COMPUTER IMPLEMENTED METHOD AND APPARATUS FOR ANALYZING GENETIC DATA - Patent application

By integrating fine-mapping algorithms with machine learning, the method addresses the limitations of PRS construction from summary statistical data, improving accuracy and robustness by identifying causative genetic variants and accounting for population-specific correlations.

JP7593999B2Active Publication Date: 2024-12-03ゲノミクス リミテッド
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2022513533
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2019-08-28
Filing Date
2020-08-28
Publication Date
2024-12-03
Estimated Expiration
2040-08-28

AI Technical Summary

Technical Problem

Existing methods for constructing polygenic risk scores (PRS) from summary statistical data are limited by uncertainties in linkage disequilibrium patterns, particularly across different populations, leading to reduced robustness and accuracy.

Method used

A computer-implemented method that combines fine-mapping algorithms with machine learning techniques to identify causative genetic variants, accounting for residual correlations and population-specific linkage disequilibrium patterns, using both summary statistical and individual-level data to enhance predictive accuracy.

Benefits of technology

The method improves the accuracy and robustness of polygenic risk scores by identifying key genetic variants and accounting for population-specific correlations, enhancing predictive power for disease risk and treatment response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007593999000012
    Figure 0007593999000012
  • Figure 0007593999000013
    Figure 0007593999000013
  • Figure 0007593999000014
    Figure 0007593999000014
Patent Text Reader

Abstract

The present disclosure relates to the analysis of genetic data. In one configuration, a method operates on input data including strengths 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. A set of one or more fine-mapped variants is identified for each association. A fine-mapping predictive model is calculated based on the input data and the set of fine-mapped variants. The effects of the set of fine-mapped variants on the target phenotype are 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.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to the analysis of genetic and phenotypic data relating to 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 subject's phenotype. [Background technology]

[0002] A PRS is a quantitative summary of the contribution of an organism's genetic DNA to the phenotype it may exhibit. A PRS can include all DNA variants associated (directly or indirectly) with the phenotype of interest, or a component part of the phenotype can be used if it is associated with a particular aspect of the organism's biological attributes (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 an organism to estimate aspects of the organism's past, present, and future biological attributes. In the context of improving human health and healthcare, PRSs have a variety of practical applications, including but not limited to predicting risk of developing a disease or phenotype, predicting age of onset of a phenotype, predicting disease severity, predicting disease subtypes, predicting response to treatment, selecting screening strategies appropriate for an individual, selecting appropriate drug interventions, and setting prior probabilities for other predictive algorithms. A PRS can be used directly as an input source in the application of artificial intelligence and machine learning approaches to make predictions or classifications from other high-dimensional input data (e.g., imaging). PRS can be used to help train these algorithms, for example, to identify predictive measures based on non-genetic data. PRSs are not only useful in making predictive statements about individuals, but can also be used to identify cohorts of individuals, including but not limited to the above applications, by calculating the PRS of a large number of individuals and then grouping the individuals based on the PRS. PRSs can also aid in the selection of individuals for clinical trials, for example optimizing the study design by recruiting individuals who are more likely to develop the relevant disease or phenotype, thereby enhancing the evaluation of the efficacy of new treatments. PRSs convey information not only about the individual for whom the PRS is calculated, but also about their relatives (who share a small portion of their genetic DNA). Information about the impact of an individual's DNA on the individual's phenotype can be derived from association assessments of the potential impact of transmitting a particular combination of DNA variants. In the following, we focus on the analysis of the recent large amount of information derived from genetic association studies (GAS).These studies systematically evaluate the potential contribution of DNA variants to the genetic basis of a phenotype.

[0003] Since the mid-2000s, GAS (usually genome-wide association studies (GWAS), or association studies targeting a single variant, or variants in a region of the genome, or GWAS restricted to a specific region of the genome) have been performed on thousands of (mostly human) phenotypes in millions of individuals, generating billions of potential associations between genotypes and phenotypes. The resulting raw data is often then simplified to generate summary statistical data. GAS summary statistical data consists, for each genetic variant (whether imputed or observed), of the estimated effect size of the genetic variant on the GAS phenotype and the standard error of the estimated effect size. In other cases, individual-level data, consisting of the complete genetic profile of the individuals in the study and information about their phenotypes, may be directly available. However, individual-level data is usually less widely available due to requirements regarding privacy of individuals' data.

[0004] In the following, phenotype is taken to be synonymous with a single study. However, very often data are obtained from several different studies of the same or similar phenotype, or from a single cohort in which several different phenotypes are measured.

[0005] A PRS is composed of a collection of effects for many genetic variants, each with a typically small individual effect, to build an aggregate prediction of the trait of interest. Variants included in such a score can be either "causal variants," meaning that the variant directly affects the trait (weakly but directly), or "tagging variants," meaning that the variant is strongly correlated with other unknown causal variants, but does not itself have a direct effect on the phenotype.

[0006] PRS can be calculated using individual-level data or summary statistical data. Although strategies for PRS construction are expanding, a widely accepted general approach to construct an accurate PRS consists of deconvoluting the signal in all regions of association by investigating the combination of variants that best captures the underlying biological association. This process describes, for each association, which variants are likely to be directly causative by assigning a probability weight to each variant. This process is called "fine mapping" and several strategies have been previously proposed to realize this task (see, for example, Benner et al., Bioinformatics 2016, 15;32(10):1493-1501).

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

[0008] Some tools for constructing PRSs are designed to take advantage of summary statistical data. One such approach refers to pruning and thresholding, where the most relevant variants are selected to contribute to the PRS and the highly correlated variants are removed. The most relevant variant among the remaining variants is then selected and the process is repeated until the significance of the remaining variants falls below a predefined threshold. A further 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 once variants are selected or removed.

[0009] The advantage of strategies based on summary statistical data is that the lack of restrictions on sharing individual-level data means that much larger sample sizes can be made available to the scientific community, and therefore many current PRS designs are based on these large summary statistical data sets.

[0010] However, for all summary statistical data based methods, correlated variants are handled by referencing external data sources that describe what the correlations between variants are expected to be. The pattern of correlation between genetic variants is called linkage disequilibrium (LD). The correlations in these external data sources do not perfectly match the correlations obtained from the individual level data used to generate the summary statistical data. This introduces further uncertainty into the fine mapping procedure due to the uncertainty about what the correct correlations should be. Thus, fine mapping based on summary statistical data is fundamentally limited by the uncertainty about the underlying LD patterns.

[0011] Another limitation of relying on external datasets to describe patterns of LD is that different populations have distinct patterns of LD. Thus, estimates made for one population are unlikely to be accurate for different populations. In other words, PRSs derived based on reference LD datasets offer limited robustness to population variation. Summary of the Invention [Problem to be solved by the invention]

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

[0013] Although patterns of LD differ between populations, a variant affecting a trait or disease in one population will generally also affect the same trait / disease in a different population. Thus, using fine-mapping techniques to identify the causative variant, or a set of variants that may contain or tag the causative variant, makes PRS more accurate, especially by making it more robust to population variation.

[0014] However, not all variants can be fine-mapped, especially the many variants with small effects on the target phenotype. Therefore, alternative techniques that focus only on the prediction problem, without the need to make precise statements about which variants are causative, are also useful for PRS construction.

