Computer-implemented method and apparatus for analyzing genetic data

Abstract

The present disclosure relates to analyzing genetic data. In one arrangement, a method operates on input data comprising the strength of association between one or more phenotypes, including a target phenotype, and a plurality of genetic variants. A fine-mapping algorithm is applied to all or a subset of the input data to identify one or more independent phenotype-variant associations. For each association, a set of one or more fine-mapped variants is identified. A fine-mapping prediction model is calculated based on the input data and the set of fine-mapped variants. The effect of the set of fine-mapped variants on the target phenotype is subtracted from the input data to obtain residual association data. A machine learning algorithm is applied to the residual association data to identify further predictive correlations between the target phenotype and the plurality of genetic variants.

Description

Technical Field

[0001] The present invention relates to analyzing genetic and phenotypic data about an organism to obtain information about the organism, particularly in the context of being able to obtain an improved polygenic risk score (PRS) for a phenotype of interest. Background Art

[0002] A PRS is a quantitative summary of the contribution of an organism's inherited DNA to the phenotypes it may exhibit. A PRS may include all DNA variants associated (directly or indirectly) with the phenotype of interest, or components thereof may be used if they are more relevant to a specific aspect of the organism's biology (including cells, tissues, or other biological units, mechanisms, or processes). A PRS can be used directly or as part of multiple measurements or records about the organism to infer aspects of its past, current, and future biology. In the context of improving human health and healthcare, PRSs have a range of practical uses, including but not limited to: predicting the risk of developing a disease or phenotype, predicting the age of phenotype onset, predicting disease severity, predicting disease subtypes, predicting response to treatment, selecting appropriate screening strategies for individuals, selecting appropriate drug interventions, and setting prior probabilities for other predictive algorithms. PRSs can be used directly as input in the application of artificial intelligence and machine learning methods to make predictions or classifications based on other high-dimensional input data (e.g., imaging). They can be used to help train these algorithms, for example, to identify predictive measurements based on non-genetic data. In addition to their utility in making predictive statements about individuals, they can also be used to identify groups of individuals (including but not limited to the applications described above) by calculating PRSs for a large number of individuals and then grouping the individuals based on the PRSs. PRSs can also help select individuals for clinical trials, for example by optimizing trial designs by recruiting individuals who are more likely to develop a related disease or phenotype, thereby enhancing the evaluation of the efficacy of new treatments. PRSs carry information about the individuals for whom they are calculated, as well as information about their relatives (who share a portion of the DNA inherited by these individuals). Information about the effect of an individual's DNA on their phenotype can be derived from any relevant assessment of the potential effects of carrying any particular combination of DNA variants. In the following, we focus on the analysis of a large amount of recent information derived from genetic association studies (GAS). These studies systematically assess the potential contribution of DNA variants to the genetic basis of a phenotype.

[0003] Since the mid-2000s, GAS (typically genome-wide association studies: GWAS, or association studies targeting single variants, or variants in genomic regions, or GWAS limited to specific regions of the genome) have been performed on tens of thousands (mainly human) phenotypes in millions of individuals, resulting in billions of potential connections between genotypes and phenotypes. The resulting raw data are then typically simplified to produce summary statistics. For each genetic variant (whether imputed or observed), the GAS summary statistics consist of the inferred effect value of the genetic variant on the GAS phenotype and the standard error of the inferred effect value. In other cases, individual-level data consisting of the complete genetic profile of the individual under study and information about their phenotype are directly available. However, due to requirements for the privacy of individual data, individual-level data are generally less widely used.

[0004] In the following, we refer to a phenotype as synonymous with a single study. However, it is very common to have data available from multiple different studies based on the same or similar phenotypes, or from a single cohort where multiple different phenotypes are measured.

[0005] A PRS consists of the aggregation of the effects of a large number of genetic variants, typically each with a small individual effect, to construct a comprehensive predictor of the trait of interest. Variants included in such a score can be "causal variants," meaning that the variant directly affects the trait (weakly, but directly), or "marker variants," meaning that they are strongly associated with other, unknown causal variants, but the marker variant itself has no direct effect on the phenotype.

[0006] PRS can be calculated using individual-level data or summary statistics. The strategies for constructing PRS are expanding, but the generally accepted approach to constructing an accurate PRS involves deconvolving the signal in all associated regions by studying the combination of variants that best captures the underlying biological association. This process assigns a probability weight to each variant for each association, thereby describing which variant or variants are likely to be directly causal. This process is called "fine-mapping", and several strategies have been proposed previously to achieve this task (see, for example, Benner et al., Bioinformatics 2016, 15; 32(10): 1493-1501).

[0007] The number of associations will vary, with many genomic regions containing a single potential association and some genomic regions containing multiple independent associations (up to 10 have been reported, but this is rare). The technical challenge of identifying the correct combination of variants responsible for all associations in a region is that these variants can be correlated with each other. The greater the correlation, the greater the number of samples required to resolve these correlations.

[0008] Some tools for constructing PRSs are designed to exploit summary statistics. One such approach is pruning and thresholding: the most relevant variants are selected to contribute to the PRS, and their highly correlated variants are removed. The most relevant of the remaining variants is then selected, and the process is repeated until the importance of the remaining variants falls below a predefined threshold. Another approach, popularized by the LDpred software (https: / / github.com / bvilhjal / ldpred), iterates multiple random selections of plausible variants across the genome and estimates the residual signal when variants are selected or removed.

[0009] The advantage of strategies based on summary statistics is that the lack of restrictions on sharing individual-level data means that larger sample sizes can be made available to the scientific community. This is why most current PRS designs are based on these large sets of summary statistics.

[0010] However, for all methods based on summary statistics, correlated variants are processed by referring to external data sources that describe the expected correlations between variants. The correlation pattern between genetic variants is called linkage disequilibrium (LD). The correlations in these external data sources will not completely match the correlations obtained from the individual-level data used to generate summary statistics. Due to the uncertainty about what the correct correlation should be, this will introduce additional uncertainty in the fine mapping process. Therefore, fine mapping based on summary statistics is fundamentally limited by the uncertainty of the underlying LD pattern.

[0011] Another limitation of relying on external datasets to describe LD patterns is that different populations have different LD patterns. Therefore, inferences made for one population are unlikely to be equally accurate for different populations. In other words, PRSs derived from a reference LD dataset have limited robustness to population variability. Summary of the Invention

[0012] It is an object of the present invention to improve the analysis of genetic data about an organism and / or to allow obtaining a more robust and / or accurate PRS for an individual.

[0013] Although the pattern of LD varies across populations, variants that affect a trait or disease in one population will often also affect the same trait / disease in a different population. Therefore, using fine-mapping techniques to identify one or more causal variants, or sets of variants that may include or label the one or more causal variants, will make PRS more accurate, particularly by improving its robustness to population variability.

[0014] However, not all variants can be finely mapped, especially the large number of variants with small effects on the target phenotype. Therefore, alternative techniques that do not require precise statement of which variants are causal and focus only on prediction problems are also useful for PRS construction.

[0015] Accurate derivation of a PRS is likely to have high clinical utility in predicting disease or predicting an individual's response to a specific drug or treatment and would therefore benefit from statistical techniques that exploit the advantages of fine mapping while also allowing the use of alternative machine learning techniques where appropriate.

[0016] According to aspects of the present invention, a computer-implemented method for analyzing genetic data about an organism to obtain information about the organism is provided, the method comprising: receiving input data, the input data comprising the strength of association between one or more phenotypes including a target phenotype and multiple genetic variants in a region of interest of the genome of the organism; applying a fine mapping algorithm to all or a subset of the input data to identify one or more independent phenotype-variant associations within the region of interest, comprising: identifying a set of one or more fine mapping variants from the multiple genetic variants for each association, and determining an estimated probability of having a causal relationship with the phenotype-variant association for each fine mapping variant, the sum of the probabilities of the fine mapping variants in the set being 1; calculating a fine mapping prediction model based on the input data and the set of fine mapping variants, the fine mapping prediction model quantifying the effect of the set of fine mapping variants on the target phenotype; using the fine mapping prediction model to subtract the effect of the set of fine mapping variants on the target phenotype from the input data to obtain residual association data; and applying a machine learning algorithm to the residual association data to identify further predictive correlations between the target phenotype and the multiple genetic variants.

