Robust SNP-based prediction of rheumatoid arthritis through machine-learning-optimized polygenic risk score

Over the past few decades, genome wide association studies (GWAS) have revolutionised the field of complex disorder genetics, with the identification of more than 70,000 significant association (p ≤ 5 × 10^–8) of variants with diverse diseases and traits (GWAS catalogue as of April 2022) [1]. While providing deep insight into complex diseases, significant challenges remain before GWAS findings can be clinically applicable. Current GWAS employs statistical approaches to identify variants associated with a phenotype [2], based on population inferred relationship between data and the outcome variable [3]. While statistical models are able to make predictions, predictive accuracy is neither their aim nor their strength [4]. As such, clinical applicability of a disease-associated variant is less clear since statistically significant association identified in one set of data may not necessarily apply in a future dataset [5‐7]. Classical statistics was designed for the analyses of data with moderate number of dependent and independent variables [3]. However, GWAS interrogates hundreds of thousands to millions of SNPs (single nucleotide polymorphisms) for disease association in an often limited number of samples. This analysis is prone to type-I errors and provides imprecise statistical inferences about the complex associations and relationships among the many variables [3]. Two strategies are used in GWAS to reduce the multiple testing and the subsequent type-I error burden. To reduce the number of variables being examined, current GWAS focuses on interrogating mainly tag-SNPs, removing other SNPs that are in strong linkage disequilibrium (LD) [8]. The second strategy to address type-I error is to employ Raw P-value Thresholding (RPVT) [9], where statistical significance of association of each SNP to phenotype is evaluated through a predefined threshold after multiple test correction. The criticisms of GWAS are that it may detect association that is spurious [10, 11], may not identify the causal variant/gene [10] and only accounts for a small fraction of the heritability of complex traits [8, 12]. These problems could be due to the removal of variants from analysis to mitigate the multiple-testing burden, as well as the treatment of variants as individual and independent, without consideration for potential higher order interactions amongst them [13, 14]. To address these criticisms, there are recent attempts to combine variants identified in GWAS to estimate the genetic risk for a trait using polygenic risk score (PRS) [15], which employs a fixed model that sums the contribution of a group of risk alleles for a specific complex disorder [16, 17], either as weighted PRS based on (p-value) and/or effect size (odds ratio) or unweighted PRS [16]. Initial attempts which employed weighted PRS achieved only limited predictive performance [16]. While relatively easy to implement and interpret, traditional PRS is based on independent, linear combination of risk alleles and assumes normal distribution of underlying data. Hence, it may not capture non-linearity or complex interactions amongst the risk alleles [17].

In contrast, machine learning (ML) is a statistics-free approach which instead focuses on the use of algorithms to identify patterns in rich and unwieldy data [3]. While statistical models often require assumptions to be made regarding the distribution of the population or the data, ML requires minimal assumptions and is effective even in the presence of complicated nonlinear interactions [3]. ML is also effective in analysing large, complex datasets with high dimensionality, which is a challenge for traditional statistical modelling methods, as in GWAS. To address the ‘curse of dimensionality’ [18‐20] in ML, feature selection can be implemented to identify a subset of features that contribute most to the prediction of a variable, restricting the overall dimensionality of the dataset to only features (SNPs) that are most relevant to the prediction variable [21]. While computationally intensive, feature selection techniques such as recursive feature selection importantly considers both joint effects of SNPs and their possible interactions, identifying a set of SNPs with the best predictive performance [21, 22]. Combining machine learning (ML) with PRS has the potential to capture non-linear and complex interactions and facilitate better clinical decision-making. Thus, ML can complement existing statistical approaches to bring the prediction of disease or trait outcomes closer to clinical application.

In this study, rheumatoid arthritis (RA) was selected as a model disease to demonstrate the robustness of ML-optimized PRS in disease prediction. Affecting ~ 1% of the population worldwide, RA is one of the more prevalent chronic inflammatory joint diseases with mortality rates up to 54% higher than the general population [23, 24]. It is a complex autoimmune disease primarily characterised by the swelling of the joints leading to joint pain, stiffness and in severe cases irreversible joint damage. This is further exacerbated by several associated comorbidities such as coronary artery diseases and hyperlipidaemia [25]. Notably, as the onset of RA occurs in middle-aged adults at their economic and productivity prime, the effects of the disease poses a major socioeconomic burden on both patient and society [26, 27]. Hence, early identification of at-risk individuals may be critical in minimising the effects of the disease, through the provision of early preventive or mitigation measures or treatment. There is no single diagnostic test for RA and experts rely on patterns of clinical presentation.

