Identifying mutation-driven changes in gene functionality that lead to venous thromboembolism

Abstract

Venous thromboembolism (VTE) is a common hematological disorder. VTE affects millions of people around the world each year and can be fatal. Earlier studies have revealed the possible VTE genetic risk factors in Europeans. The 2018 Critical Assessment of Genome Interpretation (CAGI) challenge had asked participants to distinguish between 66 VTE and 37 non-VTE African American (AA) individuals based on their exome sequencing data. We used variants from AA VTE association studies and VTE genes from DisGeNET database to evaluate VTE risk via four different approaches; two of these methods were most successful at the task. Our best performing method represented each exome as a vector of predicted functional effect scores of variants within the known genes. These exome vectors were then clustered with k-means. This approach achieved 70.8% precision and 69.7% recall in identifying VTE patients. Our second-best ranked method had collapsed the variant effect scores into gene-level function changes, using the same vector clustering approach for patient/control identification. These results show predictability of VTE risk in AA population and highlight the importance of variant-driven gene functional changes in judging disease status. Of course, more in-depth understanding of AA VTE pathogenicity is still needed for more precise predictions.

Background

Venous thromboembolism (VTE) is a disorder that causes formation blood clots, primarily affecting veins deep in the body. VTE that affects legs, groins, or arms, specifically, is designated deep vein thrombosis (DVT). Clots traveling to the lungs and blocking arteries result in pulmonary embolisms (PE) (Bartholomew, 2017). VTE affects 300,000 – 600,000 individuals in the U.S. each year (Beckman, Hooper, Critchley, & Ortel, 2010). In an estimated 100,000 – 180,000 of these, the disease is fatal (Goldhaber, 2012). VTE incidence is much higher in older people (≥ 80 years of age) compared with that in the younger ones. European and African descents are reported to have the highest VTE incidence in multiple studies (Beckman et al., 2010; Heit, Spencer, & White, 2016; Zakai et al., 2014).

The pathogenesis of VTE is not fully understood. Multiple factors contributes to development of VTE, including genetics, but also factors such as advanced age, pregnancy, obesity, trauma, surgery, hospitalization, etc. (Beckman et al., 2010). Previous studies have identified several potential genetic causes, including factor V Leiden variation (Ridker et al., 1995) and prothrombin mutations (Rosendaal et al., 1998). Despite the recent approval of new direct-acting anticoagulants (Tellor et al., 2018), warfarin is commonly used for treatment and prevention of VTE (Sterne et al., 2017), due to its low renal function impairment, good drug adherence record, and accessible cost (Burn & Pirmohamed, 2018). Warfarin is also used for clot prevention in other diseases, e.g. Atrial Fibrillation (AF) (Shariff, Aleem, Singh, Y, & S, 2012), where an abnormal heart rhythm affects blood flow and predisposes to clot formation (Zimetbaum, 2017).

CAGI (Critical Assessment of Genome Interpretation, http://genomeinterpretation.org) clotting disease challenge (https://genomeinterpretation.org/content/clotting-disease-dvt-or-pe-exomes) provided a dataset of exomes of VTE (unprovoked VTE, i.e. not caused by external factors such as trauma, surgery, etc.) and non-VTE (mostly AF) patients. All 103 patients in this cohort are African-American (AA) and all were on warfarin for treatment or prevention of thrombosis.

Some VTE risk prediction methods using genetic biomarkers have been published (Ahmad et al., 2018; Folsom et al., 2016), focusing primarily on populations of European descent. However, none of these methods specifically address VTE in AA patients. We implemented two different methods for the prediction of VTE risk in this challenge. First, we applied the commonly used genetic risk score (GRS) (Cooke Bailey & Igo, 2016), imputing missing variants and also making use of the warfarin dose information. Second, we clustered exomes using different features, e.g. functional deficiency of the known disease variant/genes (Pinero et al., 2017). Note that warfarin dosage is a confounding factor for identifying VTE signal. For example, patients with VTE often have underlying genetic defects, which may cause hypercoagulation and, thus, require a higher warfarin dose compared to patients with AF, who take warfarin as a prophylactic measure (James, Britt, Raskino, & Thompson, 1992); i.e. it is likely simpler to predict VTE status when the warfarin dose is known. All predictions submitted in this challenge were evaluated by the assessors according to the data provider labels.