[0017] By using fine-mapping techniques to identify fine-mapped variants that are potentially causal for the target phenotype, and additionally analyzing the residual signal remaining after the effects of the fine-mapped variants are accounted for (via residual association data), this approach can account for further weak correlations that may be present in the data. Including these additional correlations improves the predictive accuracy of the model.

[0018] In one embodiment, the strength of association includes an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error of each estimated effect size. The estimated effect sizes and their errors are widely available as summary statistics from a large number of studies, thereby allowing access to a large amount of data.

[0019] In one embodiment, the step of receiving input data includes: receiving individual-level data including a genotype and a corresponding phenotype for each of a plurality of individuals; and using the individual-level data, determining an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error for each of the estimated effect sizes. In some embodiments, individual-level data can be used because it is not subject to underlying assumptions about correlations between variants within a region that may be present in summary statistics, thereby reducing the chance of introducing unintentional bias or error.

[0020] In one embodiment, the identification of a set of fine-mapped variants is performed using an iterative approach, wherein each iteration comprises: identifying, based on input data, a fine-mapped variant that is different from any previously identified fine-mapped variant in a genomic region; updating the input data to account for the effect of the fine-mapped variant already identified on the target phenotype using a correlation matrix between the genetic variants in the genomic region; and determining whether to perform further iterations based on the updated input data. By using an iterative approach, multiple fine-mapped variants can be identified from residual signal not accounted for by a single fine-mapped variant, thereby maximizing the use of signal present in the aggregated data.

[0021] In one embodiment, identifying a set of fine-mapping variants comprises using a plurality of instrument traits known to affect a target phenotype, wherein the use of the instrument traits comprises: determining a set of fine-mapping variants for the instrument traits; and determining whether to include each of one or more of the fine-mapping variants for the instrument traits in the set of fine-mapping variants for the target phenotype based on the relationship between the plurality of instrument traits and the target phenotype. The relationship between the plurality of instrument traits and the target phenotype can take into account potentially complex association patterns between the instrument traits and the target phenotype. Alternatively or additionally, in other embodiments, identifying the set of fine-mapping variants comprises identifying a set of fine-mapping variants for one or more direct causal instrument traits known to affect the target phenotype. In this case, it may not be necessary to consider complex association patterns between the plurality of instrument traits and the target phenotype.

[0022] The use of instrument features can improve the accuracy of determining fine-mapped variants for a phenotype, where the genetic variant has only a small effect on the target phenotype but a larger effect on the instrument feature.

[0023] In one embodiment, calculation of the fine-mapped prediction model includes determining, for the one or more instrument features, an effect size of the set of fine-mapped variants on the one or more instrument features, and determining an effect size on the target phenotype for each fine-mapped variant for the instrument feature included in the set of fine-mapped variants for the target phenotype based on a predetermined relationship between the effect size on the instrument feature and the effect size on the target phenotype. The instrument features can also be used to improve estimates of effect sizes where the effect of a genetic variant on the instrument feature is greater than the effect on the target phenotype. This is particularly useful when the relationship between the instrument feature and the target phenotype itself has been well characterized.

[0024] In one embodiment, the effect of the set of finely mapped variants on the target phenotype is inferred using a machine learning algorithm. In this embodiment, the set of finely mapped variants and their corresponding marginal effect values ​​are input into the machine learning algorithm to generate effect values, such that the residual association data are marginal effect values ​​corresponding only to the set of finely mapped variants. The set of finely mapped variants may further include one or more variants known to have a high causal likelihood for the target phenotype.

[0025] This reduces the number of genetic variants to which the machine learning algorithm is applied by focusing it on the variants most likely to be causal. This reduces the computational load and increases the efficiency of the method.

[0026] In one embodiment, the association strength includes an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error of each estimated effect size; and the fine-mapped prediction model includes a fine-mapped effect size of each of the fine-mapped variants on the target phenotype, the fine-mapped effect size being calculated from the estimated effect size of the fine-mapped variant taking into account the estimated probability of the fine-mapped variant being causal with the phenotype-variant association. Adjusting the effect size of the fine-mapped variant based on its probability of being causal ensures that its importance is not overestimated if the certainty that the fine-mapped variant is causal is low.

[0027] In one embodiment, the association strength comprises an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error of each estimated effect size; and the step of subtracting the effect of the set of fine-mapped variants on the target phenotype from the input data comprises obtaining a residual effect size for each of the plurality of genetic variants in the input data, the residual association data comprising the residual effect sizes, wherein the residual effect size for genetic variant i is ∑ i = ... It is given by:

[0028]

[0029] where β i is the estimated marginal effect value of the gene variant i, N is the number of the finely mapped variants, p j is the probability that variant j is causal, is the fine-mapped effect size of the jth fine-mapped variant on the target phenotype, r ij is the correlation between the jth fine-mapped variant and gene variant i.

[0030] The above method enables the clear identification of the residual effects of variants in the regions of interest of the genome for further analysis by machine learning algorithms.

[0031] In one embodiment, the input data is derived from a plurality of different genetic studies, and applying a machine learning algorithm to the residual association data comprises using a prior probability for each of the plurality of genetic variants that is causally associated with the target phenotype, the prior probability being dependent on the consistency of the strength of association between each genetic variant and the target phenotype across the different genetic studies. Using a non-flat prior for the machine learning algorithm allows the method to improve its accuracy by taking into account further information about the certainty that specific data are reliable.

[0032] In one embodiment, the step of applying a machine learning algorithm to the residual association data includes using a prior probability of each of the plurality of genetic variants being causally related to the target phenotype, the prior probability being dependent on genomic annotations for the plurality of genetic variants in the region of interest. Including the genomic annotations provides further data regarding the likelihood that a particular variant is causally related to the target phenotype, thereby improving the determination of effect size.

[0033] In one embodiment, the method further comprises the step of calculating a polygenic risk score for the individual for the target phenotype using the finely mapped prediction model and further predictive correlations identified by the machine learning algorithm.

[0034] Taking into account further correlations identified by the machine learning algorithm improves the accuracy of the PRS by allowing the method to account for residual signal not explained by the set of finely mapped variants.

[0035] In one embodiment, the input data originates from multiple different groups of the organism and satisfies either or both of the following: the parts of the input data corresponding to different groups are respectively calculated to obtain multiple fine mapping prediction models matching each group; and the parts of the input data corresponding to different groups are respectively applied to the residual association data to obtain a set of further predicted correlations matching multiple groups.

[0036] Providing a fine-mapped prediction model matched to a specific population and a further set of residual predicted correlations allows the method to take into account possible changes in linkage disequilibrium (correlations between variants) within the region of interest of the genome.

[0037] In one embodiment, the method further comprises receiving input data from an individual having genes from a mixture of the different populations; calculating a polygenic risk score for the individual by performing one or both of the following operations: matching each of a plurality of population-matched fine-mapping prediction models to a corresponding portion of the input data that is population-matched to the population-matched fine-mapping prediction model, and applying each matched fine-mapping prediction model to a corresponding portion of the input data; and matching each of a plurality of sets of population-matched further predictive correlations to a corresponding portion of the input data that is population-matched to the population-matched set of further predictive correlations, and applying each matched set of further predictive correlations to a corresponding portion of the input data.

[0038] Calculating the polygenic risk score for an individual using multiple fine-mapped prediction models and / or further residual prediction correlation sets matched to different multiple corresponding portions of the input data from the individual allows the method to provide a more accurate predictive risk score that takes into account systematic differences in correlations between variants associated with different populations.

[0039] In one embodiment, the method further comprises: receiving input data from an individual having genes primarily from one of the different populations; and calculating a polygenic risk score for the individual by performing one or both of the following operations: applying a population-matched fine-mapping prediction model to all input data from the individual, wherein the population-matched fine-mapping prediction model is matched to the individual's population; and applying a set of population-matched further predictive correlations to all input data from the individual, wherein the set of population-matched further predictive correlations is matched to the individual's population.