Several non-genetic factors, including gender, smoking, pollutants, silica and asbestos have been found to modulate the risk of RA [28, 29]. The high prevalence of RA within families, with strongest risks observed in first-degree relative, suggests that genetics play an important role in RA development [30]. Since the first three RA GWAS were performed in 2007 [31], > 400 unique SNPs had been documented in the GWAS catalog (as of April 2022) [32] to be significantly associated with RA. While polymorphisms within HLA regions accounted for 11–37% of RA heritability and non-HLA risk loci were estimated to account for ~ 5% of heritability [33], > 50% of heritability remain unaccounted for [34]. Several studies have attempted to predict RA using known RA risk alleles (from previous association analyses) with some incorporating lifestyle and clinical characteristics (Additional file 2: Table S1).

Here, as a complement to GWAS, we employ ML, using similar algorithm that we previously reported for the prediction of methotrexate response in RA patients [35, 36], to select predictive SNPs that can robustly predict RA across 3 separate unseen test cohorts. We then developed an ML-optimized PRS to facilitate better clinical decision making. A summary of our overall ML strategy is presented in Fig. 1.

In this study, through rigorous ML feature selection that is tolerant to differences in sample sizes [35, 36], we identified a signature of 9 SNPs that can predict RA with excellent predictive performance not only in the training dataset (mean AUC > 0.99; mean sensitivity > 0.96, mean specificity > 0.95, mean accuracy > 0.96, and mean average precision > 0.96) assessed through cross-validation, but also in not one, but 3 independent, unseen test datasets (AUC > 0.97; sensitivity > 0.91, specificity > 0.95, F1 score > 0.90, accuracy > 0.94, and average precision > 0.93 in all 3 test datasets). This excellent predictive performance is unlikely due to random chance since the predictive performance of 1,000 9 randomly selected SNPs was poor with AUC < 0.7, and majority of the 9 randomly selected SNPs have AUC of only between 0.50 and 0.51.

To facilitate interpretability and clinical applicability of the 9 ML-identified predictive SNPs for individual patients, a PRS was developed based on these 9 ML-identified predictive SNPs. Notably, not only are the 9 ML-identified predictive SNPs significantly (P < 1 × 10^–5) associated with RA individually, the calculated PRS score (from these 9 SNPs) were also found through logistic regression to be significantly (P < 1 × 10^–6) associated with RA across all 3 unseen datasets and this single ML-optimized PRS was also found to have comparable excellent predictive performance as the 9 SNPs across all 5 selected ML models and 3 independent test sets. To further facilitate the potential clinical application, an RA ML-optimized PRS calculator based on the 9 ML-identified predictive SNPs was developed; it is accessible via this link: https://​xistance.​shinyapps.​io/​prs-ra/​. In this RA ML-optimized PRS calculator, the genotype of the 9 ML-identified predictive SNPs in a patient are entered and the PRS, odds as well as probability of the patient developing RA will be given, enabling the healthcare provider to make an earlier diagnosis when the patient presents with only minimal clinical signs.

Although exome sequencing data was interrogated, only a minority (3/9) of the predictive SNPs reside in the coding region. Two (rs1266832853 and rs60465633) of the 3 coding predictive SNPs are benign, non-synonymous susceptibility SNPs with positive effect sizes while one (rs143773270) is a potential deleterious, non-synonymous protective SNP with negative effect size. The majority of the predictive SNPs resides in the introns (6/9). Most (5/6) of these intronic predictive SNPs are potentially protective SNPs with negative effect sizes. These potentially protective predictive SNPs in non-coding regions with negative effect sizes, are potential eQTL SNPs or are in strong LD (R² > 0.8) with non-interrogated SNPs that are predicted to be potentially functional, modulating expression of the gene by being eQTL SNPs, or potentially altering transcription factor binding sites (TFBS) or intronic splicing regulatory elements (ISRE). The putative functionality of the sole intronic predictive susceptibility SNP (rs200901373) with positive effect size remains unknown. These data thus suggest that while the majority (2/3) of the non-synonymous, predictive coding SNPs have positive effect size and thus may confer susceptibility to RA through modulating protein structure/function, the majority of the (5/6) of intronic predictive SNPs have negative effect size and thus may confer protection against RA through modulating the expression of the intricate network of genes as these intronic predictive SNPs are either eQTL SNPs or have potential to alter TFBS/ISRE sites. Validating the potential function and effect of these SNPs in the RA pathway would be a worthwhile future direction.

