferation (Figure 3F). The P27Y mutation was shown to disrupt E1B binding while increasing MDM2 binding affinity (Lin et al., 1994), thus likely resulting in MDM2 sequestration and increased proliferation.
[0081] A majority of missense variants (661 / 1117, 59.2%) in DBD had high DBD-like expression scores and proliferation scores, with pathogenic variants predominantly in residues at the core of the DNA binding domain (Figure 3M). The four variants with TAD I-like signatures, S121F, S121A, S121Y and N239Y were at the periphery, towards the DNA binding interface (Figure 3M). The variants at S 121 have been previously characterized as ‘super’ mutants leading to increased apoptosis induction (Kakudo et al., 2005; Sailer, 1999), with S121F exhibiting differential specificity in activating the MDM2 promoter compared to G-rich promoters (Sailer, 1999), whereas N239Y was shown to increase the stability of the DNA-binding domain (Nikolovaet al., 1998). In our experiments, we observed lower MDM2 expression in cells expressing S121F and increased expression of PERP, a target of TP53 that promotes apoptosis (Attardi et al., 2000), in cells expressing N239Y, consistent with these effects (Figure 3K).
[0082] To get insights into the functional effects of 4D variants, we mapped both the DBD- and 4D-like expression scores onto the 4D tetramer assembly (Figure 3N). Residues with high average DBD-like expression scores were predominantly located at the interface of the four domains, with few outward-facing mutations having pathogenic effects. Conversely, residues with higher TAD I-like scores were specific to the antiparallel beta sheet and helix-helix interfaces between 4D dimers. These observations were consistent with prior studies of individual variant effects on p53 oligomerization (Fischer et al., 2016).
[0083] Finally, variants at residues D391, S392, and D393 at the C terminus, which are highly conserved across species (Laptenko et al., 2016) had elevated DBD-like expression and proliferation scores, similar to variants in 4D (Figure 3D). Phosphorylation of S392 has been shown to increase tetramer formation (Sakaguchi et al., 1997), and mutations at this residue would therefore likely mimic 4D mutations, as observed in our experiments.
[0084] Taken together, our results reveal two orthogonal axes of pathogenicity in TP53, with either TAD I-like apoptotic effects or DBD-like effects on cell proliferation. While these effects are overall domain specific, the scRNA-seq readout enabled the identification of diverging functional consequences of individual variants within each domain, in turn enabling a more accurate quantification of the degree of pathogenicity of DBD-like variants outside of the DBD. Our results also suggest that variants in the tetramerization domain are on average less pathogenic than variants in the DBD, despite perturbing the same transcriptional programs.TP53 variant classification performance
[0085] The characterization of distinct modes of pathogenicity in TP53, combined with the relatively large number of TP53 missense variants annotated as either pathogenic or benign in ClinVar (189 and 107, respectively) presented an opportunity to benchmark the performance of our assay against prior mutational scanning experimental experiments (Giacomelli et al., 2018) and computational predictions (Methods), including on distinguishing DBD- and TAD I-like pathogenicity. On the task of classifying ClinVar pathogenic vs. benign variants, our ensemble classifier based on both the proliferation and expression score (i.e., the probability of having eitherDBD- or TAD I— like pathogenicity) achieved the best performance (ROC AUC = 0.983), followed by the individual scores from our assay and predictions from AlphaMissense, PrimateAI-3D and other models, some of which were trained on ClinVar and other clinical variant databases or human allele frequencies (Figures 4A, 4B). The better performance of the ensemble classifier was also apparent in its separation of scores for pathogenic vs. benign variants (Figure 4C). While the difference between expression and proliferation was small (Figure 4A), the higher performance of the expression score nonetheless demonstrates the benefit of using expression profiles to quantify effects on proliferation. Our single cell-derived scores also achieved superior performance compared with the proliferation-based bulk assay from (Giacomelli et al., 2018), which was conducted under similar experimental conditions (i.e., overexpression of mutant TP53 alleles in A549 cells).
[0086] Next, we sought to compare classification performance for variants in TAD I, to assess how well different methods captured pathogenic effects related to apoptosis. Due to the absence of ClinVar variants in this domain, we benchmarked performance on the task of distinguishing between variants observed in the population (10) vs. variants that were never observed (134). While classification performance was worse overall compared with the ClinVar benchmark, our ensemble classifier also achieved the highest performance on this task (ROC AUC = 0.716; Figure 4D). The overall lower performance of computational predictors for TAD I variants matches our observations in the CDKN2A benchmark, where computational models generally achieved poorer performance predicting the pathogenicity of variants that affect protein-protein interactions.
[0087] While we observed strong agreement between our experimentally derived scores and the best computational predictors on the ClinVar benchmark, seven ClinVar pathogenic variants with high PrimateAI-3D were scored as benign by our ensemble classifier (Figure 4C). Three of these variants (Q165L, E171D, E326V) were scored as likely benign or ambiguous by AlphaMissense, and the remaining variants (S240N, G334W, R337L, L344P) were scored as benign in the bulk assay from (Giacomelli et al., 2018), raising the possibility that they may not have dominant negative pathogenic effects (only G334W and R337L were observed in COSMIC, in 11 and 69 individuals, respectively). Conversely, five variants annotated as benign in ClinVar with low PrimateAI-3D scores were scored as pathogenic by our ensemble classifier. Among these, three (P36Q, R110H, L252V) were only scored as borderline pathogenic by the ensemble classifier (0.50-0.55, corresponding to FDR 0.015-0.035) and one (S303G) had coverage below 99.55% ofall cDNA variants in the library, potentially producing a noisy estimate of its effects. The remaining variant, G108S, did not have significant effects on proliferation (q-value = 0.28), suggesting it may be a false positive in the expression score. Another variant, H296R, annotated as likely pathogenic in ClinVar, was classified as likely benign by both our assay (ensemble score 0.09) and PrimateAI-3D (percentile score 0.16), and was only observed in one individual in COSMIC. The other two amino acid variants (H296L, H296P) at the same position were also predicted to be benign (ensemble score < 0.31, PrimateAI-3D percentile score < 0.29), suggesting overall low pathogenicity of missense variants at position 296.
[0088] To assess how sequencing depth and the number of cells measured affected variant classification performance, we computed the expression score across a range of average numbers of transcripts captured per cell (as unique molecular identifiers (UMI)) and total cell numbers by randomly sampling from the full dataset (Methods). This showed that ~126k cells sequenced to a mean depth of ~10k UMIs were sufficient to achieve a performance on the ClinVar benchmark close to the full dataset (AUROC = 0.96 vs. 0.98; Figure 9A). Higher numbers of cells improved the correlation between the expression score and PrimateAI-3D predictions independent of sequencing depth (Figure 9B), likely due to better detection of underrepresented or depleted variants, such as TAD I variants leading to apoptosis (Figure 9C).
[0089] Taken together, these benchmarking results demonstrate state-of-the-art performance of scRNA-seq as a readout of variant pathogenicity, and enabled classifying all 559 VUS in ClinVar, with 232 as likely pathogenic and 327 as likely benign on the basis of the ensemble score, which captured both DBD- and TAD I-like pathogenicity.Distinguishing between two axes of pathogenicity in SOD1 improves classification of ALS variants
[0090] Understanding how SOD1 mutations lead to the development of ALS remains elusive. ALS-associated mutations have been identified in almost every region of SOD1 (Figure 5A) (Opie- Martin et al., 2022), with many mutations retaining partial or full dismutase activity (i.e., converting superoxide to water and hydrogen peroxide). No correlation has been observed between a reduction in dismutase activity, which is essential for cell survival (Tsherniak et al., 2017), and disease age of onset or rate of progression, leading to a consensus that disease likely arises from toxic gain-of-function effects from misfolding and / or intracellular aggregation rather than reducedenzymatic activity (Taylor et al., 2016). We sought to identify whether these two types of effects could be distinguished by their transcriptomic signature in our assay. We at first used the same experimental design as for p 16 and TP53 but did not detect any variant effects on gene expression or cell proliferation. Reasoning that this was likely due to the absence of dominant negative effects of pathogenic SOD1 variants combined with high endogenous expression, we knocked out endogenous SOD1 using CRISPR prior to lentiviral transduction of the variant library (Methods). The knockout efficiency, assayed using Sanger sequencing of the edited regions, was -77-100% across experiments (Methods). We obtained 92,465 cells expressing all 1,395 possible SOD1 SNVs using this approach, with 1,387 variants (99.4%) observed in at least 20 cells, corresponding to 1,106 of 1,113 possible amino acid variants (Figure 10A). Two-dimensional projection of the averaged expression profiles of each variant showed a gradient of pathogenicity based on PrimateAI-3D predictions with a matching enrichment of nonsense variants, but without the separation of pathogenic from benign variants that we observed for pl6 and TP53 (Figure 10B).
[0091] To identify the transcriptomic consequences of pathogenic variants, we computed residue position-level average expression profiles, which revealed domain-specific effects on gene expression that were associated with proliferation (Figure 5A). Due to the noisier effects on gene expression in these experiments, we used the position-level expression profiles to train a classifier to distinguish between synonymous- and nonsynonymous-like expression at the level of residue positions instead of individual amino acids (Figure 5A and Methods). The resulting expression score was correlated with proliferation (Spearman’s p = -0.59), with pathogenic effects broadly localized to the zinc binding and electrostatic loops of SOD1 (Figures 5 A and 5B), which are essential for dismutase function. Consistent with this, gene set enrichment analysis (Methods) showed that reactive oxygen species pathway genes were most enriched among genes upregulated by variants that decreased proliferation (module 1, Figures 5 A and 5C). Apoptosis-related pathways were also enriched, consistent with the essentiality of SOD1 (Tshemiak et al., 2017).