[0040] Calculating the polygenic risk score using a fine-mapped prediction model matched to the individual's population and a further set of residual predictive associations allows the method to provide a more accurate predictive risk score that accounts for systematic differences in associations between variants associated with different populations.

[0041] In one embodiment, identifying one or more fine-mapped variants by a fine-mapping algorithm takes into account associations between multiple genetic variants and phenotypes other than the target phenotype.

[0042] Using information about other phenotypic associations can maximize the amount of information available to identify the fine-mapped variants and their effect sizes. This further improves the accuracy of the results of the method.

[0043] According to an alternative aspect, a device for analyzing genetic data about an organism to obtain information about the organism is provided, the device comprising: a receiving unit configured to receive input data, the input data comprising strengths of association between one or more phenotypes including a target phenotype and a plurality of genetic variants in a region of interest of the genome of the organism; and a data processing unit configured to: apply a fine mapping algorithm to all or a subset of the input data to identify one or more independent phenotype-variant associations in the region of interest by identifying, for each association, a set of one or more fine mapping variants from the plurality of genetic variants, and determining, for each fine mapping variant, an estimated probability of having a causal relationship with the phenotype-variant association, the sum of the probabilities of the fine mapping variants in the set being 1; calculate a fine mapping prediction model based on the input data and the set of fine mapping variants, the fine mapping prediction model quantifying the effect of the set of fine mapping variants on the target phenotype; subtract the effect of the set of fine mapping variants on the target phenotype from the input data using the fine mapping prediction model to obtain residual association data; and apply a machine learning algorithm to the residual association data to identify further predictive correlations between the target phenotype and the plurality of genetic variants. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Embodiments of the present invention will be further described, by way of example only, with reference to the accompanying drawings, in which:

[0045] Figure 1 is a flow chart describing a method for analyzing genetic data to obtain information about an organism;

[0046] Figure 2 An apparatus for analyzing genetic data to obtain information about an organism is described;

[0047] Figure 3is a graph showing a comparison of effect sizes between coronary artery disease (CAD) and low-density lipoprotein (LDL);

[0048] Figure 4 Four graphs representing steps in a stepwise forward regression analysis used to identify four respective independent association signals for identifying finely mapped variants associated with LDL in the LPA region of chromosome 6 are shown;

[0049] Figure 5 is a description Figure 4 Plot of the combined LDL effect size versus the marginal LDL effect size estimates for the four association signals identified in ;

[0050] Figure 6 This describes the LDpred machine learning algorithm applied to Figure 4 and Figure 5 a graph of CAD PRS weights for the LPA region of chromosome 6 obtained from residual association data obtained by analyzing ; and

[0051] Figure 7 is a graph depicting the CAD PRS weights for the LPA region of chromosome 6 obtained by applying the LDpred machine learning algorithm directly to the CAD variant data without any prior fine mapping step. DETAILED DESCRIPTION

[0052]

[0014] Embodiments of the present disclosure relate to computer-implemented methods of analyzing genetic data about an organism to obtain information about the organism. Figure 1 The framework of these methods is described. Figure 2 An apparatus 6 for carrying out these methods is described.

[0053] In step S1, input data 2 is received (for example, by the receiving unit 8 of the device 6). The receiving unit 8 may include a data communication interface. The data communication interface allows the input data 2 to be provided to the data processing unit 10 of the device 6. The data processing unit 10 may include any suitable combination of computer hardware, firmware and / or software configured to perform the data processing functions described below. Optionally, a computer program provided on a computer-readable medium may be provided, which includes instructions for performing any of the methods described below. The device 6 is described as a stand-alone unit (for example, a single PC or workstation), but this is not required. In other embodiments, the device 6 includes a distributed computing system comprising multiple computers connected via a network.

[0054] In some embodiments, input data 2 comprises the strength of association between one or more phenotypes including a target phenotype and a plurality of gene variants in a region of interest of an organism's genome. In some embodiments, input data 2 comprises one or both of GWAS summary statistics and individual-level data. As will be described in more detail below, the method can use input data 2 to (i) identify high-confidence variants (referred to as fine-mapping variants) that have a direct causal effect on the target phenotype; (ii) for an individual, after adjusting for high-confidence variants and / or predicting characteristic risk (e.g., in the form of a PRS), obtain residual association data (which can be referred to as a residual signal and / or derived from a residual signal). This method is particularly advantageous when used in an embodiment where the organism is a human.

[0055] The target phenotype can be any phenotype of interest for which there are subjects of a GWAS or for which relevant individual-level genetic data are available. Examples of such phenotypes are numerous and include: the level of expression of a gene (and associated nucleotide sequence) and the regulation of expression; epigenetic features (e.g., nucleotide modifications, chromosome conformation); the abundance level of a protein or peptide; the function and / or molecular structure of a protein or peptide; the amount of a molecule (e.g., a drug, hormone, DNA molecule or RNA molecule, metabolite, vitamin) in an organism; characteristics of biochemical and metabolic processes (e.g., basal metabolic rate, prothrombin time, activated partial thromboplastin time); cell morphology and function (e.g., mean corpuscular volume of red blood cells, absolute neutrophil count); tissue morphology and function (e.g., bone mineral density, hair color); the morphology and function of organs and organ systems (e.g., left ventricular ejection fraction, any response to an external stimulus or stimuli (e.g., light, sound, touch, or any other sensory input); any response to exposure to substances or pathogens (e.g., dietary intake, medications, gases, viruses, bacteria); behavioral and lifestyle characteristics (e.g., smoking, alcohol consumption, occupation); reproductive and life course characteristics and function (e.g., age at menarche, placental weight, years of education); the onset, trajectory, and prognosis of diseases or conditions (e.g., diabetes, cardiovascular disease, obesity); measurable anatomical characteristics (e.g., body mass index, lean muscle mass, body fat percentage); measurable physiological or functional characteristics (e.g., heart rate, blood pressure, intelligence); and measurable psychological or cognitive characteristics (e.g., fluid intelligence indicators, psychiatric symptoms). Any of these measurements may be absolute or relative. Phenotypes are also often referred to as traits.

[0056] In step S2, a fine mapping algorithm is applied to all or a subset of the input data 2. In one embodiment, the fine mapping step identifies variants with high causal confidence, thereby obtaining a set of finely mapped variants. Further details about step S2 are given below.

[0057] In step S3, a fine mapping prediction model is calculated based on input data 2 and fine mapping variants. The fine mapping prediction model quantifies the effect value of the fine mapping variant on the target phenotype. The effect value refers to the degree of influence of a given variant on the risk of disease (or more generally, the "risk" of having or developing any given phenotype). For example, an effect value of 1.2 means that for a given variant, the risk of each risk allele increases by 20% (for each individual, which can be encoded as 0, 1 or 2). Therefore, the quantification of the effect value allows the fine mapping prediction model to make predictions for individuals based on genetic data from individuals. More details about S3 are given below.

[0058] In step S4, the effects of the set of finely mapped variants on the target phenotype are subtracted from the input data 2 using the finely mapped prediction model to obtain residual association data. Further details on step S4 are given below.

[0059] In step S5, a machine learning algorithm is applied to the residual association data to identify further predicted correlations between the target phenotype and multiple genetic variants of the input data 2. In the specific example below, a machine learning algorithm called LDpred is used. LDpred is well known in the field of fine mapping and PRS generation. Software for implementation is available at https: / / github.com / bvilhjal / ldpred. Further predicted correlations can quantify the effect size associated with variants other than the fine mapping variants (after taking into account the effect of the fine mapping variants), thereby allowing for refinement of predictions about individuals relative to if only the fine mapping prediction model is applied to the genetic data from the individual.