[0015] Thus, accurate derivation of a PRS, which may have high clinical utility in predicting disease or an individual's response to a particular drug or treatment, will benefit from statistical techniques that exploit the advantages of fine mapping, and also allow for the use of alternative machine learning techniques if necessary. [Means for solving the problem]

[0016] According to aspects of the invention, there is provided a computer-implemented method of analyzing genetic data for an organism to obtain information about 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 in 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 in 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 being causative of the phenotype-variant association, wherein the probabilities of the fine-mapped variants in the set sum to one; calculating a fine mapping predictive model based on the input data and the set of fine-mapped variants, the fine mapping predictive model quantifying an effect of the set of fine-mapped variants on the target phenotype; subtracting the effect of the set of fine-mapped variants on the target phenotype from the input data using the fine mapping predictive model 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 plurality of genetic variants.

[0017] By using fine-mapping techniques to identify fine-mapped variants that may be causative of the target phenotype, and further analyzing the residual signal (via the residual association data) that remains after accounting for the effect of the fine-mapped variants, the method can take into account additional weak correlations that may be present in the data. Inclusion of these additional correlations improves the predictive accuracy of the model.

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

[0019] In one embodiment, receiving the input data includes receiving individual-level data for each of a plurality of individuals, the data including a genotype and a corresponding phenotype, and using the individual-level data to determine 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. Individual-level data can be used in some embodiments because it is not subject to underlying assumptions regarding correlations between variants within regions that may be present in summary statistical data, thereby reducing the possibility of introducing unintended bias or error.

[0020] In one embodiment, identifying a set of fine-mapped variants is performed using an iterative method, with each iteration including: identifying fine-mapped variants in a region of the genome that are distinct from previously identified fine-mapped variants based on input data; updating the input data to take into account the effects of the already identified fine-mapped variants on the target phenotype using a matrix of correlations between genetic variants in the region of the genome; and determining whether to perform further iterations based on the updated input data. Using an iterative approach, multiple fine-mapped variants can be identified from residual signals not accounted for by a single fine-mapped variant, thereby maximizing the use of signals present in the summary data.

[0021] In one embodiment, identifying the set of fine-mapped variants includes using a plurality of predisposing (instrument) traits known to affect the target phenotype, and using the predisposing traits includes determining a set of fine-mapped variants for the predisposing traits, and determining whether to include each of one or more of the fine-mapped variants for the predisposing traits in the set of fine-mapped variants for the target phenotype based on the relationship between the plurality of predisposing traits and the target phenotype. The relationship between the plurality of predisposing traits and the target phenotype can take into account potentially complex patterns of association between the predisposing traits and the target phenotype. Alternatively or additionally, in other embodiments, identifying the set of fine-mapped variants includes identifying a set of fine-mapped variants for one or more causative predisposing traits known to affect the target phenotype. In such cases, it may not be necessary to take into account complex patterns of association between the plurality of predisposing traits and the target phenotype.

[0022] The use of predisposing traits can improve the accuracy of determining fine-mapped variants for a phenotype when a genetic variant has only a small effect on the target phenotype and a large effect on the predisposing trait.

[0023] In an embodiment, the step of calculating the fine-mapping predictive model includes the steps of determining the effect size of the set of fine-mapped variants for one or more predisposing traits on one or more predisposing traits, and determining the effect size of each of the fine-mapped variants for the predisposing traits included in the set of fine-mapped variants for the target phenotypes based on a predetermined relationship between the effect size of the predisposing trait and the effect size of the target phenotype. The predisposing traits can also be used to improve the estimation of the effect size when the effect of the genetic variant on the predisposing trait is larger than the effect on the target phenotype. This can be particularly effective when the relationship between the predisposing trait and the target phenotype is itself well characterized.

[0024] In an embodiment, the effect of the set of fine-mapped variants on the target phenotype is estimated using a machine learning algorithm.In such an embodiment, the set of fine-mapped variants and their corresponding marginal effect size are input into a machine learning algorithm to generate an effect size, so that the residual association data is the marginal effect size corresponding only to the set of fine-mapped variants.The set of fine-mapped variants can further include one or more variants that are known to be likely to cause the target phenotype.

[0025] This reduces the number of genetic variants to which the machine learning algorithm is applied by focusing on those variants that are most likely to be causative, thereby reducing the computational burden and improving the efficiency of the method.

[0026] In an embodiment, 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, and the fine-mapped predictive model comprises a fine-mapped effect size on the target phenotype for each of the fine-mapped variants, the fine-mapped effect sizes being calculated from the estimated effect sizes of the fine-mapped variants taking into account the estimated probability that the fine-mapped variant is causal to the phenotype-variant association. Adjusting the effect size of the fine-mapped variant according to its causal probability ensures that the significance of the fine-mapped variant is not overestimated when the certainty of causation is low.

[0027] In an embodiment, 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, and subtracting the effects 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, and the residual association data comprises residual effect sizes, and after appropriate renormalization of the effect sizes to ensure equal variances, the residual effect size for genetic variant i is

[0028]

number

[0029]

number

[0030]

number

[0031] The above approach allows the residual effects of variants in targeted regions of the genome to be clearly identified for further analysis by machine learning algorithms.

[0032] In an embodiment, the input data is derived from a plurality of different genetic studies, and applying the machine learning algorithm to the residual association data includes using a prior probability for each of the plurality of genetic variants causing the target phenotype that depends on the consistency of the strength of association between each genetic variant and the target phenotype across the different genetic studies. By using a non-flat prior distribution for the machine learning algorithm, the method can improve its accuracy by taking into account further information regarding the certainty that certain data can be trusted.

[0033] In an embodiment, applying the machine learning algorithm to the residual association data includes using a prior probability for each of a plurality of genetic variants being causative of the target phenotype that depends on a genomic annotation of the plurality of genetic variants in the region of interest. The inclusion of the genomic annotation provides further data regarding the likelihood that a particular variant is causative of the target phenotype, thereby improving the determination of the effect size.

[0034] In an embodiment, the method further comprises calculating an individual's polygenic risk score for the target phenotype using the fine mapping predictive model and additional predictive correlations identified by the machine learning algorithm.

[0035] Taking into account additional correlations identified by the machine learning algorithm improves the accuracy of the PRS by allowing the method to take into account residual signal not explained by the set of fine-mapped variants.

[0036] In an embodiment, the input data is derived from a plurality of different populations of organisms, and either or both of the following are satisfied: the step of calculating the fine mapping predictive models is performed separately for portions of the input data corresponding to the different populations to obtain a plurality of respective fine mapping predictive models of population matches, and the step of applying the machine learning algorithm to the residual associated data is performed separately for portions of the input data corresponding to the different populations to obtain a plurality of respective sets of further predictive correlates of population matches.

[0037] By providing a fine mapping predictive model and a set of additional residual predictive correlations matched to a particular population, the method is able to take into account possible variations in linkage disequilibrium (correlation between variants) within a target region of the genome.

[0038] In an embodiment, the method further includes receiving input data from an individual having genes derived from a mixture of different populations, and calculating a polygenic risk score for the individual by either or both of: matching each of a plurality of population-matched fine mapping predictive models to a corresponding portion of the input data matched to the populations of the population-matched fine mapping predictive models and applying each matching fine mapping predictive model to the corresponding portion of the input data, and / or matching each of a plurality of sets of population-matched further predictive correlations to a corresponding portion of the input data matched to the populations of the set of population-matched further predictive correlations and applying each set of matching further predictive correlations to the corresponding portion of the input data.