[0092] Based on an apparent discrepancy between our scores and the localization of ALS pathogenic variants, as well as predictions from PrimateAI-3D (Figure 5A), we hypothesized that the pathogenic effects we observed were limited to, or primarily captured, the disruption of enzymatic activity rather than ALS-associated toxic gain-of-function effects. To test this, we curated measurements of SOD1 enzymatic activity for 18 amino acid variants from the literature (all of which were associated with ALS pathogenicity), categorized the activity of these variantsinto three groups (Methods), and compared these to our proliferation scores (Figure 5D). This showed that most variants with reduced or disrupted enzymatic activity were indeed depleted in our assay. Notably, the proliferation score outperformed computational predictions on the task of distinguishing between variants with reduced vs. normal activity (ROC AUC = 0.958, Figure 10C), suggesting that computational models did not fully capture pathogenicity related to reduced dismutase activity.
[0093] Pathogenic effects detected in our assay were predominantly localized to parts of the zinc-binding and electrostatic loops and of the beta barrel closest to these domains, specifically the b4, b5, b7, and b8 sheets (Figures 5A and 5B). Mapping the proliferation score onto the 3D structure of SOD1 showed that the vast majority of residues with pathogenic effects were in close proximity to the catalytic site (Figure 5F). Observations that familial ALS-associated variants alter b-barrel folding and dimer contacts rather than catalysis (Bruijn et al., 1998; Deng et al., 1993) further suggest that the pathogenic effects captured in our experiments were largely independent of ALS-associated mechanisms. However, since we observed a nonlinear relationship between our scores and PrimateAI-3D predictions (Figure 5E), we hypothesized that adjusting the PrimateAI- 3D predictions (Figure 5H) by regressing out the proliferation and expression scores (Figures 5F and 5G) may improve the classification of ALS pathogenic variants. Indeed, the adjusted PrimateAI-3D scores showed an enrichment of variants predicted to be pathogenic in the b-barrel (Figure 51) compared with the raw scores (Figure 5H).
[0094] We next sought to benchmark these adjusted scores on the task of distinguishing ALS- associated pathogenic variants from benign ones. However, the highly variable penetrance, age of onset, and disease course of ALS observed across SOD1 mutations (Opie-Martin et al., 2022) and absence of common SOD1 missense variants in the population hinder the identification of benign variants, with no such variants annotated in ClinVar. Moreover, SOD1 variants observed in human populations were predominantly very rare and contained a high proportion of pathogenic variants (40%; Figure 10D), precluding their use as putatively benign variants. We therefore benchmarked the scores on the task of distinguishing between ClinVar pathogenic variants and variants that were not annotated as potentially pathogenic in any other database or cohort (Abel et al., 2012; Chen et al., 2021; McCann et al., 2021; Opie-Martin et al., 2022; Ruffo et al., 2022; Yamashita and Ando, 2015), reasoning that the latter set would predominantly include a combination of benign and dismutase activity-altering variants, and thus that the residualized PrimateAI-3D score wouldoutperform others on this benchmark. Indeed, we observed that while proliferation alone (i.e., effects on enzymatic activity) was not informative for classifying ALS pathogenic variants (ROC AUC = 0.496), regressing out these effects from PrimateAI-3D and other computational predictions consistently improved ALS variant classification (Figures 5 J and 10E, ROC AUC = 0.728 for the residual score vs. 0.639 for the raw score), in particular for methods that were not trained on clinical variant annotations (Figure 10F).
[0095] In summary, these results suggest that our assay specifically captured pathogenic effects related to the disruption of SOD1 dismutase function. Due to an overlap between ALS- and dismutase-associated pathogenicity and a nonlinear relationship between these effects, the scores from our assay nonetheless enabled improving prediction of ALS-associated pathogenicity.DISCUSSION
[0096] Solving variants of uncertain significance requires understanding their functional consequences and the contexts leading to their pathogenicity. To date, efforts to characterize variants at scale have largely relied on deep mutational scanning experiments and computational models that produce one-dimensional predictions of pathogenicity. Here, we demonstrate that saturation mutagenesis combined with single cell sequencing can be used as a general-purpose variant interpretation assay with state-of-the-art pathogenicity prediction tied to a functional interpretation based on perturbations of the transcriptome. Importantly, our results show that the variant-level transcriptomic readout enables the identification of independent modes of pathogenicity linked to distinct functional mechanisms. For the three genes we characterized, we observed domain-specific modes of pathogenicity linked to disruption of protein-protein interactions (pl 6 and TP53) or that were independent of the primary mode of disease-associated pathogenicity (TP53 and SOD1). The poorer performance of computational models in our benchmarks is most likely due to an inability to distinguish between such effects, combined with the absence of protein-protein interactions in the datasets used to train these models.
[0097] Although our results demonstrate the power of using single-cell transcriptomes to characterize variant pathogenicity, the implementation we proposed here has several limitations. First, as our initial results with SOD1 suggest, simple overexpression of variant transcripts is likely to work only for variants and genes with dominant negative effects or low endogenous expression. Performing a knockout of the endogenous copies prior to transduction of the variant library is apossible solution but poses additional challenges — since variant effects occur at the protein level, turnover and replacement of endogenous proteins with variants need to be considered and may further complicate the experimental design. Several alternative approaches exist for introducing single variants into cells, including saturation genome editing, which has been used in proliferation-based deep mutational scanning experiments (Findlay et al., 2018), and base editing (Bock et al., 2022), but these do not yet enable profiling all possible variants in a gene. Second, while we showed that random variant barcoding combined with amplicon sequencing yields highly accurate variant calls, short reads currently limit the assay to the -65% of genes that are at most -1300-1500 nucleotides long. As long read technologies become more scalable this limitation will disappear; alternatively dual barcoding at the 3' and 5' UTRs should enable characterizing up to 90% of all protein coding genes. Third, as our results for SOD1 showed, detecting disease- associated modes of pathogenicity may not be possible in arbitrary cell lines and at the time scales (days to weeks) that saturation mutagenesis experiments are typically conducted in. However, these limitations are independent of the specific type of assay and readout used, and identifying and optimizing these conditions is beyond the scope of this work. Lastly, due to the high level of noise and sparsity in single cell gene expression profiles, hundreds of thousands of cells must be profiled to distinguish variant- and domain-specific pathogenic effects. While high costs of droplet-based single cell library preparation methods are currently a limiting factor, split-pool approaches that enable characterizing hundreds of thousands of cells in a single experiment (Bock et al., 2022), combined with rapidly diminishing costs of sequencing, will enable scaling these experiments to hundreds of genes.
[0098] Given its ability to distinguish between independent modes of pathogenicity combined with state-of-the-art variant classification performance, we anticipate that single cell profiling will rapidly become the standard readout for saturation mutagenesis experiments. Profiling large numbers of genes — hundreds will likely be required — will in turn enable training computational models that can predict different modes of pathogenicity together with their effects on the transcriptome. Shifting towards multi-dimensional characterization of variant pathogenicity will improve our understanding of gene function, molecular pathways, and the consequences of their perturbation in disease.Methods0099] Key resources tableConstruction of variant libraries
[0100] To generate variant libraries, plasmids containing wild type ORFs were obtained from GenScript. The ORFs were cloned into the pcDNA3.1(-) backbone (Thermo Fisher Scientific V79520) between the EcoRl and BamHl sites, with the addition of a Kozak sequence (GCCACC) before the start codon. The wild type sequences used were NM000546 for TP53, NM00007 for CDKN2A, and NM000454 for SOD1. The constructed plasmids were then submitted to Twist Bioscience for the subsequent synthesis of the variant libraries. Flanking sequences (upstream: gccctctagactcgagcggccgccactgtgctggatatctgcagaattcgccacc, downstream: ggatccgagctcggtaccaagcttaagtttaaacc) were incorporated from the pcDNA3.1(-) vector to serve as cloning handles. The libraries for TP53 and SOD1 contained all 3,546 and 1,395 possible SNVs, respectively.. However, the CDKN2A library failed to pass Twist Bioscience’s internal quality controls for variants at codon 5, 41, 61, and 63. Due to low coverage of variants around these positions (Figure 6A), we excluded variants at positions 5, 39-41, and 60-64.Construction of 3' UTR barcoded lentiviral expression vector
[0101] The lentiviral expression vector sequence from (Ursu et al., 2022) was modified by deleting the original Notl and BamHl sites and adding a Multiple Cloning Site (MCS) for library cloning. This lentiviral plasmid was synthesized by GenScript. To create the barcoded lentiviral vector, a 19 bp semi-random barcode (BNSWNSKMSVNBSWNMKYN) was inserted into the MCS between the BamHl and PacI sites using Gibson assembly. The semi-random sequence was designed to minimize homopolymer formation and AATAAA generation. The barcoded vector was transformed into E. coli and subsequently plated on agar plates. After overnight incubation, bacterial DNA from around 4 million colonies was then extracted to achieve high barcode diversity. To assess barcode diversity, we performed paired-end sequencing of a small amplicon covering the barcode region, resulting in over 3.8 million different variant barcodes.Cloning of variant libraries into barcoded lentiviral vector
[0102] The variant libraries of CDKN2A, TP53, and SOD1 were cloned into the barcoded lentiviral vector by restriction digestion and ligation between the Notl and BamHl sites on the MCS. The cloned libraries were transformed into E. coli and subsequently plated on agar plates. After overnight incubation, library DNA from over 4 million different colonies was extracted to ensure high barcode diversity and full representation of the variants, with amplicon sequencing as the quality control method. All variants synthesized by Twist Bioscience were successfully clonedinto the vector, resulting in over 750,000 different variant barcode-variant combinations for each gene.Lentiviral production and titering