[0060] In step S6, the PRS model is evaluated. The PRS model may be derived in part from the fine mapping prediction model of step S3 and in part from the further prediction relevance of the machine learning performed in step S5. As will be described below, the combination of the fine mapping prediction model and the further prediction relevance from the machine learning may define a recipe for calculating the PRS, which takes the form of a weighted sum of variants, where the weights for the fine mapping variants are provided by the fine mapping prediction model and the weights for the other variants are provided by the further prediction relevance from the machine learning. This is possible if the trained machine learning algorithm can be interpreted in terms of such a weighted sum of variants. In other embodiments, the trained machine learning algorithm may be more complex and therefore represented as part of the PRS model in a different way.

[0061] The PRS model calculated in step S6 can be used to calculate a PRS score based on genetic data from an individual. The PRS model can be output as data representing the PRS model (e.g., via Figure 2The steps leading up to and including step S6 (including the training of the machine learning algorithm) may thus be performed on one device 6, and subsequent steps involving the use of the PRS model (e.g., for calculating an individual's PRS score) may be performed on another device (not shown) comprising any suitable combination of computer hardware, firmware, and / or software capable of performing the necessary data processing tasks. Alternatively, the calculation of the PRS score may be performed on the same device 6 that calculates the PRS model.

[0062] In step S7, the PRS score of the individual is calculated using the PRS model calculated in step S6. The PRS score may be output as data 4 representing the PRS score.

[0063] The calculated PRS model constitutes information about organisms at a general level (e.g., about humans in general), in the sense that it enables the calculation of a PRS score based on genetic information obtained from any individual. The PRS score constitutes information about a specific individual organism (e.g., a single human subject).

[0064] Example application scenarios

[0065] Figures 3 to 7 Depicts the use of Figure 1 and will be mentioned in a more detailed discussion of the method steps given below.

[0066] Figure 3 We show how the effect size of LDL correlates with the effect size of CAD at 95 different loci associated with LDL. This correlation is observed in part because LDL is thought to have an almost direct causal effect on CAD. Therefore, we refer to LDL as a suitable instrument for CAD, meaning that information about LDL can be used to improve the accuracy of PRS for CAD.

[0067] Figure 4 Instructions for use Figure 1 Results of LDL fine-mapping using an established method (stepwise forward regression) in the LPA region of chromosome 6. Each regression step identified an additional independent phenotype-variant association, for a total of four independent phenotype-variant associations. In each tile, black triangles represent newly identified fine-mapped variants or a credible set (CS) of fine-mapped variants. At each step, genetic variants with low fine-mapped probability (<1%) are colored gray. The first LDL association signal had four fine-mapped variants with a posterior probability greater than 1%, while the remaining three LDL association signals identified one fine-mapped variant with a fine-mapped probability >1%.

[0068] Figure 5 As shown, for the LDL instrument characteristics of this example, according to Figure 4 The four joint estimated effect sizes for the four independent phenotype-variant association estimates shown in differ slightly from the four marginal estimated effect sizes.

[0069] Figure 6 and Figure 7 Derived PRS weights for CAD of the same LPA region in chromosome 6 are depicted. Figure 6 From LDL fine mapping and effect size ( Figure 4 and Figure 5 ) extrapolated finely mapped CAD variants and subtracted them from the CAD data before LDpred analysis to capture the residual signal (indicating further predictive relevance). Figure 6 The PRS weights derived from the LDL fine map (black) are combined with the PRS weights derived from the LDpred residual signal (grey). Figure 7 In contrast, for Figure 7 , applying the standard LDpred analysis strategy based only on CAD without an initial fine mapping step. One of the consequences of the limited accuracy of the procedure without an initial LDL fine mapping is that Figure 7 No fourth signal was detected in the CAD data alone because the CAD data alone were insufficient to characterize this association.

[0070] Given below Figure 1 For further exemplary implementation details of steps S2 to S7, reference is made to the above exemplary application scenarios where appropriate for illustrative purposes.

[0071] Step S2: Fine mapping

[0072] As described above, in step S2, the method applies a fine-mapping algorithm to all or a subset of the input data 2 to identify one or more independent phenotype-variant associations within a region of interest. Identifying one or more independent phenotype-variant associations within a region of interest can include identifying high-confidence fine-mapping variants for the phenotype of interest, where these variants are high-confidence fine-mapping variants that are causal variants or tag variants of causal variants. For each association, a set of one or more fine-mapping variants is identified from a plurality of genetic variants.

[0073] Figure 4 The application of the fine mapping algorithm under the conditions of the exemplary application scenario described above is shown. In this case, the fine mapping algorithm identifies four independent phenotype-variant associations for LDL in a region on chromosome 6 (one association for each of the four tiles shown).

[0074] Fine-mapping algorithms are typically designed to capture the underlying causal biology of a target phenotype by locating one or more causal variants, or alternatively, locating one or more plausible sets of variants that contain or closely label one or more causal variants. Fine-mapping algorithms contrast with alternative purely predictive approaches, which are typically based on machine learning techniques such as LASSO, random forests, or neural networks, which capture predictive signals without providing discrete summaries of data that map to the underlying biology.

[0075] Phenotype-variant associations are independent in the following sense: even though there may be some degree of correlation between the two identified variants, the association of the second fine-mapped variant with the phenotype is not solely due to its correlation with the first fine-mapped variant, which is also associated with the phenotype. In other words, the second fine-mapped variant is associated with the phenotype even after accounting for or adjusting for the first fine-mapped variant's association with the phenotype. In contrast, multiple variants within a CS are not independent of each other because if we select one of the variants within the CS and adjust for that variant, the associations at all other variants within the CS disappear—that is, multiple associations exist only due to the high correlation between the variants.

[0076] Each independent phenotype-variant association can be associated with a single fine-mapped variant or a credible set (CS) of (multiple) fine-mapped variants. For each association, a set of one or more fine-mapped variants is identified from a plurality of genetic variants. A CS of fine-mapped variants is a set of two or more fine-mapped variants that are considered to have a high likelihood of being causal with the target phenotype. The method determines for each fine-mapped variant an estimated probability that the phenotype-variant association is causal, with the probabilities of the fine-mapped variants in the set summing to 1. In the case where only one fine-mapped variant is identified, the estimated probability will simply be 1 for that fine-mapped variant. In Figure 4 In

[15] , forward regression steps 2 to 4 show examples of identifying a single fine-mapped variant, while forward regression step 1 identifies CSs to fine-mapped variants.

[0077] In some embodiments, identifying one or more fine-mapped variants by a fine-mapping algorithm takes into account associations between multiple genetic variants and phenotypes other than the target phenotype. Input data 2 for such embodiments will therefore include the strength of associations between multiple phenotypes and multiple genetic variants in a region of interest in the genome of an organism. Using associations with multiple phenotypes helps to fully utilize data from a large number of studies, which may contain a wide range of different phenotypes, and takes advantage of the fact that many traits can share the same causal variant.

[0078] In one embodiment, input data 2 includes data describing the association between individual variants and target phenotypes in the form of marginal variant effect values ​​and standard errors. In such an embodiment, the strength of association may include an estimated effect value of each of the multiple genetic variants on the target phenotype, and the standard error of each estimated effect value. The estimated effect value is the marginal variant effect value. The marginal variant effect value refers to the effect of the variant when considered in isolation, i.e., ignoring the effects of nearby related variants. For example, a label variant may have a strong marginal effect value, but its "true" effect value is zero. Input data in this format are typically referred to as summary statistics.

[0079] In one embodiment, the fine mapping algorithm is applied to all or a subset of input data 2 to identify one or more independent phenotype-variant associations in the region of interest, including the following. By using a probability model (such as a Bayesian statistical model) in a given DNA region (i.e., the region of the genome of an organism), multiple studies (each of which contains data on the strength of association between the target phenotype and one or more gene variants) are assigned to a cluster, assuming that each cluster has a similar causal variant pattern. The space of possible cluster assignments is then explored using a Markov chain Monte Carlo algorithm or similar algorithms. Once a certain number of iterations have been performed assigning the study to a cluster, the feature set of the cluster can be used to identify the CS (i.e., the set of one or more finely mapped variants) of a single variant or gene variant, which may be a causal relationship with the phenotype assigned to the corresponding cluster. This method based on a large number of phenotypes is used to improve the ability and accuracy of identifying variants that affect phenotypes. More details of this type of method can be found in PCT application number PCT / GB2019 / 050525.