Among all methods that relied solely on genetic information (i.e. not including warfarin doses), our three clustering-based methods were ranked 1^st, 3^rd, and 4^th. Our best performing method represented each exome in the cohort as a vector of predicted effect scores of variants within the known VTE genes (extracted from the DisGeNET database). These exome vectors were then clustered into two clusters with k-means. Our second-best method (ranked 3^rd overall) had collapsed the variant effect scores into gene-level function changes, using the same vector clustering approach for patient/control identification. Our last clustering approach (ranked 4^th) represented each exome as a vector of genotypes of all variants within the known VTE genes, without any functional annotation. The genetic risk score method, based on known GWAS variants, performed worse than the clustering methods even with warfarin dosage information included but, as expected, had very high precision, albeit for a smaller number of patients identified.

Transforming our current knowledge of AA VTE variants into computationally predicted variant effects, we achieved the highest (62%) overall accuracy of prediction. However, integrating other elements into our method, including, for example, patient clinical features (Zhai et al., 2019), is likely to improve performance. Since VTE is preventable (Beckman et al., 2010), it is critical to have an accurate prediction of VTE risk for clinical use. Notably, unprovoked VTE is also a sign of other diseases, e.g. some forms of cancer (van Es et al., 2017). An early diagnosis of VTE could thus potentially lower the disease prevalence and help elevate the patients’ quality of life.

Materials and Methods

Challenge Data.

The study cohort contained 103 African-American (AA) individuals who were taking warfarin for either VTE (unprovoked VTE, 66 individuals) or non-VTE (most are AF patients, 37 individuals) treatment and/or prevention of blood clots. Specifically, CAGI participants were provided with the whole exome sequencing variant call files (VCFs) and the clinical covariates files for each individual. The clinical covariates included gender, age, height, weight, whether the individual was also taking aspirin and/or amiodarone, and his/her warfarin dose. There were 58 individuals taking a high dose (> 49 mg/week) of warfarin and 45 individuals taking a low dose (< 35 mg/week). This dataset and its detailed clinical covariate statistics were also reported in Daneshjou et al. (Daneshjou et al., 2014).

Data cleaning.

We retained only the PASS variants according to VQSR (Variant Quality Score Recalibration) standard (McKenna et al., 2010). Principle Component Analysis (PCA) of SNPs indicated three different clusters of the subjects (Supp. Figure S1A). Other analysis of the quality metrics, including individual number of variants, number of heterozygous variants, number of singletons, individual call rate, individual Ti-Tv ratio, and individual PASS (VQSR standard) rate, indicated that the separation of subjects might be due to the difference in quality of the variant calls that may result from differences in sequencing batches or other systematic errors. We removed all variants in the VCF files that failed the GTAK VQSR PASS qualification. The cleaning resulted in higher quality of data (Ts-Tv ratio rising from 2.20 to 2.48) and loss of obvious clustering of subjects (Supp. Figure S1B). While for the purposes of this challenge we did not apply further cleaning, the cleaned data still had low quality calls. Thus, we suggest that a more comprehensive clean-up could benefit all further analyses.

Genetic risk scoring (GRS) methodology (Method 1).

A variant in Protein S (PROS1), Valine in position 510 to Methionine (V510M, rs138925964) was shown to be associated of VTE (Daneshjou et al., 2016). To the best of our knowledge, there are two AA VTE genome-wide association studies (GWAS) (Heit et al., 2017; Hernandez et al., 2016). Hernandez, et al. have found three SNPs on chromosome 20 which increased the risk of VTE by 2.3-fold. These SNPs are in the eQTLs (expression quantitative trait loci) for the THBD gene; here, the VTE patients had lower THBD expression than the healthy controls. Heit, et al. found three intragenic SNPs of genome-wide significance in the LEMD3, LY86, LOC100130298 genes, respectively. Note that there were VTE GWAS done in the European population, but the identified SNPs are generally not observed in the AA population (Hotoleanu, 2017).