[0103] Lentivirus production was carried out using the ViraPower™ Lentiviral Packaging Mix (ThermoFisher Scientific K497500) in the 293FT cell line (ThermoFisher Scientific R70007) following the manufacturer's protocol. The collected virus was concentrated 100-fold after precipitation with PEG-itTM Virus Precipitation Solution (System BioSciences LV810A-1) for 3 days. The titer of the lentiviral library was determined by transducing A549 cells (ATCC CCL- 185) with the lentivirus at different dilutions in 1-ml F12k media (ATCC 30-2004) in the presence of 6 ug / ml polybrene (Sigma H9268). After selection with 1.5 ug / ml puromycin (ThermoFisher AAJ672368EQ) for 10 days, the number of puromycin-resistant colonies was counted using crystal violet (Sigma 61135) staining. The viral titer used in this study was determined to be 1.35*107 IFU per ml for TP53, 1.88*107 IFU per ml for SOD1, and 1.29*107 IFU per ml for CDKN2A.Lentiviral transduction with variant libraries
[0104] A549 cells cultured in F12k medium with 10% FBS (Sigma) were seeded at around 10% confluency on day 1. On day 2, the packaged lentiviral stocks containing the variant libraries were added to the cells at a MOI ranging from 0.003 to 0.1, along with 6 ug / ml of polybrene (Sigma- Aldrich, H9268). The medium was replaced on day 3 to remove the polybrene. On day 4, 2 ug / ml of puromycin was added to the medium to select the successfully transduced cells. After 3 days of puromycin selection, the cells were recovered in medium without puromycin for 2-3 days before being harvested for lOx Genomics single-cell RNA-seq library preparation.Single-cell RNA-seq library preparation and sequencing
[0105] Standard loading single-cell RNA-seq libraries were prepared using the lOx Genomics Chromium Next GEM Single Cell 3' Kit v3.1 (PN- 1000268) or the Chromium Next GEM Single Cell 3' HT Kit v3.1 (PN-1000370) on the Chromium controller or Chromium X. For some libraries, the lOx Cellplex kit (PN-1000261) or MULTI-seq workflow (McGinnis et al., 2019) was used for barcoding and overloading, allowing for the identification of multiplets during GEM formation. The final libraries were sequenced on NovaSeq 6000 systems, generating a minimum of 40,000 reads on average per cell. The sequencing length was 28 cycles for read 1 and 150 cycles for read 2.Amplicon sequencing to link variants to 3’ UTR barcodes
[0106] cDNA samples were generated during the lOx library preparation, 10 ng of which were used as a template for a two-step PCR to connect variants with variant barcodes. In the first PCR, each reaction uses a reverse primer mixture consisting of an equal volume of four staggered oligos that annealed to the vector backbone in the 3' region of the barcode (i.e., 3BC B15 ON, 3BC B15 IN, 3BC B15 2N, 3BC B 15 3N). To cover the 1182, 471 and 465 nucleotide long sequences of TP53, CDKN2A and SOD1, respectively, with tiling 150 bp reads, 11 amplicons were required for TP53 and 4 amplicons for CDKN2A and SOD1. For each amplicon, a gene-specific forward primer was added separately to each PCR reaction (the specific sequences are provided below). The final concentration of primers in all PCRs was 0.25 uM. The PCR products were purified using 0.8x SPRIselect beads and eluted in 52.5 ul of lOmM Tris buffer solution, pH 8.5. In the second PCR, which is an indexing PCR, 5 ul of the purified products from the first PCR were used as the template, along with 10 ul of Nextera DNA Flex Dual Index adapters (Illumina: IDT for Illumina DNA / RNA UD indexes: 20027213) as primers. The PCR products were purified using lx SPRIselect beads and eluted in 27.5 ul of lOmM Tris buffer solution, pH 8.5. Both PCRs utilized 2x KAPA HiFi HotStart ReadyMix (Kapa KK2602). The first gene-specific PCR consisted of 16 cycles (95°C for 3 min, followed by 16 cycles of 95°C for 30 seconds, 60°C for 30 seconds, 72°C for 30 seconds, and a final extension at 72°C for 5 min), while the second indexing PCR consisted of 9 cycles using the same cycling conditions. The eluted PCR products were quantitated using a Bioanalyzer or Tapestation (Agilent). The amplicons were pooled for sequencing on NovaSeq S4 flow cells using 80PM as the loading concentration with the NovaSeq Xp workflow. To increase the diversity of the amplicons, the libraries were mixed with 30% PhiX control libraries (Illumina FC-110-3001) and sequenced for 150 cycles for Read 1 and 100 cycles for Read 2. A minimum of 20 million reads were generated for each amplicon.
[0107] The reverse PCR primers on the barcode side consisted of a mixture of staggered primers, with the following sequences used for all genes:The forward primers on the variant side had the following sequences:CRISPR knockout experiments
[0108] The SOD1 CRISPR knockout experiments utilized Synthego’s Gene Knockout Kit v2 following the manufacturer's protocols. Predesigned sgRNAs specific to SOD1 and SpCas9 2NLS nuclease were ordered from Synthego. A549 cells were seeded in culture dishes or plates and incubated to obtain 50-70% confluency at the time of transfection, using either electroporation or chemical transfection. For electroporation, the Lonza 4D-NucleofectorTM System with X unit and 4D-NucleofectorTM X Kit (32 RCT) suitable for A549 cells were used. Briefly, the harvested cells were washed and resuspended in the provided nucleofector solution. Ribonucleoprotein (RNP) was formed using the sgRNA and SpCas9 nuclease, then mixed with the cell suspension. The mixture was then electroporated using the specific electroporation parameters designated by the Lonza system. In parallel, chemical transfection was performed using LipofectamineTM CRISPRMAXTM. As in the electroporation method, the harvested cells were subjected to a washing step before being resuspended. RNP -transfection solution prepared from Opti-MEM medium, LipofectamineTM CRISPRMAXTM reagent, and the formed RNP was added to the cells and incubated for 3 days to allow for transfection and recovery. The efficiency of SOD1 gene knockout was evaluated by PCR-amplifying the edited region and Sanger sequencing. The .ablfiles from Sanger sequencing from the knockout and control groups were submitted to Synthego’s ICE analysis tool to calculate a knockout score, which represents the proportion of cells that have either a frameshift-inducing indel or a >21bp indel in the protein-coding region and are likely to induce complete loss of function. The knockout scores for the SOD1 experiments ranged from 77- 100%.Variant barcode mapping from 3’ scRNA-seq libraries
[0109] To assign variant barcodes to cells, the FASTQ files generated by CellRanger (mkfastq subprogram) were parsed for reads containing the nucleotide sequence upstream of the variant barcode. These reads were then filtered to ensure that the variant barcode sequence matched the template BNSWNSKMSVNBSWNMKYN as well as the downstream sequence, allowing for up to two mismatches. For each resulting cell barcode-UMI pair, the corresponding reads were checked for concordance with a single variant barcode, allowing up to two mismatches. UMIs with multiple reads and discordant variant barcodes (i.e., likely chimeric molecules created during library preparation) were discarded (typically < 0. 1% of UMIs).Variant calling from amplicon sequencing
[0110] We developed a custom pipeline to call variants from amplicon sequencing. Each amplicon was first processed individually, starting from FASTQs, with read 1 containing the targeted portion of coding sequence and read 2 containing the variant barcode. The variant barcode was parsed from read 2 by determining the offset of the staggered primer sequence and checking the variant barcode and downstream sequences for concordance with the reference template, allowing up to one mismatch. Next, we appended the variant barcode to the read ID in read 1, and aligned this new FASTQ to the wild type reference sequence using Bowtie 1.3.1. From these alignments, we generated pileups of all reads corresponding to a variant barcode, and computed alternate allele frequencies and counts for variants found in at least 3 reads with base quality > 37. Next, we aggregated variant calls across all amplicons, retaining variant barcodes only if they were observed in all amplicons, and if they contained at least one variant (the saturation mutagenesis libraries contained between -1-22% of wild type sequences). Variant calls were required to match in regions of overlap between amplicons. Finally, we filtered variants on allele frequency (>0.6) and total coverage (>5 reads) to minimize potential template switching artifacts that may have arisen during PCR.Validation of amplicon sequencing with long reads
[0111] To test the accuracy of amplicon sequencing, we sequenced the TP53 library using PacBio HiFi long reads, produced using circular consensus sequencing. We linearized the TP53 variant library, which included the random variant barcode and mutagenized coding sequences, using Not I digestion of the plasmid DNA, and submitted this for sequencing on the PacBio Sequel® platform to Azenta Life Sciences. We only used reads that contained the full coding and barcode sequences, excluding reads with any insertions or deletions. We obtained HiFi reads matching 135,589 variant barcodes for which we observed a single variant (94.8%) or the WT sequence (5.2%) with amplicon sequencing. Of these, 132,382 (97.6%) matched the amplicon sequencing data. After filtering for the 64,206 variant barcodes that were observed in at least two HiFi reads, the concordance increased to 99.3% (63,773 matching reads).Single cell RNA-seq expression quantification and quality control
[0112] All scRNA-seq experiments were processed using CellRanger 6.1.2 (lOx Genomics) to generate gene expression matrices for downstream analyses. We processed the UMI count matrices generated by CellRanger with Scanpy (Wolf et al., 2018), first normalizing counts in each cell to 10,000, followed by loglp-transformation (i.e., generating log(CP10k+l) units). We annotated the cell cycle phase of each cell as described below. Next, we selected genes with variable expression across cells using scanpy. pp.highly_variable_genes, resulting in 1,690 genes for CDKN2A and 1,797 genes for TP53. For SOD1 we selected the top 5,000 variable genes using the 'seurat_v3' option. We regressed out the total read count and experiment from the expression matrix of variable genes, treating each experiment as a one-hot encoded batch covariate. Due to the roles of pl6 and p53 in regulating cell cycle progression, we did not include cell cycle as part of the covariate correction applied during preprocessing of the single cell gene expression data. Correcting for cell cycle did not modify the results for SOD1, and we therefore also omitted cell cycle from the covariates for SOD1. After covariate correction, gene expression was transformed to z-scores. All analyses were performed using the resulting normalized gene expression matrices with highly variable genes, unless specified otherwise.