[0080] In some cases, this method can identify at most a single fine-mapped variant or a single CS for a fine-mapped variant in a given DNA region. However, there may be more than one independent fine-mapped variant (or correspondingly more than one CS) that may be causal within a region. Identifying these additional independent fine-mapped variants will provide additional predictors of the disease or trait of interest, thereby improving the ability to predict an individual's risk of developing the disease or trait.

[0081] Alternative implementation of step S2 when only summary statistics are available

[0082] When only summary statistics are available, additional independent fine-mapped variants can be identified. In one embodiment, this is achieved by considering the correlation between genetic variants within a genomic region, typically summarized by an "LD matrix," which is a matrix of genotypes g.i and g j Correlation r at position i, j ij A matrix of phenotypes is typically obtained from a subset of a reference panel, such as the 1000 Genomes consortium or the Haplotype Reference Consortium. Methods such as FINEMAP (Benner et al., Bioinformatics 2016, 15; 32(10): 1493-501) can be appropriately adapted to situations where we consider a large number of studies and phenotypes.

[0083] Another such embodiment will identify additional causal variants (referred to herein as fine mapping variants) by updating summary statistics to illustrate the effects of the fine mapping variants identified within the DNA region and then evaluating the residual evidence for the variants of the additional fine mapping. In this case, an iterative approach is used to identify sets of fine mapping variants. Each iteration is included in identifying fine mapping variants that are different from any previously identified fine mapping variants within the genomic region based on the input data, updating the input data to explain the effects of the target phenotype of the identified fine mapping variants, using the correlation matrix between the gene variants within the genomic region, and determining whether to perform further iterations based on the updated input data (e.g., stopping when determining that the updated input data no longer contains any information of interest, e.g., stopping when no longer exceeding a predetermined significance threshold and / or P value that is relatively flat).

[0084] This method can be applied iteratively to explore the space of fine-mapped variants within a DNA region by proposing the addition or removal of at most one fine-mapped variant (https: / / projecteuclid.org / euclid.aoas / 1507168840). Thus, in some embodiments, the step of identifying a fine-mapped variant that is different from any previously identified fine-mapped variant includes removing the previously identified fine-mapped variant from the set of fine-mapped variants. Further details of these methods can be found in PCT Application No. PCT / GB2019 / 050525.

[0085] Alternative implementation of step S2 using individual-level data

[0086] An alternative fine mapping strategy is to perform fine mapping using individual-level data. In such an embodiment, the step of receiving input data includes: receiving individual-level data containing the genotype and corresponding phenotype for each of a plurality of individuals, and using the individual-level data to determine an estimated effect size of each of a plurality of genetic variants on the target phenotype and the standard error of each estimated effect size. This can be implemented using a stepwise regression method to explore the space of fine-mapped variants using forward selection, backward elimination, or a combination of the two.

[0087] Alternatively, individual-level data can be used in conjunction with summary statistics to exploit information obtained from fine-mapping methods based on summary statistics (e.g., the method described in PCT Application No. PCT / GB2019 / 050525). One way this can be achieved is to use individual fine-mapped variants / CSs obtained from methods such as those described in PCT Application No. PCT / GB2019 / 050525 and adjust these in a subsequent stepwise regression step (as before, a combination of forward selection and backward elimination can be used).

[0088] Alternatively, residual summary statistics derived from individual-level data adjusted for identified fine-mapping variants can be obtained. In this way, the effects of variants already identified as high-confidence fine-mapping variants are removed, making it possible to use residual correlations to identify more fine-mapping variants.

[0089] This is similar to the way summary statistics are adjusted when individual-level data are unavailable, with the main advantage that no LD information is required. These derived residual summary statistics can be used as input to methods such as those described in PCT Application No. PCT / GB2019 / 050525. This process can be repeated iteratively. The method can be based on using only summary statistics derived from individual-level data or combined with residual summary statistics derived from studies where individual-level data are not available using LD panels.

[0090] Alternative implementation of step S2 using one or more instrument features

[0091] An alternative implementation of step S2 utilizes instrumentation studies so that the identification of a set of fine-mapping variants includes the use of one or more instrumental features known to affect the target phenotype. We define a feature as an instrument for the target phenotype when that feature is strongly associated with the trait of interest. A special case is an instrument that is a direct causal / modifying causal ...

[0092] In many cases, the effect of a variant on the target phenotype is too small to identify a credible set (CS) of the target phenotype. However, using an appropriately powered instrumentation study, this effect may be large enough to enable fine-mapping. In other words, because the effect of the variant on the instrumented trait is larger than the effect of the variant on the target phenotype, it is easier to accurately determine whether the variant is causal for the instrumented trait. In this case, fine-mapping and causal signal identification will be based entirely on the instrumentation study, providing information about the target phenotype that would otherwise remain uncharacterized.

[0093] Based on the above insights, in one embodiment, identifying a set of fine-mapped variants includes identifying a set of fine-mapped variants of one or more directly causal instrumental features that are known to affect the target phenotype. This is a relatively simple method of using instrumental features to supplement the fine mapping of step S2, but requires that known directly causal instrumental features are available. In other embodiments, identifying a set of fine-mapped variants includes using multiple instrumental features that are known to affect the target phenotype. Then, based on the relationship between the instrumental features and the target phenotype, the method includes determining a set of fine-mapped variants of the instrumental features, and determining whether to include each of the one or more fine-mapped variants of the instrumental features in the set of fine-mapped variants of the target phenotype. In this case, the relationship between multiple instrumental features and the target phenotype can take into account potentially complex association patterns between the instrumental features and the target phenotype, thereby allowing the use of instrumental features that are not necessarily directly causal instrumental features.

[0094] Figures 3 to 5 An example is provided in which fine mapping is performed on LDL, an instrumented feature of CAD, and fine-mapped variants identified for LDL ( Figure 4 ) was used in subsequent steps where CAD was used as the phenotype of interest ( Figure 6 ).

[0095] Step S3: Compute the fine-mapped prediction model (e.g., estimate the effect size of the fine-mapped variants)

[0096] As described above, in step S3, the method calculates a fine-map prediction model based on the input data 2 and the set of fine-map variants (identified in step S2). The fine-map prediction model quantifies the effect of the set of fine-map variants on the target phenotype. The effect on the target phenotype can be quantified using a fine-map effect value for the target phenotype, in which case the fine-map prediction model includes or consists of a fine-map effect value for the target phenotype for each fine-map variant, which effect value explains the correlation between the variants.

[0097] In an embodiment where the strength of association includes summary statistics (e.g., the estimated effect value of each of the multiple genetic variants on the target phenotype, and the standard error of each estimated effect value), the fine mapping effect value can be directly obtained from the marginal summary statistics of a single GWAS from the target feature (i.e., the target phenotype). When a single fine mapping variant is identified in a region, the effect value reported in the GWAS summary statistics can be used. When the CS of the variant is identified, the GWAS summary statistics can be weighted according to the probability that the variant is causal (relative to the remaining variants in the CS). In some embodiments, taking into account the estimated probability of the fine mapping variant being causal (e.g., derived from input data 2, e.g., as weighted as described above) associated with the phenotype-variant, the effect value of each fine mapping can be calculated from the estimated effect value of the fine mapping variant (e.g., derived from input data 2). For example, the fine mapping effect value can be derived based on multiplying the estimated effect value by the probability that the fine mapping variant is causal.

[0098] Alternative implementation of step S3 when relevant associations exist

[0099] When multiple credible sets are identified in the same DNA region, capturing several independent biological associations, the effect size needs to be corrected to control the correlation between the associations. The corrected effect size is usually called the joint effect size. Figure 4 The example of fine mapping of the LPA region of chromosome 6 is illustrated in the figure. Figure 5 Figure 2 shows that for our LDL instrument characteristics, the four joint estimated effect sizes differ slightly from the four marginal estimated effect sizes. If the associations were strong, the differences could be large.