The genetic risk scoring (GRS) using SNPs identified in the Europeans does not work in African-Americans (Folsom et al., 2016). Thus, to apply the GRS strategy, we had to develop a strategy using AA relevant SNPs. We used variants from the most recent GWAS study from Heit et al. and one study from data provider (Daneshjou et al., 2016) to construct our GRS. None of the three significant GWAS variants (Heit et al., 2017) were covered by the challenge exome data. Therefore, we imputed the variants using IMPUTE2 (Howie, Donnelly, & Marchini, 2009) with reference to the 1000 Genomes Phase 3 data (NCBI build b37) (Genomes Project et al., 2015). Unfortunately, imputation accuracies (R², i.e. info value in IMPUTE2) were relatively low: 0.336, 0.206, 0.288 for the GWAS-significant variants rs138916004 (LEMD3), rs3804476 (LY86), rs142143628 (LOC100130298), respectively. We calculated three versions of scores using the GRS method (Method 1.1, Method 1.2, and Method 1.3) as described below:

$$r_i={\sum }_{j=1}^m{w}_j{x}_{ij}$$ 1

Method 1.1 (not submitted to CAGI):

Only the three significant GWAS loci from Heit et al. (imputed for our data) were included in the GRS equation (Eqn. 1), where w_j was the j^th SNP log odds ratio from the GWAS study (for a total of m SNPs) and the total risk of the i^th individual (r_i) was the sum of the weighted genotypes (x_ij) in his/her exome.

Method 1.2 (not submitted to CAGI):

We included (i) the three significant Heit et al. GWAS loci (imputed), (ii) loci reported in Heit el al. that were below GWAS significance but were covered by the challenge data, (iii) and also variants in the PROS1 gene reported in Daneshjou et al. that were covered by the challenge data, into Eqn. 1. Note here that w_j log odds ratios for (i and ii) came from the Heit et al. study, while the PROS1 variants had ratios assigned by the Daneshjou et al. study. While combining multiple studies into a single score is not ideal, we felt that additional strongly-associated variants could contribute to the resolution of the method.

Method 1.3 (submitted):

Based on the warfarin dose we adjusted the predictions of Method 1.2, such that individuals with high and low warfarin dosage (heuristically) scored 1.5-fold and 0.8-fold of the Method 1.2 predictions, respectively.

Note that since imputed variants had low imputation quality, the probabilistic values of the imputed genotypes were used in the equation instead of the hard call values (Li, Willer, Sanna, & Abecasis, 2009).

Clustering methodology (Methods 2–4).

From the DisGeNET (Pinero et al., 2017) database we extracted a full list of VTE genes. DisGeNET contains standardizes annotations of gene-disease relationships extracted from various sources. Each relationship is assigned a Gene-Disease Association Score (GDA Score) according to the level of source evidence, i.e. the higher the GDA Score, the more reliable the gene-disease relationship is. We searched (April 12^th, 2018) for keywords “Venous thromboembolisms” (“VTE”, C1861172; 111 results), “Deep vein thrombosis” (“DVT”, C0149871; 96 results), “Pulmonary embolism” (“PE”, C0034065; 71 results), obtaining three gene lists and retaining only the genes with a GDA Score ≥ 0.2 (8, 8, and 21 genes from each keyword search, respectively).

We further categorized the genes into Level 1, Level 2, and Level 3 genes (Supp. Table S1; 8, 3, 14 genes in Level 1, 2, 3 gene list, respectively), where Level 1 genes were in the VTE gene list, Level 2 genes were in both DVT and PE lists, but not in VTE list, and Level 3 genes were unique to the DVT or PE list.