[0113] To exclude cells with outlier gene expression driven by quality issues rather than variant effects, we performed Leiden clustering with a resolution of 1, based on a neighborhood graph with 15 neighbors. This overclustered the cells, enabling the identification and exclusion of a small number of outlier clusters that had low numbers of UMIs / cell, low numbers of cells, a highpercentage of mitochondrial reads (>20% on average), or only represented cells from a minority of experiments.Cell cycle annotation
[0114] To assign each cell to a cell cycle phase, we adopted the same strategy as (Ursu et al., 2022), based on marker genes for Gl / S, S, G2 / M, M, and M / Gl from (Macosko et al., 2015). Briefly, we computed the average expression of marker genes (in log(CP10k+l)) for each phase in each cell, filtering out genes with low correlation (Pearson r < 0.3) to the mean, followed by standardization. The phase with the highest score was assigned to each cell; cells with all scores < 0 were assigned to GO.Gene module definitions and gene set enrichment analyses
[0115] Gene set enrichment analyses were performed by first obtaining unsupervised clusters of genes based on the residue position-level average expression profiles. We used Leiden clustering with 5 nearest neighbors and the resolution selected for each gene to yield clusters or ‘gene modules’ with interpretable differences in the pathway enrichment results (0.2 for CDKN2A, 0.4 for TP53, and 0.1 for SOD1). For each module, we computed gene set enrichments with Enrichr (Kuleshov et al., 2016) and GSEApy (Fang et al., 2023), using the MSigDB_Hallmark_2020 gene set and retaining enrichments with adjusted p-value < 0.05 for further analysis.Proliferation scores
[0116] Proliferation scores for missense and nonsense variants were computed as the ratio of cells containing a given variant over its abundance in the library, normalized using synonymous variants. This was performed by aggregating the number of cells for each variant across all available experiments. The cell counts were then scaled by the median number of barcodes for synonymous variants in the library, divided by the median number of cells per synonymous variant. To identify variants with a statistically significant change in proliferation, we first computed an empirical p-value for each missense or nonsense variant based on the distribution of proliferation scores for synonymous variants. We then used these p-values to compute q-values (Storey and Tibshirani, 2003) with the ‘lambda’ parameter set to 0.5 (i.e., using the p-value interval [0.5, 1] to estimate the proportion of true null p-values, pO), and applied a 0.05 FDR threshold.Expression-derived pathogenicity scores
[0117] To generate pathogenicity scores from expression, we performed unsupervised clustering and trained supervised models to predict cluster labels.
[0118] To generate the cluster labels for CDKN2A, we excluded amino acid variants with insufficient library coverage and low cell counts (<30 cells) and then performed UMAP embedding on missense and synonymous variants using the first 50 PCs of the amino acid-level average gene expression profiles computed using all available cells for each variant. We used 20 nearest neighbors and a minimum distance between embedded points of 1. We then applied Leiden clustering with a resolution of 0.2 on the UMAP neighborhood graph, resulting in two distinct clusters of variants, which we labeled as ‘Synonymous’ and ‘Non-synonymous’, respectively, given that the former was grouped with synonymous variants. For TP53, we calculated average expression profiles for each amino acid variant using all available cells. These profiles were then averaged by position and further averaged within a 12A radius in the TP53 AlphaFold structure. We then performed UMAP embedding as described for CDKN2A, but with a minimum distance of 1.5, followed by Leiden clustering with a resolution of 0.2 to identify variant clusters. Clusters were labeled using majority voting; if a cluster predominantly contained synonymous variants, all variants within it were labeled as ‘synonymous’, and if the majority represented the DBD, the cluster was labeled as ‘DBD’. This process resulted in four distinct classes: Synonymous, TAD I, DBD, and 4D variants. For SOD1, we followed the same procedure but limited averaging across neighboring residues in the SOD1 AlphaFold structure to a radius of 6 A. We performed UMAP embedding as described for CDKN2A, followed by Leiden clustering with a resolution of 0.25, resulting in three clusters of variants. We labelled the cluster overlapping synonymous variants as ‘synonymous’, and the other two as ‘non-synonymous’.
[0119] Next, we trained XGBoost models separately for each gene with hyperparameter selection to predict variant classes. We used averaged expression of amino acid variants from all available cells for CDKN2A and TP53 as input, and augmented the synonymous class with wildtype samples (averaged across randomly selected, non-overlapping samples of 200 cells). For SOD1, due to the higher level of noise, we used the expression averaged across neighboring residues as the input. We held out 10% of the data for testing. For the remaining 90%, we trained an XGBoost model with hyperparameter selection with five-fold cross validation using multi :softprob as the objective. To utilize the gradient in the expression profile, we added sample weighting when training the TP53 model. Sample weights for missense variants were calculated based on their distance from the center of the first two principal components of synonymous variants divided by the median distance of all missense variants from the synonymous center,ensuring higher weights for variants farther away from synonymous variants. WT and synonymous samples were assigned a sample weight of 1.
[0120] We selected the best model based on minimal multi-class log loss of the cross-validation runs. Predictions for the held-out test set, nonsense variants, and missense variants not included in training (CDKN2A variants with fewer than 30 cells) were averaged across five models from the five folds. We refer to the predicted probability of being in a class as the expression score for that class. For CDKN2A and SOD1, the expression score is the probability of being non-synonymous. For TP53, in addition to expression scores for TAD I, DBD, and 4D, a general expression score was defined as 1 - probability of being synonymous.
[0121] We computed the ensemble score (Expression & Proliferation) by normalizing the proliferation score to the unit interval, adding it to the expression score, and normalizing the resulting sum to the unit interval. Since TP53 TAD I variants were associated with reduced proliferation, we instead started by normalizing the negative proliferation score to the unit interval for these variants.
[0122] To select variants with expression scores or the ensemble scores that differed significantly from synonymous variants, we calculated an empirical p-value for each variant based on the distribution of scores for synonymous variants, computed adjusted p-values using Benjamini-Hochberg to correct for multiple testing, and classified variants with adjusted p-values < 0.05 as pathogenic.Variant annotations
[0123] We used HGVS conventions (den Dunnen et al., 2016) for all variant descriptions, omitting the ‘p’ prefix for amino acid variants for simplicity.Benchmarking
[0124] ClinVar annotations for the three genes were downloaded from ncbi.nlm.nih dot gov / clinvar / on 09 / 30 / 2023 and filtered for coding SNVs, excluding VUSs. We also excluded variants with SpliceAI scores > 0.5 (Jaganathan et al., 2019), for example the c.457G>T mutation in CDKN2A (Rutter et al., 2003), since pathogenic effects from alternative splicing were not captured in our experiments. For all benchmarks, we merged variants annotated as ‘Pathogenic’, ‘Pathogenic / Likely pathogenic’ or ‘Likely pathogenic’ into a single ‘Pathogenic’ category, and similarly, variants annotated as ‘Benign’, ‘Benign / Likely benign’, ‘Likely benign’ into a single ‘Benign’ category.
[0125] Predictions were evaluated using the rank scores downloaded from the database for functional prediction dbNSFP4.2a (Liu et al., 2020) for the following methods: DEOGEN2 (Raimondi et al., 2017), MutationAssessor (Reva et al., 2011), PROVEAN converted (PROVEAN) (Choi et al., 2012), SIFT_converted (SIFT) (Ng and Henikoff, 2001), SIFT4G_converted (SIFT4G) (Vaser et al., 2016), MutPred (B. Li et al., 2009), MPC (Samocha et al., 2017), LIST-S2 (Malhis et al., 2020), LRT_converted (LRT) (Chun and Fay, 2009), DANN (Quang et al., 2015), MutationTaster_converted (Mutation Taster) (Steinhaus et al., 2021), M-CAP (Jagadeesh et al., 2016), fathmm-MKL_coding (fathmm-MKL) (Shihab et al., 2015), fathmm- XF_coding (fathmm-XF) (Rogers et al., 2018) , BayesDel_noAF (BayesDel) (Feng, 2017), REVEL (loannidis et al., 2016), CADD raw (CADD)(Rentzsch et al., 2021), VEST4 (Carter et al., 2013b), MetaLR and MetaSVM (Dong et al., 2015). We also benchmarked five scores that were not in the dbNSFP database, including PrimateAI-3D (Gao et al., 2023), AlphaMissense (Cheng et al., 2023), EVE and EVE VAE (Frazer et al., 2021), and ESMlv (Meier et al., 2021). Due to the unavailability of full mutation effect predictions of the ESMlv model, we reproduced ESMlv scores by applying the ESMlv model from githubdot com / facebookresearch / esm with pretrained-weights to all human protein sequences.
[0126] For CDKNA, 37 missense variants were annotated as pathogenic in ClinVar, and 4 missense variants as benign. We excluded two pathogenic variants due to insufficient library coverage, and omitted D153Y / c.457G>T based on having a SpliceAI score of 0.76. To expand the set of benign variants, we included variants observed across human populations in gnomAD, TOPMed, and UK Biobank. Since most of these variants were rare (MAE < 0.001), we first compared the performance of our expression and proliferation scores, PrimateAI-3D, and AlphaMissense across a range of allele frequency thresholds for selecting benign variants (i.e., variants above the threshold), on the task of distinguishing pathogenic variants from benign ones using the area under the receiver operating characteristic curve (AUROC). Since the performance of the four scores was largely consistent across allele frequency thresholds (Figure 7F), and to maximize the set of putatively benign variants, we fixed the threshold at 5e-5, adding 7 variants to the benign set, to comprehensively benchmark a wide range of computational scores (of the 10 variants with allele frequencies > 5e-5, 3 were already in ClinVar and one was excluded due to low coverage in the library). MutationAssessor, LRT, DeepSequence, EVE, and MutPred were excluded from the analysis due to a high number (> 27) of missing values for both pathogenic andbenign variants.