[0039] By calculating a polygenic risk score for an individual using multiple fine mapping predictive models and / or a set of additional residual predictive correlations matching multiple distinct respective portions of input data from the individual, the method can provide a more accurate predicted risk score that takes into account systematic differences in correlations between variants associated with different populations.

[0040] In an embodiment, the method includes receiving input data from an individual having genes that are primarily derived from one of the different populations, and calculating a polygenic risk score for the individual by either or both of: applying a population-matched fine-mapping predictive model to all of the input data from the individual and matching the population-matched fine-mapping predictive model to the individual's population, and / or applying a set of population-matched further predictive correlations to all of the input data from the individual and matching the set of population-matched further predictive correlations to the individual's population.

[0041] By calculating a polygenic risk score using a fine mapping predictive model and an additional set of residual predictive correlations matched to populations of individuals, the method can provide a more accurate predicted risk score that takes into account systematic differences in correlations between variants associated with different populations.

[0042] In an embodiment, the step of identifying one or more fine-mapped variants by a fine-mapping algorithm considers association of multiple genetic variants with phenotypes other than the target phenotype.

[0043] The use of information about correlations with other phenotypes maximizes the amount of available information that can be used to identify fine-mapped variants and their effect sizes, further improving the accuracy of the method results.

[0044] According to another aspect, there is provided an apparatus for analysing genetic data relating to an organism to obtain information about the organism, the apparatus comprising: a receiver configured to receive 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 in the genome of the organism; and a data processor 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-mapped variants from the plurality of genetic variants and determining, for each fine-mapped variant, an estimated probability of being causative of the phenotype-variant association, such that the probabilities of the fine-mapped variants in the set sum to one; calculate a fine mapping predictive model that quantifies an effect of the set of fine-mapped variants on the target phenotype based on the input data and the set of fine-mapped variants; subtract the effect of the set of fine-mapped variants on the target phenotype from the input data using the fine mapping predictive 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.

[0045] Embodiments of the present invention will now be further described, by way of example only, with reference to the accompanying drawings, in which: [Brief description of the drawings]

[0046] [Figure 1] 1 is a flow chart illustrating a method for analyzing genetic data to obtain information about an organism. [Diagram 2] FIG. 1 illustrates an apparatus for analyzing genetic data to obtain information about an organism. [Diagram 3] 1 is a graph showing a comparison of effect sizes between coronary artery disease (CAD) and low density lipoprotein (LDL). [Figure 4] FIG. 13 is four graphs representing steps in a stepwise forward regression analysis to identify four respective independent association signals to identify fine-mapped variants associated with LDL in the LPA region of chromosome 6. [Diagram 5] FIG. 5 is a graph showing estimates of LDL binding effect size versus LDL marginal effect size for the four association signals identified in FIG. 4. [Figure 6] FIG. 6 is a graph showing PRS weights of CAD for the LPA region of chromosome 6, obtained by applying the LDpred machine learning algorithm to the residual association data obtained using the analysis of FIGS. 4 and 5 . [Figure 7] FIG. 13 shows the PRS weights of CAD for the LPA region of chromosome 6, obtained by applying the LDpred machine learning algorithm directly to the CAD variant data without a preceding fine mapping step. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0047] The embodiments of the present disclosure relate to computer-implemented methods of analysing genetic data relating to an organism to obtain information about the organism. Figure 1 shows a framework of these methods. Figure 2 shows an apparatus 6 for carrying out the methods.

[0048] In step S1, input data 2 is received (e.g., by a receiving unit 8 of the device 6). The receiving unit 8 may comprise a data communication interface. The data communication interface allows the input data 2 to be provided to a data processing unit 10 of the device 6. The data processing unit 10 may comprise any suitable combination of computer hardware, firmware, and / or software configured to perform the data processing functions described below. A computer program, optionally provided on a computer readable medium, may be provided containing instructions for performing any of the methods described below. Although the device 6 is shown as a stand-alone unit (e.g., a single PC or workstation), this is not required. In other embodiments, the device 6 comprises a distributed computing system including multiple computers connected by a network.

[0049] In some embodiments, input data 2 includes the strength of association of one or more phenotypes, including the target phenotype, with a plurality of genetic variants in a region of interest in the genome of the organism. In some embodiments, input data 2 includes either or both of GWAS summary statistics and individual-level data. As described in more detail below, the method can use input data 2 to (i) identify variants with high confidence that they have a direct causal effect on the target phenotype (referred to as fine-mapped variants), and (ii) obtain residual association data (referred to as and / or derived from residual signals) after conditioning on the high confidence variants, and / or predict individual trait risk (e.g., in the form of a PRS). This method is particularly advantageous when used in embodiments in which the organism is a human.

[0050] The target phenotype can be any phenotype of interest that has been the subject of a GWAS or for which relevant individual-level genetic data is available. There are many examples of such phenotypes, including levels and modulation of expression of genes (and associated nucleotide sequences), epigenetic features (e.g., nucleotide modifications, chromosomal conformation), abundance levels of proteins or peptides, function and / or molecular structure of proteins or peptides, molecular weights of organisms (e.g., drugs, hormones, DNA or RNA molecules, metabolites, vitamins), 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, absolute neutrophil count), tissue morphology and function (e.g., bone mineral density, hair color), organ and organ system morphology and function (e.g., left ventricular ejection fraction, pulmonary circulation ... These include: pulmonary function, pulmonary function (e.g., fractional, forced vital capacity), response to external stimuli (e.g., light, sound, touch, or other sensory input), response to exposure to substances or pathogens (e.g., dietary intake, drugs, 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), onset, history, and prognosis of disease or conditions (e.g., diabetes, cardiovascular disease, obesity), measurable anatomical characteristics (e.g., body mass index, lean muscle mass, percentage of body fat), measurable physiological or functional characteristics (e.g., heart rate, blood pressure, intelligence), and measurable psychological or cognitive characteristics (e.g., indices of fluid intelligence, psychotic symptoms). Any of these measurements can be absolute or relative. Phenotypes are also often referred to as traits.

[0051] In step S2, a fine mapping algorithm is applied to all or a subset of the input data 2. In an embodiment, the fine mapping step identifies high confidence causal variants, thereby obtaining a set of fine mapped variants. Further details regarding step S2 are provided below.

[0052] In step S3, a fine-mapping predictive model is calculated based on the input data 2 and the fine-mapped variants. The fine-mapping predictive model quantifies the effect size of the fine-mapped variants on the target phenotype. Effect size refers to how much a given variant affects disease risk (or, more generally, the "risk" of having or developing a given phenotype). For example, an effect size of 1.2 means a 20% increase in risk per risk allele (which may be coded as 0, 1, or 2 for each individual) for that given variant. Thus, the quantification of effect size allows the fine-mapping predictive model to make predictions about individuals based on genetic data from the individuals. Further details regarding S3 are provided below.

[0053] In step S4, the fine-mapping predictive model is used 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. Further details regarding step S4 are provided below.

[0054] In step S5, a machine learning algorithm is applied to the residual association data to identify further predictive correlations between the target phenotype and the multiple genetic variants in the input data 2. In the specific example below, a machine learning algorithm called LDpred is used. LDpred is known in the art of fine mapping and PRS generation. Implementation software is available at https: / / github.com / bvilhjal / ldpred. The further predictive correlations can quantify effect sizes associated with variants other than the fine-mapped variants (after taking into account the effect of the fine-mapped variants), thereby improving predictions for an individual relative to when the fine-mapping predictive model is applied only to genetic data from the individual.