[0100] When multiple fine-mapped variants are associated with a trait independently of one another, there may still be some correlation between them. The marginal effect sizes of these independent fine-mapped variants need to be adjusted to account for the correlation between the variants. So, in other words, the joint effect size is the effect size of multiple variants for a trait that takes into account the correlation between the variants, such as the four fine-mapped variants in the LDL example, which account for some correlation between the four variants.

[0101] The correction of this joint effect value estimate can be applied using summary statistics (such as Yang et al., Nature Genetics 2012, 44 (4): 369-75), provided that the variant correlation (or LD) pattern (population-specific) in the DNA region is well characterized. Alternatively, this correction can be applied using individual level data, thereby using a regression model to jointly fit all selected fine mapping variants identified in the DNA region. If several different associations are related to a credible set of associated variants, this correction is necessary.

[0102] Alternative implementation of step S3 using one or more instrument features

[0103] An alternative approach to effect size estimation is to utilize instrument studies. In this type of embodiment, the identification of a set of fine-mapped variants in step S2 includes determining a set of fine-mapped variants that are known to affect one or more instrumental features of the target phenotype. The calculation of the fine-mapped prediction model then includes determining the effect size of the set of fine-mapped variants for one or more instrumental features on the one or more instrumental features, and determining the effect size of each fine-mapped variant for the instrumental feature included in the set of fine-mapped variants for the target phenotype on the target phenotype based on a predetermined relationship between the effect size on the instrumental feature and the effect size on the target phenotype. Because the effect of a genetic variant on the instrumental feature is higher than the effect on the target phenotype, it is easier to estimate the effect size of the variant on the instrumental feature than on the target phenotype.

[0104] Therefore, if external or genome-wide data allow accurate characterization of the relationship between the instrument trait effect size and the target phenotype effect size, then the better estimated effect size for the instrument trait can be exploited to more accurately estimate the effect size for the target phenotype. One way to characterize the relationship between the instrument trait effect size and the target phenotype effect size is to perform a linear regression on the effect sizes of variants defined as finely mapping to both the instrument trait and the target phenotype.

[0105] Figure 3 An example is shown in which a large set of LDL-associated variants is used to infer a relationship between the effect size of LDL and the effect size of CAD. In this example, LDL serves as an instrumental feature for CAD.

[0106] Alternative implementation of step S3 using all studies / phenotypes as potential instrument features

[0107] An alternative to step S3 is to use independent fine-mapped variants (or CSs) identified for all studies used to train the above probabilistic model (as described in detail in PCT application number PCT / GB2019 / 050525). This results in a set of variants that are likely to be causal to at least one disease / trait.

[0108] The machine learning algorithm of step S5 can then be applied to the set of variants (typically, the number of variants in the set is much smaller than the number used for step S5). Therefore, a machine learning algorithm is used to infer the effect of the set of finely mapped variants on the target phenotype, which algorithm is preferably the same algorithm as the algorithm used in step S5. The input to this embodiment of step S3 is the marginal effect value of each finely mapped variant, i.e., no signal subtraction is applied at this stage. The output of this embodiment of step S3 is the same as the output of step S5, i.e., a set of weights of residual effect values ​​based on the uncertainty of the effect value estimate and the probability of the variant being causal with the focal phenotype. These weights calculated for the subset of finely mapped variants are then subtracted from the effect values ​​of multiple genetic variants, thereby generating residual association data comparable to other embodiments of step S4.

[0109] In some embodiments, the set of fine-mapped variants can be combined with a set of variants reported in the literature that have a high likelihood of a causal relationship with the disease / trait. Thus, the set of fine-mapped variants further includes one or more variants known to have a high likelihood of a causal relationship with the target phenotype.

[0110] Alternative implementation of step S3 using cross-population data

[0111] Assumptions can be made about the consistency of effect sizes across populations. At one extreme, we can assume that effect sizes are constant across populations. At the other extreme, if sufficient data are available, effect sizes can only be estimated using population-specific datasets, using any of the methods described above in matched populations.

[0112] The intermediate procedure is a hierarchical model that borrows information about effect sizes across groups while allowing for some variability in inferred effect sizes when supported by the data.

[0113] Steps S4 and S5: Subtraction and Machine Learning

[0114] In steps S4 and S5, the method includes using the fine-mapped prediction model to subtract the effects of the set of fine-mapped variants on the target phenotype from the input data 2 to obtain residual association data, and applying a machine learning algorithm to the residual association data to identify further predictive associations between the target phenotype and the plurality of gene variants.

[0115] In one embodiment, the machine learning algorithm includes the model proposed by LDpred and only requires summary statistics to identify the residual signal.

[0116] In this example, we define three types of effect sizes for each variant:

[0117] β irefers to the marginal effect of variant i, that is, the effect size estimated from the summary statistics when summary statistics are used;

[0118] ·p j is the probability that variant j of the refined mapping is causal (the sum of the probabilities in the credible set adds up to 1).

[0119] · refers to the causal effect of the fine-mapping variant j inferred based on the fine-mapping step, and thus corresponds to the jth (j th )

[0120] The estimated fine-mapped effect size of the fine-mapped variant on the target phenotype. Most variants will have no causal effect, but fine-mapped variants in the credible set will have a non-zero value and therefore a non-zero value of p j ;

[0121] · is the residual effect value of variant i, that is, the marginal effect of variant i, but minus the effects of the related variants in the credible set.

[0122] Use these symbols and add the standardized effect size β i After making their variances equal, we can perform the subtraction:

[0123]

[0124] where r ij captures the correlation between variants i and j, which is population-specific and is often referred to as linkage disequilibrium patterns. For all fine-mapping probabilities p j This subtraction is performed for variants that are not zero. Thus, in this embodiment, the step of subtracting the effect of the set of finely mapped variants on the target phenotype from the input data includes: subtracting a weighted sum of the effect values ​​from the estimated effect value of each of the plurality of genetic variants on the target phenotype to obtain a residual effect value for each of the plurality of genetic variants. In this embodiment, the residual association data includes the residual effect value.

[0125] The machine learning step of estimating these residual effect values ​​can then be performed in the same manner as would be performed if there were no refined mapping (i.e., if steps S2 and S3 were not performed and the machine learning step operated directly on the input data). The addition of refined mapping can lead to significant differences in the output of the machine learning process. For example, these significant differences can be seen where the output from the machine learning algorithm is used to calculate the PRS weights (defined below), as shown in Figure 2. Figure 6 (showing the PRS weights derived using the method with refined mapping) and Figure 7(shown are the differences seen between PRS weights derived using the method without fine mapping.) Furthermore, the fine-mapped signal will approximate the true causal variant, which is typically shared between groups, leading to better robustness to group differences.

[0126] The machine learning step S5 can output a set of weights for non-fine-mapped variants (i.e., variants included in the input data 2 but not identified as fine-mapped variants in step S2), which weights indicate the importance assigned to the variants based on the residual signal, while accounting for the correlation between variants. This process is significantly influenced by the population-specific correlation patterns between variants, resulting in population-specific sets of variants and weights. Thus, in embodiments where the input data are derived from multiple different populations of organisms, the correlation r between the i-th variant and the j-th variant is ij It is group-dependent.

[0127] Figure 6 and Figure 7 It shows how the machine learning / LDpred weights are broadly distributed over the region, in contrast to the fine-mapping output that accurately characterizes variants that are inferred to be causal or at least closely associated with true causal variants.

[0128] Incorporating relevant trait association data into variant-specific priors

[0129] Bayesian machine learning algorithms for gene prediction such as LDpred typically rely on prior values ​​that capture the probability that a variant is causal. Typically, the same prior value is assigned to all variants. This is called a flat prior. Low prior values ​​assigned to all variants result in a sparse model in which most weights are small or zero, while higher values ​​result in a more diffuse model in which the prediction weights are spread over a larger number of variants. An alternative to the standard LDpred model (assuming a flat prior for each variant) is to exploit cross-trait information to adjust the prior probabilities in a variant-specific manner.

[0130] One possible implementation is to use a logistic regression model; the binary outcome variable represents the consistency in the direction of marginal variant effect sizes between a well-powered GWAS for the target phenotype and a GWAS of the same target phenotype using an independent cohort of individuals.