[0127] For TP53, the ClinVar annotation contained 189 pathogenic and 107 benign missense variants. One pathogenic variant (G105V) with SpliceAI score > 0.5 was excluded from the analysis. Due to differences between the number of variants scored across computational methods, we performed two separate benchmarks: one with better coverage of ClinVar variants, excluding MutPred, DeepSequence, and EVE due to a high number (> 22) of variants with missing scores, and another with better coverage of methods, but on a smaller common set of 182 pathogenic and 66 benign variants. Since no ClinVar variants were annotated in TAD I, we evaluated methods on the task of distinguishing between TAD I variants observed in the population (n = 10) as a proxy for likely benign variants, from those not observed in the population (n = 134) as proxy for likely pathogenic variants.
[0128] For SOD1, we evaluated two scores for each model: the raw score and a residualized score obtained by regressing out the proliferation and expression scores from the raw score. Due to the absence of benign variants in ClinVar, we benchmarked scores on the task of distinguishing between 56 ClinVar pathogenic variants and 692 variants that were not classified as pathogenic in other databases or cohorts, as described in the main text. We further excluded variants that were not scored by all the included methods or that had a SpliceAI score > 0.5, resulting in a common set of 56 pathogenic and 675 benign variants.
[0129] To assess statistical differences in AUROCs for each set of benchmarks, we conducted bootstrapping with replacement (sampling 1,000 times) on the shared set of pathogenic and benign / non-pathogenic variants used in each benchmark. Subsequently, Mann-Whitney U tests were used to identify significant differences in AUROCs between pairs of scores.TP53 downsampling analysis
[0130] To benchmark variant classification performance as a function of total cell numbers and sequencing depth, we randomly sampled the TP53 dataset across a range of mean UMIs per cell (by sampling read counts from a binomial distribution) and cell numbers. We computed an expression score for each downsampling experiment as described in the ‘Expression-derived pathogenicity scores’ section above, with the following exceptions to streamline the analysis: 1) the optimal number of neighbors for UMAP embeddings (10, 15 or 20) and Leiden clustering resolution (ranging from 0.2 to 1 in steps of 0.2) were selected based on the highest concordance, measured by adjusted Rand index, between clusters and TP53 domain annotations; 2) WT sampleswere not included in training; 3) no sample weighting was applied; and 4) the same set of XGBoost hyperparameters was used for training all downsampled datasets.
[0131] For the ClinVar benchmark, we used the common set of variants shared among all downsampling experiments. When calculating Spearman’s correlations between expression and PrimateAI-3D scores, we included all missense variants to account for variants observed in fewer cells, such as variants in TAD I.Literature-based annotation of SOD1 enzymatic activity
[0132] To test for a correlation between variant effects detected in our assay and effects on SOD1 enzymatic activity, we searched the literature for prior studies that measured the enzymatic activity of individual amino acid variants (Borchelt et al., 1994; Fahmy et al., 2023; Gurney et al., 1994; Hayward et al., 2002; Liochev et al., 1998; Stathopulos et al., 2003). Since different assays were used to measure activity across studies (e g., pulse radiolysis, gel-based assays, immunoblots), we categorized the enzymatic activity of each variant as normal, reduced, strongly reduced, or ambiguous if there was conflicting evidence (e.g., A5V). We found measures of enzymatic activity for 20 variants, two of which had ambiguous effects and were excluded from further analyses.
[0133] REFERENCESAbel, O., Powell, J.F., Andersen, P.M., Al-Chalabi, A., 2012. ALSoD: A user-friendly online bioinformatics tool for amyotrophic lateral sclerosis genetics. Hum. Mutat. 33, 1345— 1351.Adzhubei, I., Jordan, D M., Sunyaev, S.R., 2013. Predicting functional effect of human missense mutations using PolyPhen-2. Curr. Protoc. Hum. Genet. 76.Attardi, L.D., Reczek, E.E., Cosmas, C., Demicco, E.G., McCurrach, M.E., Lowe, S.W., Jacks, T., 2000. PERP, an apoptosis-associated target of p53, is a novel member of the PMP- 22 / gas3 family. Genes Dev. 14, 704-718.Bock, C., Datlinger, P., Chardon, F., Coelho, M.A., Dong, M B., Lawson, K.A., Lu, T., Maroc, L., Norman, T.M., Song, B., Stanley, G., Chen, S., Garnett, M., Li, W., Moffat, J., Qi, L.S., Shapiro, R.S., Shendure, J., Weissman, J.S., Zhuang, X., 2022. High-content CRISPR screening. Nat. Rev. Methods Primers 2.Boettcher, S., Miller, P.G., Sharma, R., McConkey, M., Leventhal, M., Krivtsov, A.V., Giacomelli, A.O., Wong, W , Kim, J., Chao, S., Kurppa, K.J., Yang, X., Milenkowic, K., Piccioni, F., Root, D.E., Rucker, F.G., Flamand, Y., Neuberg, D., Lindsley, R.C., Janne, P.A., Hahn, W.C., Jacks, T., Dohner, H., Armstrong, S.A., Ebert, B.L., 2019. A dominant-negative effect drives selection of TP53 missense mutations in myeloid malignancies. Science 365, 599-604.Borchelt, D.R., Lee, M.K., Slunt, H.S., Guarnieri, M., Xu, Z.S., Wong, P.C., Brown, R.H., Jr, Price, D.L., Sisodia, S.S., Cleveland, D.W., 1994. Superoxide dismutase 1 with mutationslinked to familial amyotrophic lateral sclerosis possesses significant activity. Proc. Natl. Acad. Sci. U. S. A. 91, 8292-8296. doi dot org / 10.1073 / pnas.91.17.8292Brandes, N., Goldman, G., Wang, C.H., Ye, C.J., Ntranos, V., 2023. Genome-wide prediction of disease variant effects with a deep protein language model. Nat. Genet. 55, 1512-1522.Bruijn, L.I., Houseweart, M.K., Kato, S., Anderson, K.L., Anderson, S.D., Ohama, E., Reaume, A.G., Scott, R.W., Cleveland, D.W., 1998. Aggregation and motor neuron toxicity of an ALS-linked SOD1 mutant independent from wild-type SOD1. Science 281, 1851-1854.Carter, H., Douville, C., Stenson, P.D., Cooper, D.N., Karchin, R., 2013a. Identifying Mendelian disease genes with the Variant Effect Scoring Tool. BMC Genomics 14.Carter, H., Douville, C., Stenson, P.D., Cooper, D.N., Karchin, R., 2013b. Identifying Mendelian disease genes with the variant effect scoring tool. BMC Genomics 14 Suppl 3, S3.Chen, L.-X., Xu, H.-F., Wang, P.-S., Yang, X.-X., Wu, Z.-Y., Li, H.-F., 2021. SOD1 mutation spectrum and natural history of AES patients in a 15-year cohort in southeastern China. Front. Genet. 12.Chene, P., 2003. Inhibiting the p53-MDM2 interaction: an important target for cancer therapy. Nat. Rev. Cancer 3, 102-109.Cheng, J., Novati, G., Pan, J., Bycroft, C., Zemgulyte, A., Applebaum, T., Pritzel, A., Wong, L.H., Zielinski, M., Sargeant, T., Schneider, R.G., Senior, A.W., Jumper, J., Hassabis, D., Kohli, P., Avsec, Z., 2023. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381.Choi, Y., Sims, G.E., Murphy, S., Miller, J.R., Chan, A.P., 2012. Predicting the functional effect of amino acid substitutions and indels. PLoS One 7, e46688.Chun, S., Fay, J.C., 2009. Identification of deleterious mutations within three human genomes. Genome Res. 19, 1553-1561.Cleveland, D.W., Rothstein, J.D., 2001. From charcot to lou gehrig: deciphering selective motor neuron death in als. Nat. Rev. Neurosci. 2, 806-819.Dabrowski, M., Bukowy-Bieryllo, Z., Zietkiewicz, E., 2015. Translational readthrough potential of natural termination codons in eucaryotes— The impact of RNA sequence. RNA Biol. 12, 950-958. den Dunnen, J.T., Dalgleish, R., Maglott, D.R., Hart, R.K., Greenblatt, M.S., McGowan-Jordan, J., Roux, A.-F., Smith, T., Antonarakis, S.E., Taschner, P.E.M., on behalf of the Human Genome Variation Society (HGVS), the Human Variome Project (HVP), and the Human Genome Organisation (HUGO), 2016. HGVS recommendations for the description of sequence variants: 2016 update. Hum. Mutat. 37, 564-569.Deng, H.X., Hentati, A., Tainer, J.A., Iqbal, Z., Cayabyab, A., Hung, W.Y., Getzoff, E.D., Hu, P., Herzfeldt, B., Roos, R.P., 1993. Amyotrophic lateral sclerosis and structural defects in Cu,Zn superoxide dismutase. Science.Dong, C., Wei, P., Jian, X., Gibbs, R., Boerwinkle, E., Wang, K., Liu, X., 2015. Comparison and integration of deleteriousness prediction methods for nonsynonymous SNVs in whole exome sequencing studies. Hum. Mol. Genet. 