[0055] In step S6, the PRS model is evaluated. The PRS model may be derived in part from the fine mapping predictive model of step S3 and in part from the machine learning further predictive correlation performed in step S5. As described below, the combination of the fine mapping predictive model and the machine learning further predictive correlation may define a means for calculating the PRS, which takes the form of a weighted sum over variants, with the weights of the fine-mapped variants being provided by the fine mapping predictive model and the weights of the other variants being provided by the machine learning further predictive correlation. This is possible if the trained machine learning algorithm can be interpreted in terms of such weighted sums over variants. In other embodiments, the trained machine learning algorithm is more complex and may therefore be represented differently as part of the PRS model.

[0056] The PRS model calculated in step S6 can be used to calculate a PRS score based on genetic data from the individual. The PRS model can be output (e.g., via a data communication interface of device 6 of FIG. 2) as data representative of the PRS model. Thus, the steps up to step S6 (including training of the machine learning algorithm) can be performed on one device 6, and subsequent steps (e.g., to calculate the individual's PRS score) involving use of the PRS model can be performed on another device (not shown) that includes 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 can be performed on the same device 6 that calculated the PRS model.

[0057] In step S7, a PRS score for 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.

[0058] The calculated PRS model constitutes information about organisms at a general level (e.g., about humans in general) in the sense that a PRS score can be calculated from genetic information obtained from any individual, whereas the PRS score constitutes information about a specific individual organism (e.g., a single human subject).

[0059] [Example of application scenario] 3-7 illustrate the use of the method of FIG. 1 in an example scenario and are referenced in the more detailed description of the method steps presented below.

[0060] Figure 3 shows how the effect size of LDL correlates with the effect size of CAD at 95 separate loci associated with LDL. This correlation is observed in part because LDL is understood to have a near direct causal effect on CAD. Therefore, we call LDL a good predisposing factor for CAD, which means that information about LDL can be used to improve the accuracy of the PRS for CAD.

[0061] Figure 4 shows the results of fine-mapping LDL using the method of Figure 1 in the LPA region of chromosome 6 using an established method (stepwise forward regression). Each regression step identifies an additional independent phenotype-variant association, for a total of four independent phenotype-variant associations. In each plot, 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-mapping probability (<1%) are in grey. The first LDL-associated signal has four fine-mapped variants with a backward probability >1%, while the remaining three LDL-associated signals identify a single fine-mapped variant with a fine-mapping probability >1%.

[0062] Figure 5 shows that for this example LDL-predisposing trait, the four joint effect size estimates estimated from the four independent phenotype-variant associations shown in Figure 4 differ insignificantly from the four marginal effect size estimates.

[0063] Figures 6 and 7 show the PRS weights of CAD derived for the same LPA region of chromosome 6. In Figure 6, we show that the fine-mapped CAD variants were extrapolated from the fine mapping and effect sizes of LDL (Figures 4 and 5) and subtracted from the CAD data before LDpred analysis to capture the residual signal (representing further predictive correlations). Thus, Figure 6 combines the PRS weights derived from LDL fine mapping (black) with the PRS weights derived from the LDpred residual signal (grey). This is in contrast to Figure 7, where a standard LDpred analysis strategy based only on CAD was applied, without an initial fine mapping step. One consequence of the limited accuracy of the process without an initial LDL fine mapping is the failure to detect the fourth signal in Figure 7. This is because the CAD data alone is insufficient to characterize this association.

[0064] Further examples of implementation details of steps S2-S7 of FIG. 1 are provided below, where appropriate, with reference to the above example application scenarios for illustrative purposes.

[0065] [Step S2: Fine mapping] As discussed 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 the region of interest. Identifying one or more independent phenotype-variant associations within the region of interest may include identifying high confidence fine-mapped variants that are variants with high confidence that are causative variants or tagging variants of causative variants for the phenotype of interest. For each association, a set of one or more fine-mapped variants is identified from the plurality of genetic variants.

[0066] Figure 4 illustrates the application of the fine mapping algorithm in the context of the example application scenario above, where the fine mapping algorithm identifies four independent phenotype-variant associations for LDL within a region of chromosome 6 (one for each of the four plots shown).

[0067] Fine mapping algorithms are typically designed to capture the causal biological attributes underlying a target phenotype by locating a causal variant or a set of reliable variants that contain or tightly tag the causal variant. Fine mapping algorithms are in contrast to alternative purely predictive approaches, which are typically based on machine learning techniques such as LASSO, random forests, or neural networks, that capture predictive signals without providing a separate summary of the data that maps to the underlying biological attributes.

[0068] The phenotype-variant association is independent in the sense that even if some degree of correlation may exist between the two identified variants, the association of the second fine-mapped variant with the phenotype is not solely due to the correlation with the first fine-mapped variant associated with the phenotype. In other words, the second fine-mapped variant is associated with the phenotype even after considering or conditioning on the first fine-mapped variant associated with the phenotype. On the other hand, if one of the variants in the CS is selected and conditioned on this variant, the association in all other variants in the CS disappears, i.e., multiple associations exist only due to high correlation between the variants, and therefore multiple variants in the CS are not independent of each other.

[0069] Each independent phenotype-variant association may be associated with a single fine-mapped variant, or a credible set (CS) of (multiple) fine-mapped variants. Thus, for each association, a set of one or more fine-mapped variants is identified from the multiple genetic variants. A CS of fine-mapped variants is a set of two or more fine-mapped variants that are considered likely to be causative of the target phenotype. The method determines for each fine-mapped variant an estimated probability of being causative of the phenotype-variant association, such that the sum of the probabilities of the fine-mapped variants in the set equals 1. If only one fine-mapped variant is identified, the estimated probability is simply 1 for that fine-mapped variant. In FIG. 4, forward regression steps 2-4 show an example of identifying a single fine-mapped variant, while forward regression step 1 identifies a CS of a fine-mapped variant.

[0070] In some embodiments, the identification of one or more fine-mapped variants by the fine mapping algorithm considers the association of multiple genetic variants with phenotypes other than the target phenotype.Accordingly, the input data 2 of such an embodiment includes the strength of association of multiple phenotypes with multiple genetic variants in the target region of the genome of the organism.The use of associations with multiple phenotypes makes it easier to leverage data from multiple studies, covering a wide range of different phenotypes, and taking advantage of the fact that many traits can share the same causal variants.

[0071] In an embodiment, input data 2 includes data describing the association of individual variants with the target phenotype in the form of marginal variant effect size and standard error. In such an embodiment, the strength of association can include an estimated effect size of each of the multiple genetic variants on the target phenotype, and a standard error of each of the estimated effect sizes. The estimated effect size is the marginal variant effect size. The marginal variant effect size refers to the effect of a variant when considered alone, i.e., ignoring the effects of nearby correlated variants. For example, a tag variant can have a strong marginal effect size, but its "true" effect size is zero. This form of input data is commonly referred to as summary statistical data.