[0131] This means that in the case where the input data is derived from multiple different genetic studies, the step of applying the machine learning algorithm to the residual association data can include using a prior probability that each of the multiple genetic variants is causally related to the target phenotype, which depends on the consistency of the strength of association between each genetic variant and the target phenotype between the different genetic studies. The strength of association from GWAS for the relevant features (e.g., P-value) is used as an input / predictor variable. The resulting linear combination of the regression coefficients (wherein each regression coefficient captures the degree of prediction of the relevant features to the target phenotype) weighted by the input variables (i.e., fitted values) and then subjected to a normalization process can serve as a variant-specific prior. As a result, the machine learning algorithm will generate higher weights for variants that have evidence of association with the features that are most relevant to the target phenotype.

[0132] Another option for defining variant-specific weights is to incorporate genomic annotations from external genomic studies other than GWAS. In such cases, the step of applying a machine learning algorithm to the residual association data involves using a prior probability for each of the multiple genetic variants that each of the multiple genetic variants is causally associated with the target phenotype, where the prior probability depends on the genomic annotations of the multiple genetic variants in the region of interest. Such functional information, such as the presence of protein-coding variants or DNA binding sites for relevant transcription factors, can be combined with the priors defined from the GWAS data to further enhance the machine learning algorithm and improve predictive performance.

[0133] Steps S6 and S7: Calculate PRS model and PRS

[0134] In one embodiment, the method further comprises calculating the PRS of the individual for the target phenotype using the fine mapping prediction model (calculated in step S3) and the further predictive correlations identified by the machine learning algorithm (in step S5). In one embodiment, the fine mapping prediction model and the further predictive correlations identified by the machine learning algorithm are used to define a PRS model (step S6). The PRS model can be used to calculate the PRS of the individual given the genetic data 3 from the individual (step S7). In an embodiment, the PRS model is a weighted sum of variants, where the weights are provided by the fine mapping prediction model and the further predictive correlations identified by the machine learning algorithm. In implementation, the PRS is calculated as follows:

[0135]

[0136] where L is the number of variants contributing to the PRS, each of which is either included in the fine-mapping prediction model or included in further prediction relevance from the machine learning algorithm, and x l is the genotype of variant 1, α lis a PRS weight that quantifies the predicted impact of variant 1 on the target phenotype (i.e., quantifies the strength of association of variant 1 with the target phenotype). The PRS weight is related to the effect size and can be assigned by a fine-mapping prediction model (as calculated in step S3) or by further predicted correlations from a machine learning algorithm (obtained in step S5).

[0137] For the fine-mapping variant, the PRS weight α l Usually related to the effect size β of variant l on the target phenotype l Direct correlation, given by the probability p that the variant is causal l The weighting is performed, so:

[0138] α l =p l β l .

[0139] If instrument characteristics are used and a relationship has been established between the effect size of the instrument and the effect size of the target (e.g., proportional beta l =Kβ′ l , where β′ l is the effect size of the instrument study), then the PRS weight is based on the instrument:

[0140] α l =p l Kβ′ l .

[0141] For variants assigned PRS weights by a machine learning algorithm, the relationship between effect size and PRS weight may be less straightforward and depend on the specifics of the algorithm.

[0142] In some embodiments, a polygenic risk score for an individual can be derived from a combination (e.g., a sum) of a first partial polygenic risk score provided by applying a fine-mapped prediction model to genetic data from the individual (e.g., based only on fine-mapped variants in the genetic data) and a second partial polygenic risk score provided by applying further predictive correlations from a machine learning algorithm to genetic data from the individual (e.g., based on variants in the genetic data in addition to the fine-mapped variants).

[0143] The machine learning steps leading to the calculation of the PRS weights can be population-specific, meaning that different PRSs can be applied to different individuals based on their ancestry, which can be identified using genetic data.

[0144] In some embodiments, input data 2 is derived from multiple different populations of organisms (e.g., different levels of lineage) and satisfies either or both of the following:

[0145] i) performing calculations on the refined mapping prediction models for the portions of input data corresponding to different groups to obtain a plurality of refined mapping prediction models matching the respective groups; and

[0146] ii) for the portions of input data corresponding to different populations, performing respectively applying a machine learning algorithm to the residual association data to obtain a set of further predicted correlations for a plurality of respective population matches.

[0147] The PRS for an individual from one of multiple populations (e.g., an individual having genes primarily from one of the different populations) can be calculated as follows. Input data is received from the individual. The PRS is calculated for the individual by performing one or both of the following operations:

[0148] i) applying the group-matched fine-map prediction model to all input data from the individual, the group-matched fine-map prediction model being matched to the individual's group; and

[0149] ii) applying the set of group-matched further predictive correlations to all input data from individuals whose group is matched to that individual.

[0150] Alternative implementation for calculating mixed individual PRS

[0151] For individuals who are admixed with two or more well-defined ancestry groups, such as African Americans, different chromosome segments can be assigned to each of these ancestries. A key motivation for fine-mapping methods is to identify causal variants and CSs that are more likely to remain consistent across populations. However, with the exception of fine-mapping, the class of prediction algorithms that includes machine learning methods relies on linkage disequilibrium patterns and, therefore, on the target population. Consequently, different PRSs will be derived for different populations.

[0152] The field of population genetics has established methods for matching individual chromosome segments to the different populations from which these segments originated. This process is called "chromosome painting." To correctly handle admixed individuals, we apply this chromosome painting step to the genotype data of related individuals. Instead of assigning individuals to a single population, we construct a mixed PRS that considers maternal and paternal chromosome copies separately, which applies the relevant, population-specific PRS to the appropriate chromosome segments.

[0153] In this type of implementation, input data is received from an individual (having a mixture of genes from different populations). The PRS for the individual is calculated by performing one or both of the following operations:

[0154] i) matching each of the plurality of group-matched fine mapping prediction models to a corresponding portion of the input data that is group-matched to the group-matched fine mapping prediction model, and applying each matched fine mapping prediction model to the corresponding portion of the input data; and

[0155] ⅱ) matching each of the plurality of sets of group-matched further predictive correlations with a corresponding portion of the input data, the corresponding portion of the input data being matched to a group in the set of group-matched further predictive correlations, and applying each matched set of further predictive correlations to the corresponding portion of the input data.

[0156] In practice, fine-mapped prediction models are expected to be largely consistent across populations, so that the set of fine-mapped variants, and even in some cases, the effect sizes of the fine-mapped variants, will be unique, with cross-population information used to make them correct. Therefore, in the above-described method, it is expected to be most valuable to match the populations for sets of further predicted correlations. Therefore, in one embodiment, a fine-mapped prediction model is established by combining data from multiple available population data sets for one or both of the following: i) the selection of fine-mapped variants and ii) the effect sizes associated with these variants. In such an embodiment, a polygenic risk score can be obtained by applying a shared population-consistent fine-mapped prediction model (i.e., a fine-mapped prediction model that is valid for multiple individuals, regardless of which population or populations they belong to) to input data from the individual, wherein further predicted correlations are established only in a population-specific manner.

Claims

1. A computer-implemented method for analyzing genetic data about an organism to calculate a polygenic risk score for the organism, the method comprising: receiving input data comprising strengths of association between one or more phenotypes including a target phenotype and a plurality of genetic variants in a region of interest of a genome of the organism; applying a fine-mapping algorithm to all or a subset of the input data to identify one or more independent phenotype-variant associations within the region of interest, comprising: identifying, for each association, a set of one or more fine-mapped variants from the plurality of genetic variants, and determining, for each fine-mapped variant, an estimated probability of having a causal relationship with the phenotype-variant association, the probabilities of the fine-mapped variants in the set summing to 1; computing a fine-mapped prediction model based on the input data and the set of fine-mapped variants, the fine-mapped prediction model quantifying the effect of the set of fine-mapped variants on the target phenotype; subtracting the effects of the set of fine-mapped variants on the target phenotype from the input data using the fine-mapped prediction model to obtain residual association data; applying a machine learning algorithm to the residual association data to identify further predictive associations between the target phenotype and the plurality of genetic variants; and A polygenic risk score for the individual for the target phenotype is calculated by applying the fine-mapped prediction model and the further predictive correlations identified by the machine learning algorithm to genetic data from the individual. 2 . The method of claim 1 , wherein the strength of association comprises an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error of each of the estimated effect sizes.