Further, VCF variants were annotated with ANNOVAR (K. Wang, Li, & Hakonarson, 2010) to retain those that affected the genes of interest. We applied three different clustering strategies to cluster exomes in our cohort: (i) each individual was represented as a vector of genotypes (0/0, 0/1, or 1/1) of all variants within Level 1 and 2 genes (125 variants within 11 genes, Method 2) and then clustered using k-modes (Huang, 1997) clustering; (ii) k-means (Hartigan & Wong, 1979) clustering by SNAP (Bromberg & Rost, 2007) predicted functional effects of all non-neutral variants within all three gene lists (67 non-neutral variants in total, Method 3); (iii) k-means clustering by gene function deficiency scores (27 genes total) with the 67 non-neutral variants in (ii) (Method 4). The gene function deficiency score for each gene was calculated via Eqn. 2, where score_i was the SNAP prediction, normalized to range 0 to 1, for the i^th variant (for a total of N variants) within the gene. We heuristically set a factor of 0.35 (het_i) for heterozygous genotype to account for the fact that heterozygous variants are generally less functionally effective than homozygous variants. The gene score approximated the amount of gene function left after the mutation(s).

$$genescore={\prod }_{i=1}^N\left(1-he{t}_i\times scor{e}_i\right)$$ 2

Note that k-modes was used in (i) for clustering nominal features (genotypes) and k-means was used in (ii) and (iii) for clustering numerical features (function deficiency scores for variant or gene). For all three clustering methods, we chose the bigger cluster as the VTE group and smaller one as non-VTE group for the submissions, as we knew that there were more VTE than non-VTE individuals in the dataset (Daneshjou et al., 2014).

Prediction evaluation.

Method 1 produced numeric predictions (ranging from 0 to 1) while Methods 2–4 results were binary (0 is non-VTE and 1 is VTE). We used the ROC (receiver operating characteristic) curve AUC (area under the curve) to evaluate the prediction performance of Method 1. For all methods, we calculated the overall accuracy (Eqn. 3), precision and recall (Eqn. 4), and Matthews correlation coefficient (MCC) (Eqn. 5), where TP (true positives) were the correctly classified VTE patients, FP (false positives) were the incorrectly classified non-VTE controls, TN (true negatives) were the correctly classified non-VTEs, and FN (false negatives) were the incorrectly classified VTEs.

$$Accuracy=\frac{TP+TN}{TP+FP+TN+FN}$$ 3

$$Precision=\frac{TP}{TP+FP}Recall=\frac{TP}{TP+FN}$$ 4

$$MCC=\frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}$$ 5

Results

In post-CAGI evaluation, the performance of the 3-locus GRS method (Method 1.1) was worse than random guessing (ROC/PR AUC = 0.432/0.610). The performance of Method 1.2 (not submitted to CAGI) was also poor (ROC/PR AUC = 0.546/0.688). This is not unexpected as the score included low significance variants of low imputation confidence and also heuristically combined variant log-odds scores across different studies. The submitted GRS+warfarin (Method 1.3) predictions were significantly better (Figure 1, Table 1, ROC/PR AUC = 0.646/0.788). However, since Method 1.3 used warfarin dosage information in prediction, it was not in the final ranking of evaluated submissions.

Figure 1. GRS (Method 1.3) predictions separate VTE and non-VTE individuals.

GRS (genetic risk scores) were normalized to a 0 to 1 range (Y axis). The samples are scattered within a status along the X-axis for better visibility. Note that non-VTE individuals, on average, score lower than VTE ones.

Table 1.