[0072] In an embodiment, applying a fine mapping algorithm to all or a subset of the input data 2 to identify one or more independent phenotype-variant associations in the region of interest includes: Using a probability model (e.g., a Bayesian statistical model) in a given DNA region (i.e., a region of the genome of an organism), studies (each containing data on the strength of association between the target phenotype and one or more genetic variants) are assigned to clusters, with each cluster assumed to have a similar pattern of causal variants. The space of possible cluster assignments is then explored using a Markov chain Monte Carlo algorithm or the like. Once a set number of iterations of assigning studies to clusters have been performed, a set of cluster features can be used to identify a single variant or a CS of genetic variants (i.e., a set of one or more fine-mapped variants) that may be responsible for the phenotype assigned to the corresponding cluster. Using this method based on a large number of phenotypes increases the power and accuracy of identifying variants that affect the phenotype. Further details of this type of method can be found in PCT Application No. PCT / GB2019 / 050525.

[0073] In some cases, the method identifies at most a single fine-mapped variant or a single CS of a fine-mapped variant for a given DNA region. However, there may be multiple independent fine-mapped variants (or corresponding multiple CSs) that may be causative within a region. Identification of these additional independent fine-mapped variants provides additional prediction of the disease or trait of interest, thus improving the ability to predict the risk of an individual developing the disease or trait.

[0074] [Alternative implementation of step S2 when only summary statistical data is available] When only summary statistical data is available, additional independent fine-mapped variants can be identified. In an embodiment, this is achieved by considering the correlation between genetic variants in a region of the genome, which is usually summarized by a "LD matrix", which is a matrix of correlations r_ij of genotypes g_i, g_j at positions i, j, often obtained from a subpopulation 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):10):1493-501) can be appropriately adapted to this setting, considering a large number of studies and phenotypes.

[0075] Another such embodiment identifies additional causal variants (referred to herein as fine-mapped variants) by updating summary statistics to take into account the effects of already identified fine-mapped variants in the DNA region, and then evaluating the residual evidence of the additional fine-mapped variants. In this case, identifying a set of fine-mapped variants is performed using an iterative method. Each iteration includes: based on input data, identifying fine-mapped variants in a region of the genome that are different from the previously identified fine-mapped variants; updating the input data to take into account the effects of the already identified fine-mapped variants on the target phenotype using a matrix of correlations between genetic variants in the region of the genome; and based on the updated input data, deciding whether to perform further iterations (e.g., stopping when the updated input data is determined to no longer contain information of interest, such as when it no longer exceeds a predefined significance threshold and / or when the P-values ​​are all relatively flat).

[0076] This approach 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 a previously identified fine-mapped variant comprises 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.

[0077] [Alternative implementation of step S2 using individual-level data] An alternative fine mapping strategy is to use individual level data to perform fine mapping. In such an embodiment, the step of receiving input data includes receiving individual level data including genotypes and corresponding phenotypes 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 a target phenotype, and a standard error of each of the estimated effect sizes. This can be achieved by using a stepwise regression method to explore the space of fine-mapped variants using forward selection, backward elimination, or a combination of the two.

[0078] Alternatively, individual level data can be used in combination with summary statistical data to exploit information obtained by fine mapping methods based on summary statistics such as those described in PCT Application No. PCT / GB2019 / 050525. One way this can be achieved is to use single fine mapped variants / CSs obtained by methods such as those described in PCT Application No. PCT / GB2019 / 050525 and condition on these in a subsequent stepwise regression step (which can use a combination of forward selection and backward elimination as described above).

[0079] Alternatively, residual summary statistics can be obtained that are derived from individual-level data and conditioned on previously identified fine-mapped variants, thus removing the effects of variants that have already been identified as high-confidence fine-mapped variants, allowing the residual correlations to be used to identify further fine-mapped variants.

[0080] This is done in a similar way to the conditioning done on summary statistics when individual level data are not available, with the major advantage that no LD information is required. These derived residual summary statistics can be used as input for a method such as that described in PCT Application No. PCT / GB2019 / 050525. This procedure can be repeated iteratively. The method can be based on using only summary statistics derived from individual level data, or can be combined with residual summary statistics derived using LD panels from studies where no individual level data exist.

[0081] Alternative implementation of step S2 using one or more predisposing traits An alternative implementation of step S2 utilizes predisposition studies, such that identifying the set of fine-mapped variants includes using one or more predisposition traits known to affect the target phenotype. A trait is defined as a predisposition for the target phenotype when it is strongly correlated with the trait of interest. Special cases are predispositions that directly cause / modify the target phenotype. For example, LDL can be considered a predisposition trait for coronary artery disease, which is a predisposition trait for overall survival. Predisposition studies provide information about the strength of association between predisposition traits and multiple genetic variants being considered with respect to the target phenotype.

[0082] In many cases, the effect of a variant on a target phenotype is too small to identify a reliable set of target phenotypes (CS). However, this effect may be sufficient to make fine mapping feasible using appropriately powered predisposition studies. In other words, it becomes easier to accurately determine whether a variant is causal for a predisposition trait because the effect of the variant on the predisposition trait is larger than the effect of the variant on the target phenotype. In that context, fine mapping and identification of causal signals provide information about a target phenotype that would otherwise be uncharacterized based solely on predisposition studies.

[0083] Based on the above findings, identifying the set of fine-mapped variants includes identifying a set of fine-mapped variants for one or more causative predisposing traits known to affect the target phenotype. This is a relatively simple method of using predisposing traits to supplement the fine mapping of step S2, but requires the availability of known causative predisposing traits. In another embodiment, identifying the set of fine-mapped variants includes using multiple predisposing traits known to affect the target phenotype. The method then includes determining a set of fine-mapped variants for the predisposing traits, and determining whether to include each of one or more of the fine-mapped variants for the predisposing traits in the set of fine-mapped variants for the target phenotype based on the relationship between the predisposing traits and the target phenotype. In this case, the relationship between the multiple predisposing traits and the target phenotype can take into account potentially complex patterns of association between the predisposing traits and the target phenotype, allowing the use of predisposing traits that are not necessarily causative predisposing traits.

[0084] Figures 3-5 show an example of fine mapping for LDL, a predisposing trait for CAD, and the fine-mapped variants identified for LDL (Figure 4) are used in a subsequent step using CAD as the phenotype of interest (Figure 6).

[0085] [Step S3: Calculation of a fine-mapping predictive model (e.g., to estimate effect sizes of fine-mapped variants)] As previously described, in step S3, the method calculates a fine-mapping predictive model based on the input data 2 and the set of fine-mapped variants (identified in step S2). The fine-mapping predictive model quantifies the effect of the set of fine-mapped variants on a target phenotype. The effect on the target phenotype may be quantified using a fine-mapped effect size for the target phenotype, where the fine-mapping predictive model consists of or includes a fine-mapped effect size on the target phenotype for each of the fine-mapped variants that takes into account correlations between the variants.

[0086] In embodiments where the strength of association includes summary statistics (e.g., the estimated effect size of each of multiple genetic variants on the target phenotype, and the standard error of each of the estimated effect sizes), the fine-mapped effect size can be obtained directly from the marginal summary statistics from a single GWAS for the target trait (i.e., the target phenotype). When a single fine-mapped variant is identified in a region, the effect size reported in the GWAS summary statistics can be used. When a variant CS is identified, the GWAS summary statistics can be weighted according to the probability that the variant is causal (relative to the remaining variants of the CS). In some embodiments, each fine-mapped effect size can thus be calculated from the estimated effect size of the fine-mapped variant (e.g., derived from input data 2) taking into account the estimated probability that the fine-mapped variant is causal to the phenotype-variant association (e.g., derived from input data 2 as a weighting, as described above). For example, the fine-mapped effect size can be derived based on multiplying the estimated effect size by the probability that the fine-mapped variant is causal.