3. The method of claim 1 , wherein the step of receiving input data comprises: receiving individual-level data comprising a genotype and a corresponding phenotype for each of a plurality of individuals; and Using the individual-level data, an estimated effect size of each of the plurality of genetic variants on the target phenotype and a standard error of each of the estimated effect sizes are determined.

4. The method of claim 1 , wherein identifying the set of refined mapping variants is performed using an iterative approach, wherein each iteration comprises: identifying, based on the input data, a fine-mapped variant in a region of the genome that differs from any previously identified fine-mapped variant; updating the input data to account for the effects of the identified fine-mapped variants on the target phenotype using a correlation matrix between the genetic variants within the genomic region; and Determines whether to perform further iterations based on updated input data.

5. The method of claim 1 , wherein identifying the set of finely mapped variants comprises using a plurality of instrument features known to affect the target phenotype, the use of the instrument features comprising: determining a set of fine-mapping variants for said instrument features; and Based on the relationship between the plurality of instrument features and the target phenotype, it is determined whether each of the one or more fine-mapped variants for the instrument features is to be included in a set of fine-mapped variants for the target phenotype.

6. The method of claim 1, wherein identifying a set of fine-mapping variants comprises identifying a set of fine-mapping variants for one or more direct causal instrumental traits known to affect the target phenotype.

7. The method according to claim 5 or 6, wherein calculating the refined mapping prediction model comprises: determining, for the one or more instrument features, an effect size of the set of fine-mapped variants on the one or more instrument features, and An effect size on the target phenotype for each of the fine-mapped variants for the instrument feature included in the set of fine-mapped variants for the target phenotype is determined based on a predetermined relationship between the effect size on the instrument feature and the effect size on the target phenotype.

8. The method according to claim 1, wherein: The association strength includes an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error of each of the estimated effect sizes; and The fine-mapped prediction model comprises a fine-mapped effect value for each of the fine-mapped variants on the target phenotype, the fine-mapped effect value being calculated from the estimated effect values ​​of the fine-mapped variants given an estimated probability that the fine-mapped variant is causal for the phenotype-variant association.

9. The method of claim 1, wherein the effect of the set of finely mapped variants on the target phenotype is inferred using a machine learning algorithm.

10. The method of claim 9, wherein the set of fine-mapped variants further comprises one or more variants known to have a high causal likelihood for the target phenotype.

11. The method according to claim 1 , wherein: The association strength includes an estimated effect size of each of the plurality of genetic variants on the target phenotype, and a standard error of each of the estimated effect sizes; and The step of subtracting the effects of the set of finely mapped variants on the target phenotype from the input data comprises obtaining a residual effect value for each of a plurality of genetic variants in the input data, the residual association data comprising the residual effect value, where, after appropriate renormalization of the effect sizes to ensure equal variances, the gene variants The residual effect value It is given by: in The gene variant The estimated marginal effect value of is the number of fine-mapping variants, is a variant is the probability of causality, It is The fine-mapped effect size of each fine-mapped variant on the target phenotype, It is the Fine-mapped variants and genetic variants The correlation between them.

12. The method of claim 1 , wherein the input data originates from a plurality of different genetic studies, and the step of applying a machine learning algorithm to the residual association data comprises: A prior probability of each of the plurality of genetic variants being causally associated with the target phenotype is used, the prior probability being dependent on the consistency of the strength of association between each genetic variant and the target phenotype across the different genetic studies.

13. The method of claim 1 , wherein the step of applying a machine learning algorithm to the residual association data comprises using a prior probability of each of the plurality of genetic variants being causally associated with the target phenotype, the prior probability being dependent on the genomic annotation of the plurality of genetic variants in the region of interest.

14. The method of claim 1, wherein the polygenic risk score is given by the weighted sum of: in Yes the number of contributing variants, each variant included in the fine-mapping prediction model or included in further prediction relevance from the machine learning algorithm, Quantized variants a strength of association to the target phenotype, the strength of association being specified by the fine-mapped prediction model or by a further predicted correlation from the machine learning algorithm, and is a variant genotype.

15. The method of claim 1 , wherein the polygenic risk score for the individual is derived from a combination of a first partial polygenic risk score provided by applying the fine-mapping prediction model to genetic data from the individual and a second partial polygenic risk score provided by applying further predictive correlations of the machine learning algorithm to genetic data from the individual.

16. The method of claim 1, wherein the input data originates from a plurality of different populations of the organisms and satisfies either or both of the following: performing calculations of fine mapping prediction models on portions of the input data corresponding to different groups respectively to obtain a plurality of fine mapping prediction models matching the respective groups; and The machine learning algorithm is applied to the residual association data for portions of the input data corresponding to different populations to obtain a set of further predicted correlations for a plurality of respective population matches.

17. The method according to claim 16, further comprising: receiving input data from an individual having a mixture of genes from the different populations; and A polygenic risk score is calculated for the individual by performing one or both of the following: matching each of a plurality of group-matched fine mapping prediction models to a corresponding portion of the input data that is group-matched to the group-matched fine mapping prediction model, and applying each matched fine mapping prediction model to the corresponding portion of the input data; and Each of a plurality of sets of further predicted correlations for group matches is matched to a corresponding portion of the group matches in the set of further predicted correlations for group matches for the input data, and each matched set of further predicted correlations is applied to the corresponding portion of the input data.

18. The method of claim 17, wherein matching of each of a plurality of population-matched further predictive correlation sets is performed without performing matching of each of a plurality of population-matched fine-mapped prediction models, the calculation of the polygenic risk score comprising applying a shared population-consistent fine-mapped prediction model to input data from the individual.

19. The method according to claim 16, further comprising: receiving input data from an individual having genes primarily from one of the different populations; and A polygenic risk score is calculated for the individual by performing one or both of the following: applying a population-matched fine-mapped prediction model to all input data from the individual, the population-matched fine-mapped prediction model being population-matched to the individual; and A set of further predicted correlations of the population match to the population match of the individual is applied to all input data from the individual.

20. The method of claim 19, wherein applying the set of population-matched further predictive correlations is performed without applying the population-matched fine-mapping prediction model, the calculation of the polygenic risk score comprising applying a shared population-consistent fine-mapping prediction model to the input data from the individual.

21. The method of claim 1, wherein identifying the one or more fine-mapped variants by the fine-mapping algorithm takes into account associations between the plurality of genetic variants and phenotypes other than the target phenotype.

22. The method of claim 1, wherein the organism is a human.

23. An apparatus for analyzing genetic data about an organism to calculate a polygenic risk score for the organism, the apparatus comprising: a receiving unit configured to receive input data comprising association strengths between one or more phenotypes including a target phenotype and a plurality of gene variants in a region of interest of a genome of the organism; and A data processing unit configured as follows: applying a fine-mapping algorithm to all or a subset of the input data to identify one or more independent phenotype-variant associations within the region of interest by identifying, for each association, a set of one or more fine-mapped variants from the plurality of genetic variants, and determining, for each fine-mapped variant, an estimated probability of having a causal relationship with the phenotype-variant association, the probabilities of the fine-mapped variants in the set summing to one; computing a fine-mapped prediction model based on the input data and the set of fine-mapped variants, the fine-mapped prediction model quantifying the effect of the set of fine-mapped variants on the target phenotype; subtracting the effects of the set of fine-mapped variants on the target phenotype from the input data using the fine-mapped prediction model to obtain residual association data; applying a machine learning algorithm to the residual association data to identify further predictive associations between the target phenotype and the plurality of genetic variants; and A polygenic risk score for the individual for the target phenotype is calculated by applying the fine-mapped prediction model and the further predictive correlations identified by the machine learning algorithm to genetic data from the individual.

24. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 22.

Images