24, 2125-2137.Fahmy, N., Muller, K., Andersen, P.M., Marklund, S.L., Otto, M., Ludolph, A.C., Hamdi, N., 2023. A novel homozygous p.Ser69Pro SOD1 mutation causes severe young-onset ALS with decreased enzyme activity. J. Neurol. 270, 1770-1773.Fang, Z., Liu, X., Peltz, G., 2023. GSEApy: a comprehensive package for performing gene set enrichment analysis in Python. Bioinformatics 39.Feng, B.-J , 2017. PERCH: A unified framework for disease gene prioritization. Hum. Mutat. 38, 243-251.Findlay, G.M., Daza, R.M., Martin, B., Zhang, M.D., Leith, A.P., Gasperini, M., Janizek, J.D., Huang, X., Starita, L.M., Shendure, J., 2018. Accurate classification of BRCA1 variants with saturation genome editing. Nature 562, 217-222.Fischer, N.W., Prodeus, A., Malkin, D., Gariepy, J., 2016. P53 oligomerization status modulates cell fate decisions between growth, arrest and apoptosis. Cell Cycle 15, 3210-3219.Fowler, D M., Fields, S., 2014. Deep mutational scanning: a new style of protein science. Nat. Methods 11, 801-807.Frazer, J., Notin, P., Dias, M., Gomez, A., Min, J.K., Brock, K., Gal, Y., Marks, D.S., 2021. Disease variant prediction with deep generative models of evolutionary data. Nature 599, 91-95.Gao, H., Hamp, T., Ede, J., Schraiber, J.G., McRae, J., Singer-Berk, M., Yang, Y., Dietrich, A.S.D., Fiziev, P.P., Kuderna, L.F.K., Sundaram, L., Wu, Y , Adhikari, A., Field, Y., Chen, C., Batzoglou, S., Aguet, F., Lemire, G., Reimers, R., Balick, D., Janiak, M.C., Kuhlwilm, M., Orkin, J.D., Manu, S., Valenzuela, A., Bergman, J., Rousselle, M., Silva, F.E., Agueda, L., Blanc, J., Gut, M., de Vries, D., Goodhead, I., Harris, R.A., Raveendran, M., Jensen, A., Chuma, I.S., Horvath, J.E., Hvilsom, C., Juan, D., Frandsen, P., de Melo, F.R., Bertuol, F., Byrne, H., Sampaio, I., Farias, I., do Amaral, J.V., Messias,M., da Silva, M.N.F., Trivedi, M., Rossi, R., Hrbek, T., Andriaholinirina, N., Rabarivola, C.J., Zaramody, A., Jolly, C.J., Phillips-Conroy, J., Wilkerson, G., Abee, C., Simmons, J.H., Fernandez-Duque, E., Kanthaswamy, S., Shiferaw, F., Wu, D., Zhou, L., Shao, Y., Zhang, G., Keyyu, J.D., Knauf, S., Le, M.D., Lizano, E., Merker, S., Navarro, A., Bataillon, T., Nadler, T., Khor, C.C., Lee, J., Tan, P., Lim, W.K., Kitchener, A.C., Zinner, D., Gut, I., Melin, A., Guschanski, K., Schierup, M.H., Beck, R.M.D., Umapathy, G , Roos, C., Boubli, J.P., Lek, M., Sunyaev, S., O’Donnell-Luria, A., Rehm, H.L., Xu, J., Rogers, J., Marques-Bonet, T., Farh, K.K.-H., 2023. The landscape of tolerated genetic variation in humans and primates. Science 380.Giacomelli, A.O., Yang, X., Lintner, R.E., McFarland, J.M., Duby, M., Kim, J., Howard, T.P., Takeda, D.Y., Ly, S.H., Kim, E., Gannon, H.S., Hurhula, B., Sharpe, T., Goodale, A., Fritchman, B., Steelman, S., Vazquez, E, Tsherniak, A., Aguirre, A.J., Doench, J.G., Piccioni, F., Roberts, C.W.M., Meyerson, M., Getz, G., Johannessen, C.M., Root, D.E., Hahn, W.C., 2018. Mutational processes shape the landscape of TP53 mutations in human cancer. Nat. Genet. 50, 1381-1387.Gil, J., Peters, G., 2006. Regulation of the INK4b-ARF-INK4a tumour suppressor locus: all for one or one for all. Nat. Rev. Mol. Cell Biol. 7, 667-677.Goldstein, A.M., Fraser, M.C., Struewing, J.P., Hussussian, C.J., Ranade, K., Zametkin, D.P., Fontaine, L.S., Organic, S.M., Dracopoli, N.C., Clark, W.H., Jr, Tucker, M.A., 1995. Increased risk of pancreatic cancer in melanoma-prone kindreds with / ? / 6INK4mutations.N. Engl. J. Med. 333, 970-975.Goutman, S.A., Hardiman, O., Al-Chalabi, A., Chio, A., Savelieff, M.G., Kiernan, M.C., Feldman, E.L., 2022. Emerging insights into the complex genetics and pathophysiology of amyotrophic lateral sclerosis. Lancet Neurol. 21, 465-479.Gurney, M.E., Pu, H., Chiu, A.Y., Dal Canto, M.C., Polchow, C.Y., Alexander, D.D., Caliendo, J., Hentati, A., Kwon, Y.W., Deng, H.X., 1994. Motor neuron degeneration in mice that express a human Cu,Zn superoxide dismutase mutation. Science 264, 1772-1775.Hafner, A., Bulyk, M.L., Jambhekar, A., Lahav, G., 2019. The multiple mechanisms that regulate p53 activity and cell fate. Nat. Rev. Mol. Cell Biol. 20, 199-210.Hayward, L.J., Rodriguez, J. A., Kim, J.W., Tiwari, A., Goto, J. J., Cabelli, D.E., Valentine, J.S., Brown, R.H., Jr, 2002. Decreased metallation and activity in subsets of mutant superoxide dismutases associated with familial amyotrophic lateral sclerosis. J. Biol. Chem. 277, 15923-15931.Hussussian, C.J., Struewing, J.P., Goldstein, A.M., Higgins, P.A., Ally, D.S., Sheahan, M.D., Clark, W.H., Jr, Tucker, M.A., Dracopoli, N.C., 1994. Germline pl6 mutations in familial melanoma. Nat. Genet. 8, 15-21.Ikediobi, O.N., Davies, H., Bignell, G., Edkins, S., Stevens, C., O’Meara, S., Santarius, T., Avis, T., Barthorpe, S., Brackenbury, L., Buck, G., Butler, A., Clements, J., Cole, J., Dicks, E., Forbes, S., Gray, K., Halliday, K., Harrison, R., Hills, K., Hinton, J., Hunter, C., Jenkinson, A., Jones, D., Kosmidou, V., Lugg, R., Menzies, A., Mironenko, T., Parker, A., Perry, J., Raine, K., Richardson, D., Shepherd, R., Small, A., Smith, R., Solomon, H., Stephens, P., Teague, J., Tofts, C., Varian, J., Webb, T., West, S., Widaa, S., Yates, A., Reinhold, W., Weinstein, J.N., Stratton, M.R., Futreal, P.A., Wooster, R., 2006. Mutation analysis of 24 known cancer genes in the NCI-60 cell line set. Mol. Cancer Ther. 5, 2606-2612. loannidis, N.M., Rothstein, J.H., Pejaver, V., Middha, S., McDonnell, S.K., Baheti, S., Musolf, A., Li, Q., Holzinger, E., Karyadi, D., Cannon-Albright, L.A., Teerlink, C.C., Stanford, J.L., Isaacs, W.B., Xu, J., Cooney, K.A., Lange, E.M., Schleutker, J., Carpten, J.D., Powell, I. J., Cussenot, O., Cancel-Tassin, G., Giles, G.G., Maclnnis, R.J., Maier, C., Hsieh, C.-L., Wiklund, F., Catalona, W.J., Foulkes, W.D., Mandal, D., Eeles, R.A., Kote- Jarai, Z., Bustamante, C.D., Schaid, D.J., Hastie, T., Ostrander, E.A., Bailey-Wilson, J.E., Radivojac, P., Thibodeau, S.N., Whittemore, A.S., Sieh, W., 2016. REVEL: An ensemble method for predicting the pathogenicity of rare missense variants. Am. J. Hum. Genet. 99, 877-885.Jagadeesh, K.A., Wenger, A.M., Berger, M.J., Guturu, H., Stenson, P.D., Cooper, D.N., Bernstein, J. A., Bejerano, G., 2016. M-CAP eliminates a majority of variants of uncertain significance in clinical exomes at high sensitivity. Nat. Genet. 48, 1581-1586.Jaganathan, K., Kyriazopoulou Panagiotopoulou, S., McRae, J.F., Darbandi, S.F., Knowles, D., Li, Y.I., Kosmicki, J.A., Arbelaez, J., Cui, W., Schwartz, G.B., Chow, E.D., Kanterakis, E., Gao, H., Kia, A., Batzoglou, S., Sanders, S.J., Farh, K.K.-H., 2019. Predicting splicing from primary sequence with deep learning. Cell 176, 535-548. e24.Kakudo, Y., Shibata, H., Otsuka, K., Kato, S., Ishioka, C., 2005. Lack of correlation between p53-dependent transcriptional activity and the ability to induce apoptosis among 179 mutant p53s. Cancer Res. 65, 2108—2114.Kimura, H., Paranal, R.M., Nanda, N., Wood, L.D., Eshleman, J.R., Hruban, R.H., Goggins, M.G., Klein, A.P., Roberts, N.J., The Familial Pancreatic Cancer Genome Sequencing Project, 2022. Functional CDKN2A assay identifies frequent deleterious alleles misclassified as variants of uncertain significance. Elife 11.Kuleshov, M.V., Jones, M.R., Rouillard, A.D., Fernandez, N.F., Duan, Q., Wang, Z., Koplev, S., Jenkins, S.L., Jagodnik, K.M., Lachmann, A., McDermott, M.G., Monteiro, C.D., Gundersen, G.W., Ma’ayan, A., 2016. Enrichr: a comprehensive gene set enrichment analysis web server 2016 update. Nucleic Acids Res. 44, W90-W97.Landrum, M.J., Lee, J.M., Benson, M., Brown, G., Chao, C., Chitipiralla, S., Gu, B., Hart, J., Hoffman, D., Hoover, J., Jang, W., Katz, K., Ovetsky, M., Riley, G , Sethi, A., Tully, R., Villamarin-Salomon, R., Rubinstein, W., Maglott, D.R., 2016. ClinVar: public archive of interpretations of clinically relevant variants. Nucleic Acids Res. 44, D862-D868. doi dot org / 10.1093 / nar / gkvl222Langmead, B., Trapnell, C., Pop, M., Salzberg, S.L., 2009. Ultrafast and memory-efficient alignment of short DNA sequences to the human genome. Genome Biol 10, R25. doi dot org / 10.1186 / gb-2009-l 0-3-r25Laptenko, O., Tong, D.R., Manfredi, J., Prives, C., 2016. The tail that wags the dog: How the disordered C-terminal domain controls the transcriptional activities of the p53 tumorsuppressor protein. Trends Biochem. Sci. 41, 1022-1034. doi dot org / 10.1016 / j.tibs.2016.08.011Li, B., Krishnan, V.G., Mort, M.E., Xin, F., Kamati, K.K., Cooper, D.N., Mooney, S.D., Radivojac, P , 2009. Automated inference of molecular mechanisms of disease from amino acid substitutions. Bioinformatics 25, 2744-2750. doi dot org / 10.1093 / bioinformatics / btp528Li, H., Handsaker, B., Wysoker, A., Fennell, T., Ruan, J., Homer, N., Marth, G., Abecasis, G., Durbin, R., 1000 Genome Project Data Processing Subgroup, 2009. The Sequence Alignment / Map format and SAMtools. Bioinformatics 25, 2078-2079. doi dot org / 10.1093 / bioinformatics / btp352Lin, J., Chen, J., Elenbaas, B., Levine, A.J., 1994. Several hydrophobic amino acids in the p53 amino-terminal domain are required for transcriptional activation, binding to mdm-2 and the adenovirus 5 E1B 55-kD protein. Genes Dev. 8, 1235-1246. doi dot org / 10.1101 / gad.8.10.1235Liochev, S.I., Chen, L.L., Hallewell, R.A., Fridovich, I., 1998. The familial amyotrophic lateral sclerosis-associated amino acid substitutions E100G, G93A, and G93R do not influence the rate of inactivation of copper- and zinc-containing superoxide dismutase by H2O2. Arch. Biochem. Biophys. 