Method performance summary

Method#	Method	Known variants/genes used	Function effect	Accuracy†	Precision†	Recall†	MCC†
1.1	GRS	Heit et al. (3 loci)	No	37.9%	100.0%	3.0%	0.105
1.2	GRS	Heit et al. and PROS1 variants	No	47.6%	68.8%	33.3%	0.065
1.3	GRS+warfarin dose	Heit et al. and PROS1 variants	No	50.5%	89.5%	25.8%	0.252
2	Kmodes clustering	Level 1, 2 genes	No	52.4%	62.3%	65.2%	0.052
3	Kmeans clustering	Level 1, 2, 3 genes	variant-level	62.1%	70.8%	69.7%	0.182
4	Kmeans clustering	Level 1, 2, 3 genes	protein-level	58.3%	65.8%	72.7%	0.054

^†Overall accuracy (Eqn. 3), precision and recall (Eqn. 4), and MCC (Eqn. 5). Default cutoff of 0.5 was used for calling an exome VTE or non-VTE. The best performance among four methods is indicated in bold.

For Method 2, we considered all possible variants within the known Level 1 and 2 disease genes as features (Materials and Methods). By choosing the larger cluster to be VTE, 69 and 34 individuals were predicted to be VTEs and non-VTEs, respectively. There was also an obvious separation between the two clusters by Multiple Correspondence Analysis (Husson, Lê, & Pagès, 2017) (Supp. Figure S2).

For Method 3, we chose only the functional significant (as per SNAP predictions variants as clustering features. Only the 67 non-neutral variants from SNAP predictions were kept for k-means clustering. Here, 65 and 38 individuals were predicted to be VTEs and non-VTEs, respectively. Since this CAGI challenge assessors chose the overall accuracy (Eqn. 3) as the primary metric for ranking methods, we report that this method (accuracy = 62%) was ranked the first among all methods (ours and others’) that did not use warfarin dosage.

For Method 4, we converted the functional annotation from the variant-level to the protein-level, taking the product of all variant function scores within one gene as the protein function deficiency score (Eqn. 2). By k-means clustering of the 27 gene function deficiency scores, 73 and 30 individuals were predicted to be VTEs and non-VTEs, respectively. This method was our second-best and ranked 3^rd overall (accuracy = 58%).

The GRS method (Method 1.1) had the highest precision (100%) among all predictions, but it recognized only very few VTE patients (recall=3%; 2 of 66 patients). On the other hand, Method 3 had more balanced precision and recall values, identifying 70% of the VTE patients with moderately high precision (71%). Method 4 identified more VTE patients (73%) but less accurately (precision of 66%). Note that the two methods that included variant functional annotations had higher accuracies than the GRS methods and Method 2, the clustering method without functional annotation.

Discussion

Similarity between VTE and AF complicates prediction.

In the challenge, non-VTE samples came mostly from the AF (Atrial Fibrillation patients) population. The reason for this choice is clear – it is uncommon to find a study where individuals would be taking a drug for no reason. Warfarin in VTE is mainly used for treatment of blood clots in veins (for DVT) or in the lungs (for PE), while in non-VTE cases it is often prophylactic. However, separating two types of patients is arguably a more difficult task than separating patients from healthy controls.

VTE and AF are both blood-related polygenic conditions and are not fully understood genetically (Bapat, Anderson, Ellinor, & Lubitz, 2018; Hotoleanu, 2017). They share genetic risk factors and patho-physiological bases for clot formation (Shariff et al., 2012). In fact, DisGeNET (Pinero et al., 2017) suggests 39 overlapping disease risk genes between the two diseases, e.g. coagulation factor F5 (Tang et al., 2013; Zateyshchikov, Brovkin, Chistiakov, & Nosikov, 2010), CEPT (cholesteryl ester transfer protein) (Asselbergs et al., 2006; Deguchi, Banerjee, Elias, & Griffin, 2016), and TFPI (tissue factor pathway inhibitor) (Efthymiou et al., 2018; Xie et al., 2017) are involved in thromboembolic risk of both VTE and AF. Additionally note that VTE often results from the activation of the coagulation system (Shariff et al., 2012), where tissue factors (TF) play an important role in initiating the clot formation. Similarly TF are over-expressed in AF patients with thromboembolism (Watson, Shantsila, & Lip, 2009). This observation may explain why an anticoagulative drug warfarin is effective in preventing blood clots in both VTE and AF, as well as why studies found that the two diseases often co-occur and that the presence of one increases the risk for another (Enga et al., 2015; Lutsey et al., 2018; Sundboll et al., 2017).