[0087] [Alternative implementation of step S3 in the presence of correlation associations] When multiple reliable sets capturing several independent biological associations are identified within the same DNA region, it is desirable to apply corrections to the effect sizes to control for correlation between the associations. The corrected effect size is commonly referred to as the joint effect size. This is illustrated in the example of fine mapping in the LPA region of chromosome 6 previously described with reference to Figure 4. Figure 5 shows that for LDL predisposing traits, the four joint effect size estimates differ slightly from the four marginal effect size estimates. The differences can be large when the associations are closely correlated.

[0088] When multiple fine-mapped variants are associated with a trait independently of each other, there may still be some correlation between them. The marginal effect size of these independent fine-mapped variants needs to be adjusted to take into account the correlation between the variants. Thus, in other words, the joint effect size is the effect size of multiple variants for one trait taking into account the correlation between the variants, for example, the effect size of four fine-mapped variants taking into account that there is some correlation between the four fine-mapped variants in the LDL example.

[0089] This correction for joint effect size estimates can be applied using summary statistical data (as described in Yang et al., Nature Genetics 2012, 44(4):369-75), provided that the patterns of variant correlation (or LD) in population-specific DNA regions are well characterized. Alternatively, this correction can be applied using individual-level data, whereby all selected fine-mapped variants identified within a DNA region are jointly fitted using a regression model. This correction is necessary when several separate associations are associated with a set of correlated reliable variants.

[0090] Alternative implementation of step S3 using one or more predisposing traits An alternative method for estimating effect size utilizes predisposition studies. In this type of embodiment, the step S2 of identifying a set of fine-mapped variants includes determining a set of fine-mapped variants for one or more predisposition traits known to affect the target phenotype. The calculation of the fine-mapped predictive model then includes determining the effect size of the set of fine-mapped variants for one or more predisposition traits for the one or more predisposition traits, and determining the effect size of the target phenotype for each of the fine-mapped variants for the predisposition traits included in the set of fine-mapped variants for the target phenotype based on a predetermined relationship between the effect size of the predisposition trait and the effect size of the target phenotype. Since the effect of a gene variant on a predisposition trait is greater than that of the target phenotype, it is easier to estimate the effect size of the variant on the predisposition trait than on the target phenotype.

[0091] Therefore, if external data or genome-wide data allows accurate characterization of the relationship between predisposing traits and the effect size of the target phenotype, the better estimated effect size of the predisposing trait can be utilized to more accurately estimate the effect size of the target phenotype.One way to characterize the relationship between the effect size of the predisposing trait and the effect size of the target phenotype is to perform linear regression on the effect size of the variants defined to be fine-mapped for both the predisposing trait and the target phenotype.

[0092] Figure 3 shows an example of estimating the effect size relationship between LDL and CAD using a large set of LDL-associated variants, where LDL serves as a predisposing trait for CAD.

[0093] Alternative implementation of step S3 using all studies / phenotypes as potential predisposing traits An alternative to step S3 is to utilize the independent fine-mapped variants (or CS) identified for all studies used to train the probabilistic model described above (and detailed in PCT Application No. PCT / GB2019 / 050525), which results in a set of variants that are likely to be causative of at least one disease / trait.

[0094] A machine learning algorithm of step S5 can then be applied to this set of variants (usually the number of variants in this set is much smaller than the number used in step S5). The effect of the set of fine-mapped variants on the target phenotype is thereby estimated using a machine learning algorithm, which is preferably the same as the algorithm used in step S5. The input of this embodiment of step S3 is the marginal effect size for each of the fine-mapped variants, 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 based on the residual effect size that takes into account the uncertainty of the effect size estimate and the probability that the variant is causal of the focal phenotype. These weights calculated for the subset of fine-mapped variants are then subtracted from the effect sizes of the multiple genetic variants, thereby generating residual association data equivalent to the other embodiments of step S4.

[0095] In some embodiments, the set of fine-mapped variants can be combined with a set of literature-reported variants that are likely to be causative of a disease / trait, such that the set of fine-mapped variants further includes one or more variants known to be likely to be causative of the target phenotype.

[0096] [Alternative implementation of step S3 using between-population data] Assumptions can be made about the consistency of effect sizes across populations. On the one hand, effect sizes can be assumed to be constant across populations. On the other hand, if sufficient data are available, effect sizes can be estimated using any of the above methods in matched populations using only population-specific data sets.

[0097] The intermediate process is a hierarchical model that borrows information about effect sizes across populations, allowing for some variation in estimated effect sizes if the data support this.

[0098] [Steps S4 and S5: Subtraction and Machine Learning] In steps S4 and S5, the method includes using the fine mapping predictive 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 correlations between the target phenotype and the multiple genetic variants.

[0099] In an embodiment, the machine learning algorithm includes the model proposed by LDpred, which requires only summary statistical data to identify the residual signal.

[0100] In that illustrative context, we define three effect sizes for each variant. ·β i refers to the marginal effect of variant i, which is the estimated effect size from the summary statistics when using summary statistics. p j refers to the probability that fine-mapped variant j is causal (probabilities in the credible set sum to 1). ·

[0101]

number

[0102]

number

[0103] These notations allow for the effect size β i can be normalized so that their variances are equal before the subtraction.

[0104]

number

[0105] A putative machine learning step can then be performed on these residual effect sizes in a manner identical to how it would be performed in the absence of fine mapping (i.e., steps S2 and S3 are not performed and the machine learning step operates directly on the input data). Adding fine mapping can result in large differences in the output of the machine learning process. These large differences are seen, for example, when the output from the machine learning algorithm is used to calculate the PRS weights (defined below), as shown by the difference between Figure 6 (showing PRS weights derived using a method with fine mapping) and Figure 7 (showing PRS weights derived using a method without fine mapping). Furthermore, the fine mapped signals provide better robustness to population differences, as they approximate the true causal variants that are typically shared across populations.

[0106] The machine learning step S5 can output a set of weights for non-fine-mapped variants (i.e. variants that were included in the input data 2 but were not identified as fine-mapped variants in step S2) that indicate the significance assigned to the variants based on the residual signal while taking into account the correlation between the variants. This process is heavily influenced by population-specific correlation patterns between variants, resulting in population-specific sets of variants and weights. Thus, in an embodiment where the input data is derived from multiple different populations of organisms, the correlation r between the i-th variant and the j-th variant can be calculated by the weights of the i-th variant and the j-th variant. ij depends on the population.

[0107] Figures 6 and 7 show how the machine learning / LDpred weights are widely distributed across the region, in contrast to fine-mapping outputs that accurately characterize putative causal variants, or at least correlate closely with the true causal variants.

[0108] Incorporating relevant trait association data in variant-specific prior distributions Bayesian machine learning algorithms for genetic prediction, such as LDpred, typically rely on priors that capture the probability that a variant is causal. Typically, the same prior is assigned to all variants; this is called a flat prior. Small priors assigned to all variants lead to a sparse model where most weights are small or equal to zero, while larger values ​​lead to a more diffuse model where the predictive weights are spread across a large number of variants. An alternative to the standard LDpred model, which assumes a flat prior for each variant, is to exploit cross-trait information to adapt the priors in a variant-specific manner.