352, 237-239. doi dot org / 10.1006 / abbi.1998.0616Liu, X., Li, C., Mou, C., Dong, Y., Tu, Y., 2020. dbNSFP v4: a comprehensive database of transcript-specific functional predictions and annotations for human nonsynonymous and splice-site SNVs. Genome Med. 12, 103. doi dot org / 10.1186 / sl3073-020-00803-9Luecken, M.D., Theis, F.J., 2019. Current best practices in single-cell RNA-seq analysis: a tutorial. Mol. Syst. Biol. 15. doi dot org / 10.15252 / msb.20188746Macosko, E.Z., Basu, A., Satija, R., Nemesh, J., Shekhar, K., Goldman, M., Tirosh, I., Bialas, A.R., Kamitaki, N., Martersteck, E.M., Trombetta, J. J., Weitz, D.A., Sanes, J.R., Shalek, A.K., Regev, A., McCarroll, S.A., 2015. Highly parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell 161, 1202-1214. doi dot org / 10.1016 / j. cell.2015.05.002Malhis, N., Jacobson, M., Jones, S.J.M., Gsponer, J., 2020. LIST-S2: taxonomy based sorting of deleterious missense mutations across species. Nucleic Acids Res. 48, W154-W161. doi dot org / 10.1093 / nar / gkaa288Matreyek, K.A., Starita, L.M., Stephany, J. J., Martin, B., Chiasson, M.A., Gray, V.E., Kircher, M., Khechaduri, A., Dines, J.N., Hause, R.J., Bhatia, S., Evans, W.E., Relling, M.V., Yang, W., Shendure, J., Fowler, D.M., 2018. Multiplex assessment of protein variant abundance by massively parallel sequencing. Nat. Genet. 50, 874-882. doi dot org / 10.1038 / s41588-018-0122-zMcBride, K.A., Ballinger, M L., Killick, E., Kirk, J., Tattersall, M.H.N., Eeles, R.A., Thomas, D M., Mitchell, G., 2014. Li-Fraumeni syndrome: cancer risk assessment and clinical management. Nat. Rev. Clin. Oncol. 11, 260-271. doi dot org / 10.1038 / nrclinonc.2014.41McCann, E.P., Henden, L., Fifita, J. A., Zhang, K.Y., Grima, N., Bauer, D.C., Chan Moi Fat, S., Twine, N.A., Pamphlett, R., Kiernan, M.C., Rowe, D.B., Williams, K.L., Blair, I.P., 2021. Evidence for polygenic and oligogenic basis of Australian sporadic amyotrophic lateral sclerosis. J. Med. Genet. 58, 87-95. doi dot org / 10.1136 / j medgenet-2020- 106866McGinnis, C.S., Patterson, D.M., Winkler, J., Conrad, D.N., Hein, M.Y., Srivastava, V., Hu, J.L., Murrow, L.M., Weissman, J.S., Werb, Z., Chow, E.D., Gartner, Z.J., 2019. MULTI- seq: sample multiplexing for single-cell RNA sequencing using lipid-tagged indices. Nat. Methods 16, 619-626. doi dot org / 10.1038 / s41592-019-0433-8Meier, J., Rao, R., Verkuil, R., Liu, J., Sercu, T., Rives, A., 2021. Language models enable zeroshot prediction of the effects of mutations on protein function. bioRxiv. doi dot org / 10.1101 / 2021.07.09.450648Ng, P.C., 2003. SIFT: predicting amino acid changes that affect protein function. Nucleic Acids Res. 31, 3812-3814. doi dot org / 10.1093 / nar / gkg509Ng, P.C., Henikoff, S., 2001. Predicting deleterious amino acid substitutions. Genome Res. 11, 863-874. doi dot org / 10.1101 / gr.176601Nikolova, P.V., Henckel, J., Lane, D.P., Fersht, A.R., 1998. Semirational design of active tumor suppressor p53 DNA binding domain with enhanced stability. Proc. Natl. Acad. Sci. U. S. A. 95, 14675-14680. doi dot org / 10.1073 / pnas.95.25.14675Opie-Martin, S., lacoangeli, A., Topp, S.D., Abel, O., Mayl, K., Mehta, P.R., Shatunov, A., Fogh, I., Bowles, H., Limbachiya, N., Spargo, T.P., ALKhleifat, A., Williams, K.L., Jockel -B al sarotti, J., Bali, T., Self, W., Henden, L., Nicholson, G.A., Ticozzi, N., McKenna- Yasek, D., Tang, L., Shaw, P.J., Chio, A., Ludolph, A., Weishaupt, J.H., Landers, J.E., Glass, J.D., Mora, J.S., Robberecht, W., Van Damme, P., McLaughlin, R., Hardiman, O., van den Berg, L., Veldink, J.H., Corcia, P., Stevie, Z., Siddique, N., Silani, V., Blair, I.P., Fan, D.-S., Esselin, F., de la Cruz, E., Camu, W., Basak, N.A., Siddique, T., Miller, T., Brown, R.H., ALChalabi, A., Shaw, C.E., 2022. The SODl-mediated ALS phenotype shows a decoupling between age of symptom onset and disease duration. Nat. Commun. 13, 6901. doi dot org / 10.1038 / s41467-022-34620-yQuang, D., Chen, Y., Xie, X., 2015. DANN: a deep learning approach for annotating the pathogenicity of genetic variants. Bioinformatics 31, 761-763. doi dot org / 10.1093 / bioinformatics / btu703Raimondi, D., Tanyalcin, I., Ferte, J., Gazzo, A., Orlando, G., Lenaerts, T., Rooman, M., Vranken, W., 2017. DEOGEN2: prediction and interactive visualization of single amino acid variant deleteriousness in human proteins. Nucleic Acids Res. 45, W201-W206. doi dot org / 10.1093 / nar / gkx390Raj, N., Attardi, L.D., 2017. The transactivation domains of the p53 protein. Cold Spring Harb. Perspect. Med. 7, a026047. doi dot org / 10.1101 / cshperspect.a026047Rentzsch, P., Schubach, M., Shendure, J., Kircher, M., 2021. CADD-Splice-improving genomewide variant effect prediction using deep learning-derived splice scores. Genome Med. 13, 31. doi dot org / 10.1186 / S13073-021-00835-9Rentzsch, P., Witten, D., Cooper, G M., Shendure, J., Kircher, M., 2019. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Res. 47, D886- D894. doi dot org / 10.1093 / nar / gky 1016Replogle, J.M., Saunders, R.A., Pogson, A.N., Hussmann, J.A., Lenail, A., Guna, A., Mascibroda, L., Wagner, E.J., Adelman, K., Lithwick- Yanai, G., Iremadze, N., Oberstrass, F., Lipson, D., Bonnar, J.L., Jost, M., Norman, T.M., Weissman, J.S., 2022. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb- seq. Cell 185, 2559-2575.e28. doi dot org / 10.1016 / j.cell.2022.05.013Reva, B., Antipin, Y., Sander, C., 2011. Predicting the functional impact of protein mutations: application to cancer genomics. Nucleic Acids Res. 39, el 18. doi dot org / 10.1093 / nar / gkr407Rogers, M.F., Shihab, H.A., Mort, M., Cooper, D.N., Gaunt, T.R., Campbell, C., 2018. FATHMM-XF: accurate prediction of pathogenic point mutations via extended features. Bioinformatics 34, 511-513. doi dot org / 10.1093 / bioinformatics / btx536Rosen, D.R., Siddique, T., Patterson, D., Figlewicz, D.A., Sapp, P., Hentati, A., Donaldson, D., Goto, J., O’Regan, J.P., Deng, H.X., 1993. Mutations in Cu / Zn superoxide dismutase gene are associated with familial amyotrophic lateral sclerosis. Nature 362, 59-62. doi dot org / 10.1038 / 362059a0Ruffo, P., Perrone, B., Conforti, F.L., 2022. SOD-1 variants in amyotrophic lateral sclerosis: Systematic re-evaluation according to ACMG-AMP guidelines. Genes (Basel) 13, 537. doi dot org / 10.3390 / genesl3030537Russo, A. A., Tong, L., Lee, J.O., Jeffrey, P.D., Pavletich, N.P., 1998. Structural basis for inhibition of the cyclin-dependent kinase Cdk6 by the tumour suppressor pl 6INK4a. Nature 395, 237-243. doi dot org / 10.1038 / 26155Rutter, J.L., Goldstein, A.M., Davila, M.R., Tucker, M.A., Struewing, J.P., 2003. CDKN2A point mutations D153spl(c.457G>T) and IVS2+1G>T result in aberrant splice products affecting both pl6INK4a and pl4ARF. Oncogene 22, 4444-4448. doi dot org / 10.1038 / sj . one.1206564Sakaguchi, K., Sakamoto, H., Xie, D., Erickson, J.W., Lewis, M.S., Anderson, C.W., Appella, E., 1997. Effect of phosphorylation on tetramerization of the tumor suppressor protein p53. J. Protein Chem. 16, 553-556. doi dot org / 10.1023 / a: 1026334116189Sailer, E., 1999. Increased apoptosis induction by 12 IF mutant p53. EMBO J. 18, 4424-4437. doi dot org / 10.1093 / emboj / 18.16.4424Samocha, K.E., Kosmicki, J. A., Karczewski, K.J., O’Donnell-Luria, A.H., Pierce-Hoffman, E., MacArthur, D.G., Neale, B.M., Daly, M.J., 2017. Regional missense constraint improves variant deleteriousness prediction. bioRxiv. doi dot org / 10.1101 / 148353Schmidt, R., Steinhart, Z., Layeghi, M., Freimer, J.W., Bueno, R., Nguyen, V.Q., Blaeschke, F., Ye, C.J., Marson, A., 2022. CRISPR activation and