Thus, while AF is indeed different from VTE, including a separate cohort of healthy controls may help better explore the disease pathogenesis of either VTE or AF in future studies.

VTE genetics vary between European and African-American populations.

VTE-associated genetic risk loci were reported in several studies. The first VTE GWAS (Tregouet et al., 2009) done in the European population. Later, more VTE GWA studies were done (Germain et al., 2015; Germain et al., 2011; Heit et al., 2012; Hinds et al., 2016; Tang et al., 2013), but no African-American VTE GWAS was published until 2016 (Hernandez et al., 2016); another AA GWAS closely followed (Heit et al., 2017).

In all European population GWAS, variants in genes F2, F5, F11, FGG, FGA, ABO, ZFPM2, LCN1P2, NME7, etc. were found to be significantly VTE-associated. However, the allele frequencies (AFs) of the risk alleles in the European population are different in people of African ancestry (AFR in 1000 Genomes). For example, the T allele in variant rs6025 (gene F5) is significantly associated with VTE in European and other ancestry cases (no AA included, log-odds ratio of up to 3.57 (Heit et al., 2012)), but this allele is not observed at all in the AFR population. On the other hand, the two AA-specific GWAS (Heit et al., 2017; Hernandez et al., 2016) have found significantly VTE-associated SNPs in LEMD3, LY86, LOC100130298 or CLVS1, LOC102723446, CD93 genes (Table 2). These variants were not discovered in any other GWA studies in European populations.

Table 2.

Significantly VTE-associated variants in AA population

Variant	Mapped gene	Risk allele	AF† in AFR	AF† in EUR
rs138916004	LEMD3	G	2%	0%
rs3804476	LY86	G	8%	44%
rs142143628	LOC100130298, CLVS1	T	1%	0%
rs73692310	LOC102723446	T	7%	0%
rs1998081	CD93	T	22%	8%
rs2144940	N.A.	C	26%	8%

^†AFs (allele frequencies) of the risk allele from 1000 Genomes Project Phase 3.

Thus, the differences in genetic underpinnings of this disease in different populations require population-specific model building for further insight. That is, we suspect that our predictors built for this challenge will not work as well for Europeans as for African-Americans. However, we can likely use the same approaches with European-specific data.

Warfarin dose predicts VTE risk.

One consideration for this challenge was whether to take the warfarin dose and/or warfarin dose-related genes into consideration during prediction. At the time of the challenge, all 103 individuals were taking warfarin for treatment or prevention of thrombosis. For the VTE status prediction in the current CAGI challenge, warfarin dose alone (without any genetic data) performed the best. That is, high-dose individuals were more likely to be VTE and low-dose warfarin individuals were more likely to be non-VTE. Note that as the purpose of the CAGI challenge was to interpret the genetic data, the assessor had excluded from assessment all methods that used the provided warfarin dose information.

This result, however, highlights two questions: (1) Does warfarin dose vary from disease to disease? and (2) Do the warfarin dose-related genes affect VTE risk? In response to the first question, James et al. have found that AF patients required lower and VTE patients require higher doses of warfarin (James et al., 1992). This trend remained significant even after adjusting for age and other factors. Thus, if used in hind-sight, once the warfarin dosage is established it is arguably easier to identify the patient’s VTE status. This approach, however, has limited, if any, clinical utility.