[0109] One way to achieve this is to use a logistic regression model in which a binary outcome variable shows the consistency of the direction of marginal variant effect sizes between adequately powered GWAS of the target phenotype and GWAS of the same target phenotype using independent cohorts of individuals.

[0110] This means that when 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 for each of multiple genetic variants to be causative of the target phenotype that depends on the consistency of the strength of association of each genetic variant with the target phenotype across different genetic studies. The strength of association (e.g. P-value) from the GWAS performed on the relevant traits is used as the input / predictor variable. The resulting linear combination of regression coefficients (where each regression coefficient captures how predictive the associated trait is of the target phenotype) is weighted by the input variables (i.e., fit values) and then subjected to a normalization procedure, which can then serve as the variant-specific prior distribution. As a result, the machine learning algorithm generates higher weights for those variants with evidence of association to the traits most relevant to the target phenotype.

[0111] Another option for the definition of variant-specific weights is the incorporation of genomic annotations derived from external genomic studies that are not GWAS. In such a case, applying the machine learning algorithm to the residual association data includes using a prior probability for each of multiple gene variants that are causative of the target phenotype, which depends on the genomic annotation of multiple gene variants in the region of interest. To further enhance the machine learning algorithm and improve the prediction performance, such functional information, e.g., the presence of protein-coding variants or DNA binding sites for associated transcription factors, can be combined with the prior distributions defined from GWAS data.

[0112] [Steps S6 and S7: Calculation of PRS model and PRS] In an embodiment, the method further comprises calculating the individual's PRS for the target phenotype using the fine mapping predictive model (calculated in step S3) and the further predictive correlations identified by the machine learning algorithm (step S5). In an embodiment, the fine mapping predictive 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 individual's PRS taking into account genetic data 3 from the individual (step S7). In an embodiment, the PRS model is a weighted sum over the variants, with weights provided by the fine mapping predictive model and the further predictive correlations identified by the machine learning algorithm. In an embodiment, the PRS is calculated as follows:

[0113]

number

[0114] For fine-mapped variants, PRS weight α l is usually the probability that the variant is the cause p l The effect size of variant l on the target phenotype, weighted by β l is directly related to, and therefore α l =p l β l It becomes.

[0115] Predisposing traits were used to estimate the relationship between predisposing and target effect sizes (proportionality β l =Kβ' l etc. Here, β' l is the effect size for a predisposition study) has been established, the weight of the PRS is based on this predisposition. α l =p l Kβ' l

[0116] For variants with PRS weights assigned by machine learning algorithms, the relationship between effect size and PRS weight may be less direct and depends on the details of the algorithm.

[0117] In some embodiments, an individual's polygenic risk score may be derived from a combination (e.g., the sum) of a first partial polygenic risk score (e.g., based on fine-mapped variants of the genetic data only) provided by applying a fine mapping predictive model to the genetic data from the individual, and a second partial polygenic risk score (e.g., based on variants of the genetic data other than the fine-mapped variants) provided by applying additional predictive correlations from a machine learning algorithm to the genetic data from the individual.

[0118] The machine learning steps that lead 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.

[0119] In some embodiments, the input data 2 is derived from multiple different populations of organisms (eg, different types of ancestors), and either or both of the following are satisfied: i) Calculation of the fine mapping predictive models is performed separately on portions of the input data corresponding to different populations to obtain a plurality of respective population-concordant fine mapping predictive models. ii) application of the machine learning algorithm to the residual associated data is performed separately on portions of the input data corresponding to different populations to obtain multiple respective sets of further predictive correlations of population agreement.

[0120] The PRS of an individual from one of the populations (e.g., an individual with genes that are primarily from one of the different populations) may be calculated as follows: Input data is received from the individual. The PRS of the individual is calculated by performing either or both of the following: i) A population-matched fine-mapping predictive model is applied to all of the input data from individuals, and the population-matched fine-mapping predictive model is matched to the population of individuals. ii) applying the population-matched further set of predictive correlations to all of the input data from individuals, matching the population-matched further set of predictive correlations to the population of individuals.

[0121] [Alternative implementation for calculating PRS of mixed individuals] For individuals who are a mixture of two or more distinct ancestral groups, such as African-American individuals, different segments of chromosomes can be assigned to each of these ancestries. An important motivation for fine-mapping approaches is to identify causal variants and CSs that are more likely to be consistent across populations. However, beyond fine-mapping, the types of predictive algorithms, including machine learning methods, depend on the patterns of linkage disequilibrium and thus on the target population. Thus, different PRSs are derived for different populations.

[0122] The field of population genetics has established methods to match chromosome segments of individuals to the distinct populations from which they originated. This process is called "chromosome painting." To properly handle admixed individuals, we apply this chromosome painting step to the genotype data of the individuals. Rather than assigning individuals to a single population, we construct an admixture PRS that considers the maternal and paternal chromosome copies separately and applies the relevant population-specific PRS to the appropriate chromosome segments.

[0123] In such embodiments, input data from an individual (with genes derived from a mixture of different populations) is received. The PRS for the individual is calculated by doing one or both of the following: i) matching each of a plurality of population-matched fine mapping predictive models to a corresponding portion of the input data that is matched to the population of the population-matched fine mapping predictive models, and applying each matching fine mapping predictive model to the corresponding portion of the input data. ii) matching each of the multiple sets of population-matched further predictive correlations to a corresponding portion of the input data that matches the population of the set of population-matched further predictive correlations, and applying each set of matching further predictive correlations to the corresponding portion of the input data.

[0124] In practice, fine-mapping predictive models are expected to be largely consistent across populations, with the set of fine-mapped variants, and in some cases even the effect sizes of the fine-mapped variants, being specific, and using cross-population information to correctly capture them. Thus, the method is expected to be most valuable in matching with the population for a set of additional predictive correlations. Thus, in an embodiment, the fine-mapping predictive model is established by combining data from multiple available population data sets for either or both of i) the selection of fine-mapped variants and ii) the effect sizes associated with these variants. In such an embodiment, polygenic risk scores can be derived by applying a shared population-consistent fine-mapping predictive model (i.e., a fine-mapping predictive model that is valid for multiple individuals regardless of the population they belong to) to input data from individuals, and only the additional predictive correlations are established in a population-specific manner. [Explanation of symbols]

[0125] 2. Input Data 3. Genetic data 4. Data representing PRS scores 6 equipment 8 Receiving section 10 Data Processing Section

Claims

1. 1. A computer-implemented method of analyzing genetic data relating to an organism to obtain information about the organism, 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 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, the fine mapping algorithm 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 being causative of the phenotype-variant association, wherein the probabilities of the fine-mapped variants in the set sum to one; calculating a fine mapping predictive model based on the input data and the set of fine mapped variants, the fine mapping predictive model quantifying an effect of the set of fine mapped variants on the target phenotype; using the fine-mapping predictive model to subtract effects of the set of fine-mapped 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 additional predictive correlations between the target phenotype and the plurality of genetic variants.

2. 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 step of receiving input data comprises: receiving individual-level data for each of a plurality of individuals, the individual-level data including a genotype and a corresponding phenotype; and using the individual-level data to determine an estimated effect size for each of the plurality of genetic variants on the target phenotype and a standard error for each of the estimated effect sizes.