interference screens decode stimulation responses in primary human T cells. Science 375. doi dot org / 10.1126 / science.abj 4008Shihab, H.A., Rogers, M.F., Gough, J., Mort, M., Cooper, D.N., Day, I.N.M., Gaunt, T.R., Campbell, C., 2015. An integrative approach to predicting the functional effects of noncoding and coding sequence variation. Bioinformatics 31, 1536-1543. doi dot org / 10.1093 / bioinformatics / btv009Stathopulos, P.B., Rumfeldt, J.A.O., Scholz, G.A., Irani, R.A., Frey, H.E., Hallewell, R.A., Lepock, J.R., Meiering, E.M., 2003. Cu / Zn superoxide dismutase mutants associated withamyotrophic lateral sclerosis show enhanced formation of aggregates in vitro. Proc. Natl. Acad. Sci. U. S. A. 100, 7021-7026. doi dot org / 10.1073 / pnas,1237797100Steinhaus, R., Proft, S., Schuelke, M., Cooper, D.N., Schwarz, J.M., Seelow, D., 2021. MutationTaster2021. Nucleic Acids Res. 49, W446-W451. doi dot org / 10.1093 / nar / gkab266Storey, J.D., Tibshirani, R., 2003. Statistical significance for genomewide studies. Proc. Natl.Acad. Sci. U. S. A. 100, 9440-9445. doi dot org / 10.1073 / pnas,1530509100Taylor, J.P., Brown, R.H., Jr, Cleveland, D.W., 2016. Decoding ALS: from genes to mechanism. Nature 539, 197-206. doi dot org / 10.1038 / nature20413Tsherniak, A., Vazquez, F., Montgomery, P.G., Weir, B.A., Kryukov, G., Cowley, G.S., Gill, S., Harrington, W.F., Pantel, S., Krill-Burger, J.M., Meyers, R.M., Ali, L., Goodale, A., Lee, Y., Jiang, G., Hsiao, J., Gerath, W.F.J., Howell, S., Merkel, E., Ghandi, M., Garraway, L.A., Root, D.E., Golub, T.R., Boehm, J.S., Hahn, W.C., 2017. Defining a cancer dependency map. Cell 170, 564-576. el6. doi dot org / 10.1016 / j . cell.2017.06.010Ursu, O., Neal, J.T., Shea, E., Thakore, P.I., Jerby-Arnon, L., Nguyen, L., Dionne, D., Diaz, C., Bauman, J., Mosaad, M.M., Fagre, C., Lo, A., McSharry, M., Giacomelli, A.O., Ly, S.H., Rozenblatt-Rosen, O., Hahn, W.C., Aguirre, A.J., Berger, A.H., Regev, A., Boehm, J.S., 2022. Massively parallel phenotyping of coding variants in cancer with Perturb-seq. Nat. Biotechnol. 40, 896-905. doi dot org / 10.1038 / s41587-021-01160-7Vaser, R., Adusumalli, S., Leng, S.N., Sikic, M., Ng, P C., 2016. SIFT missense predictions for genomes. Nat. Protoc. 11, 1-9. doi dot org / 10.1038 / nprot.2015.123Wolf, F.A., Angerer, P., Theis, F.J., 2018. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 19. doi dot org / 10.1186 / sl 3059-017- 1382-0Yamashita, S., Ando, Y., 2015. Genotype-phenotype relationship in hereditary amyotrophic lateral sclerosis. Transl. Neurodegener. 4. doi dot org / 10.1186 / s40035-015-0036-y
[0134] A number of embodiments have been described. Nevertheless, it will be understood that various modifications may be made. Accordingly, other embodiments are within the scope of the following claims.
Claims
What is claimed is:
1. A method of simultaneously determining the effects of a plurality of variants in a gene comprising: a) generating a mutant expression vector for each of the plurality of variants; b) expressing each variant in a cell in a plurality of single cells; and c) sequencing the RNA expressed in each single cell and evaluating expression patterns that result in each single cell to generate single cell RNA sequencing data.
2. The method of claim 1, wherein each mutant expression vector comprises a random barcode for each variant.
3. The method of any of the preceding claims, further comprising evaluating the single cell RNA sequencing data to determine the pathogenicity of each variant.
4. The method of any of the preceding claims, further comprising generating a cell barcode during reverse transcription, resulting in said barcode being coupled to the cDNA.
5. The method of any of the preceding claims, wherein generating the single cell RNA sequencing data comprises generating a plurality of amplicons that comprise both the variant and the random barcode.
6. The method of any of the preceding claims, wherein generating the single cell RNA sequencing data comprises generating a plurality of amplicons that comprise both the variant and the cell barcode.
7. The method of any of the preceding claims, wherein evaluating the single cell RNA sequencing data comprises use of machine learning.
8. The method of any of the preceding claims, variants are classified as either pathogenic, likely pathogenic, likely benign, or benign.
9. The method of any of the preceding claims, wherein each cell of the plurality of single cells is configured to express a single variant of the plurality of variants.
10. The method of any of the preceding claims, wherein each single variant of the plurality of variants is expressed in more than one cell, thereby creating redundancy of the single cell RNA sequencing data.
11. The method of any of the preceding claims, wherein the plurality of single cells comprises at least 30,000, at least 40,000, at least 50,000, at least 60,000, at least 70,000, at least80,000, at least 90,000 or at least 100,000, 150,000, 200,000 or 250,000 single cells.
12. The method of any of the preceding claims, wherein the plurality of variants comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1100, 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900, 2000, 2100, 2200, 2300, 2400, 2500, 2600, 2700, 2800, 2900 or at least 3000 variants.
13. The method of any of the preceding claims, wherein gene is selected from genes associated with rare disease.
14. The method of any of the preceding claims, wherein gene is selected from genes associated with oncology.
15. The method of any of the preceding claims, wherein the mutant expression vector is generated using CRISPR.
16. The method of any of the preceding claims, wherein the mutant expression vector is generated using CRISPRi.
17. The method of any of the preceding claims, wherein the mutant expression vector is generated using CRISPR.
18. The method of any of the preceding claims, wherein the RNA from each single cell is identified using barcoding and sequencing.
19. The method of any of the preceding claims, further comprising repeating steps a) - c) in more than one set of primary cells or cell lines.
20. A method of simultaneously determining the effects of a gene comprising: d) perturbing the function of a gene in a cell in a plurality of single cells; and e) sequencing the RNA expressed in each single cell and evaluating expression patternsthat result in each single cell to generate single cell RNA sequencing data.
21. The method of claim 20, further comprising evaluating the single cell RNA sequencing data to determine the pathogenicity of each perturbed gene.
22. The method of claim 21, wherein perturbing the function of a gene comprises using one or more drug compounds.
23. The method of claim 22, wherein a different drug compound is administered to each single cell.
24. The method of claim 22, wherein the plurality of drug compounds comprises at least 100, at least 200, at least 300, at least 500, at least 1,000, at least 2,000, at least 3,000 at least 4,000, at least 5,000, or at least 10,000 compounds25. The method of claim 20, wherein evaluating the single cell RNA sequencing data comprises use of machine learning.
26. A method of improving prediction of pathogenicity of one or more variants in a gene comprising: a) generating a mutant expression vector for each of the plurality of variants; b) expressing each variant in a cell in a plurality of single cells; c) sequencing the RNA expressed in each single cell and evaluating expression patterns that result in each single cell to generate single cell RNA sequencing data; d) obtaining a quantitative value of enzymatic function of the gene for each variant to yield an enzyme function score; and e) training a machine learning model with the enzyme function scores, thereby refining the predictive power of a pathogenicity prediction model.
Citation Information
Patent Citations
Methods and compositions for massively parallel variant and small molecule phenotyping
US20200283843A1
High-throughput methods for identifying gene interactions and networks
WO2019113499A1
Use of gene expression data and gene signaling networks along with gene editing to determine which variants harm gene function
WO2020198732A1
Clinical variant classifier models, machine learning systems and methods of use
WO2021076790A1
Analyzing expression of protein-coding variants in cells
WO2022192191A1
Cited By
Method and system for evaluating interference degree of splicing factors on single cell level function
CN121545586A