As for the second question, aside from clinical factors such as age, weight, sex, etc., two major categories of genes affect warfarin dose: the pharmacodynamic genes (e.g. VKORC1, EPHX1, GGCX, CALU), which warfarin works on to block the vitamin K dependent clotting pathway, and the pharmacokinetic genes (CYP3A4 for R-warfarin, CYP2C9 for S-warfarin, and other Cytochrome P450 enzymes), which metabolize warfarin (Whirl-Carrillo et al., 2012). One variant in VKORC1 (1173 C > T) was shown to be VTE-associated (Lacut et al., 2007). Another in VKORC1 (−1639 G > A) was shown to be DVT-associated (Vesa, Trifa, Crisan, & Buzoianu, 2016). To make a generalizable conclusion about relationship between warfarin dose-relevant genes and VTE risk across populations, further specifically designed experiments are needed.

Use of prediction method in clinical practice needs more work.

Although the GRS result (Method 1.1, Supp. Figure S3) had the lowest overall accuracy, it is likely more useful in the clinical diagnostic settings, where the number of healthy individuals is much larger than the number of VTE-affected people. Here, precision in “diagnosing” someone with VTE is of utmost importance as it guides further treatments and interventions. While identifying only two patients with VTE (3% of all patients), GRS VTE diagnosis precision was 100%--that is no healthy people were misdiagnosed. Our clustering-based Method 3, on the other hand, identified significantly more VTE patients (46 people; 70%), but had a precision (71%) likely useless in the clinic. Thus, our methods are not yet ready for prime-time in real world applications.

Note further that clustering-based approaches require two pieces of information to be practical: the presumed number of clusters (i.e. phenotypes to split people by) and knowledge of which cluster represents which phenotypic group. Of course, this would not be a problem if a classification (e.g. clustering) method was developed using an unrelated “training set”, for which both of these pieces of information were available. However, this approach would then be further complicated by the need to ensure biological, and methodological similarity between training and testing samples (Y. Wang et al., 2017).

Variant/gene functional changes are important for prediction.

Our two function-annotated methods (Methods 3 and 4) performed better than the one without function annotation (Method 2). In terms of predicting disease risk, using variant/gene function has two important merits over simply using variant genotypes: (1) decreasing the number of features and (2) giving biologically meaningful weight to individual features.

High-dimensionality of feature space and small numbers of samples consistently plague high-throughput experimentation. For complex diseases, such as VTE, hundreds or even thousands of genetic loci are within the known disease genes, but we generally have much fewer individuals to analyze. For example, in this challenge, we had only 103 individuals evaluate as compared to the 561 variants in the known disease-related genes. Keeping only the 67 variants that had a predicted effect on protein function significantly changed the prediction space. Moreover, as most clustering algorithms assume roughly equal importance for all features (Dash & Liu, 2000), the variants of low or no likely contribution to disease should be ignored as noise. These observations could explain the reduced Method 2 performance where all variants were treated equally.

When comparing the performance of top two methods (Methods 3 and 4), Method 4 had a slightly lower performance. One way to explain this is that the heuristic gene function deficiency score used here possibly did not represent function change accurately; e.g. due to the semi-probabilistic combination of the multiple variants per gene. Nevertheless, the fact that Method 4 still worked better than Method 2 suggests that weighing genes in a biologically meaningful fashion is a more reasonable approach than simply using genotypes. Current work in the lab is promising a more meaningful gene score formula in the near future.

Future prospects.

Since African-American and European populations have very different genetic architecture (Park, Cheng, & Haiman, 2018) precise and validated biomarkers of VTE useful in the AA population need more targeted GWAS. Equipped with this new information, GRS calculations can be modified to be more accurate. However, it is worth repeating that the GWAS significant variants are often only biomarkers of the disease instead of disease-causing mutations. Methods that consider variants in the marked regions from a functional level could contribute to our understanding of disease and, potentially, outperform allele-count based approaches. Moreover, as the next-generation sequencing techniques develop and drop in price, we will be able to access more variants, including the rare and individual ones, within an individual and across larger cohorts. Information contained in these new data sets cannot, by definition, be assessed with strictly statistically driven methods. On the other hand, new and specifically targeted tools, could help reveal previously unseen disease-causative variants.