4. Identifying the set of fine-mapped variants is performed using an iterative method, each iteration comprising: identifying fine-mapped variants in regions of the genome that are distinct from previously identified fine-mapped variants based on the input data; updating the input data to take into account the effect of the previously identified fine-mapped variants on the target phenotype using a matrix of correlations between the genetic variants within the regions of the genome; and determining whether to perform further iterations based on the updated input data.

5. wherein identifying the set of fine-mapped variants comprises using a plurality of predisposing traits known to influence the target phenotype, said step of using the predisposing traits comprising: determining a set of fine-mapped variants for said predisposing trait; and determining whether to include each of one or more of the fine-mapped variants for the predisposing traits in the set of fine-mapped variants for the target phenotype based on a relationship between the plurality of predisposing traits and the target phenotype.

6. 6. The method of any one of claims 1 to 5, wherein identifying the set of fine-mapped variants comprises identifying a set of fine-mapped variants for one or more causative predisposing traits known to influence the target phenotype.

7. Calculating the fine mapping predictive model comprises: determining an effect size for the one or more predisposing traits of the set of fine-mapped variants for the one or more predisposing traits; and determining an effect size of the target phenotype of each of the fine-mapped variants for the predisposing trait included in the set of fine-mapped variants for the target phenotype based on a predetermined relationship between the effect size of the predisposing trait and the effect size of the target phenotype.

8. 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; 8. The method of any one of claims 1 to 7, wherein the fine-mapped predictive model comprises a fine-mapped effect size on the target phenotype for each of the fine-mapped variants, the fine-mapped effect size being calculated from the estimated effect sizes of the fine-mapped variants taking into account the estimated probability that the fine-mapped variant is responsible for the phenotype-variant association.

9. The method of any one of claims 1 to 8, wherein the effect of the set of fine-mapped variants on the target phenotype is predicted using a machine learning algorithm.

10. 10. The method of claim 9, wherein the set of fine-mapped variants further comprises one or more variants known to be likely causative of the target phenotype.

11. 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; wherein the step of subtracting the effects of the set of fine-mapped variants on the target phenotype from the input data comprises obtaining a residual effect size for each of a plurality of the genetic variants in the input data, wherein the residual association data comprises the residual effect sizes, and after appropriate renormalization of the effect sizes to ensure equal variances, the residual effect size for genetic variant i is [0010] is given by the following formula: [0025] Here, β i is the estimated marginal effect size of gene variant i, N is the number of fine-mapped variants, and p j is the probability that variant j is causative, [0030] is the fine-mapped effect size of the jth fine-mapped variant on the target phenotype; r ij The method of any one of claims 1 to 10, wherein j is the correlation between the jth fine-mapped variant and the ith genetic variant.

12. 12. The method of any one of claims 1 to 11, wherein the input data is derived from a plurality of different genetic studies, and wherein the step of applying a machine learning algorithm to the residual association data comprises using a prior probability for each of the plurality of genetic variants being causative of the target phenotype that depends on the consistency of the strength of association of each genetic variant with the target phenotype across the different genetic studies.

13. 13. The method of any one of claims 1 to 12, wherein applying the machine learning algorithm to the residual association data comprises using a prior probability for each of the plurality of genetic variants being causative of the target phenotype that depends on genomic annotation of the plurality of genetic variants in the region of interest.

14. 14. The method of any one of claims 1 to 13, further comprising calculating an individual's polygenic risk score for the target phenotype by applying the fine mapping predictive model and the further predictive correlations identified by the machine learning algorithm to genetic data from the individual.

15. The polygenic risk score is a weighted sum of: [0045] where L is the number of variants contributing to the PRS, each variant being included in the further predicted correlation from the fine mapping predictive model or the machine learning algorithm, and α l quantifies the strength of association of variant l to the target phenotype, the strength of association being specified by the fine mapping predictive model or by the further predicted correlation from the machine learning algorithm; X l The method of claim 14, wherein is the genotype of mutant type l.

16. 16. The method of any one of claims 14 or 15, 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 predictive model to genetic data from the individual and a second partial polygenic risk score provided by applying the further predictive correlation of the machine learning algorithm to the genetic data from the individual.

17. the input data is derived from a plurality of different populations of organisms; 17. The method of claim 1, wherein either or both of the steps of calculating the fine mapping predictive models are performed separately for portions of the input data corresponding to different populations to obtain a plurality of respective fine mapping predictive models of population matches, and the step of applying the machine learning algorithm to residual associated data is performed separately for portions of the input data corresponding to different populations to obtain a plurality of respective sets of further predictive correlations of population matches.

18. The method further comprising: receiving input data from individuals having genes derived from a mixture of different populations; matching each of a plurality of population-matched fine mapping predictive models to a corresponding portion of the input data that matches the population of the population-matched fine mapping predictive models, and applying each matching fine mapping predictive model to the corresponding portion of the input data; and matching each of a plurality of sets of population-matching further predictive correlations to a corresponding portion of the input data that matches a population of the set of population-matching further predictive correlations and applying each set of matching further predictive correlations to the corresponding portion of the input data; and calculating a polygenic risk score for the individual by performing either or both of the following steps:

19. 20. The method of claim 18, wherein the step of matching each of a plurality of sets of population-consistent further predictive correlates is performed, the step of matching each of a plurality of population-consistent fine-mapping predictive models is not performed, and the step of calculating the polygenic risk score comprises applying a shared population-consistent fine-mapping predictive model to the input data from the individuals.

20. The method further comprising: receiving input data from individuals having genes predominantly derived from one of the different populations; applying a population-matched fine-mapping predictive model to all of the input data from the individual, matching the population-matched fine-mapping predictive model to the population of the individual; and applying a set of population-matched further predictive correlations to all of said input data from said individuals, and matching said set of population-matched further predictive correlations to said population of individuals; and calculating a polygenic risk score for the individual by performing either or both of the following steps:

21. 21. The method of claim 20, wherein the step of applying a set of further predictive correlations of the population match is performed, the step of applying a fine mapping predictive model of the population match is not performed, and the step of calculating the polygenic risk score comprises applying a shared population consistent fine mapping predictive model to the input data from the individuals.

22. 22. The method of any one of claims 1 to 21, wherein the step of 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.

23. The method according to any one of claims 1 to 22, wherein the organism is a human.

24. 1. An apparatus for analyzing genetic data relating to an organism to obtain information about the organism, comprising: a receiver configured to receive 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 in a genome of the organism; a data processing unit, applying a fine-mapping algorithm to all or a subset of said input data to identify one or more independent phenotype-variant associations within said region of interest by identifying, for each association, a set of one or more fine-mapped variants from said plurality of genetic variants, and determining, for each fine-mapped variant, an estimated probability that it is causative of the phenotype-variant association, such that the probabilities of the fine-mapped variants in said set sum to one; calculating a fine-mapping predictive model based on the input data and the set of fine-mapped variants, the fine-mapping predictive model quantifying an effect of the set of fine-mapped variants on the target phenotype; using the fine mapping predictive model to subtract the effects of the set of fine mapped variants on the target phenotype from the input data to obtain residual association data; and The apparatus is configured to 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.

25. A computer program comprising instructions which, when said program is executed by a computer, cause said computer to carry out a method according to any one of claims 1 to 23.

26. A computer readable medium containing instructions which, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 23.