None of the 9 predictive SNPs have previously been reported to be associated with RA. To gain further insight into these 9 predictive SNPs, numerous GWAS databases were interrogated to evaluate if any of these 9 predictive SNPs were previously reported via GWAS to be associated with any disease/phenotype/function which may help explain the role of these SNPs/genes in RA. As GWAS mainly interrogate tag-SNPs and this study interrogates exomic SNPs, only 3 of these 9 predictive SNPs were reported by 3 GWAS databases (BioBank Japan PheWeb [53], IEU Open GWAS Project [54], GWAS Atlas [55]) to be significantly associated (P < 1 × 10^–5) (Additional file 2: Table S4) with several different functions (including eQTL association) and some diseases, which can be categorized into 6 different themes (Additional file 2: Table S2). Notably, most of these association were consistent with the characteristics or phenotype of RA. For example, rs11385557 in the intronic region of BAIAP2L1 was found in OpenGWAS database to be significantly (P < 7.69 × 10^–5) associated with peripheral nerve disorders (Additional file 2: Table S4, #8) which is consistent with RA patients often experiencing peripheral neuropathy with pain, numbness, and muscle weakness [56]. Similarly, rs11385557 was also reported by OpenGWAS database to be significantly (P < 9.08 × 10^–5) associated with lymphocyte and monocyte counts (Additional file 2: Table S4, #1–4). This is consistent with reports of lymphopenia (low lymphocyte counts) commonly observed in RA patients [57], as well as monocytes activation and migration into joints in early RA [58]. Hence, it may be worthwhile to further investigate their roles in RA.

Although these 9 SNPs were not previously associated with RA, 2 of them reside in disease susceptibility genes (Additional file 2: Table S2, Row 5 and 9). rs11385557 is found in the intron of the BAIAP2L1, an insulin receptor tyrosine kinase gene (Additional file 2: Table S2). The expression of the BAIAP2L1 gene is positively correlated with C-reactive protein (CRP) levels within fibroblast-like synovial cells from RA patients [61]. CRP is an immune regulator that is commonly used as a marker for systemic inflammation in RA [62]. Although the intronic SNP was not predicted to be potentially functional, it was found to be in strong linkage disequilibrium (r² > 0.8) with 19 potentially functional SNPs, most of which are associated with modulation of gene expression (eQTL SNPs), while some are predicted to alter intronic splice regulatory elements (ISRE) (Additional file 2: Table S3). Hence, the above observation is consistent with SNPs in LD with rs11385557 modulating gene expression of the BAIAP2L1 gene, which in turn alters the CRP levels in RA patients. rs112667995 resides within the intron of the PRKN (Parkin RBR E3 Ubiquitin Protein Ligase) gene. PRKN deficiency ameliorates inflammatory arthritis through the suppression of p53 degradation [63]. Similarly, although this intronic SNP (rs79555231) is not predicted to be potentially functional (Additional file 2: Table S2), it is in strong LD (r² > 0.8) with 21 potentially functional SNPs that are mainly predicted to alter transcription factor binding sites (TFBS) (Additional file 2: Table S3). Thus, SNPs in LD with rs79555231 may influence the expression of PRKN which in turn modulates inflammatory arthritis. Taken together, both these intronic SNPs are in LD with markers that modulate gene expression of either BAIAP2L1 or PRKN that is associated with RA. Since majority of the predictive SNPs are in non-coding regions and many of the potentially functional SNPs likely reside beyond the regions that were sequenced through exome sequencing, it may thus be worthwhile to build models from WGS data.

While these 9 SNPs displayed excellent predictive performance for RA in Singaporean Chinese population, future work could explore the generalizability of the predictive performance of these 9 SNPs in other populations. It may be worthwhile to initially determine whether these SNPs exhibit population differentiation between Singaporean Chinese and another population [64] before general adoption. Further studies could also focus on the characterization of the roles of the predictive coding SNPs in conferring predisposition to RA as well as the roles of the predictive intronic SNPs in conferring protection against RA.

These 9 SNPs have potential to be clinically applicable for diagnosing individual patients with RA through the development of rapid genotyping assays for these SNPs. With improvements of WGS technology coupled with its rapidly declining cost, it will not be surprising that most, if not all, individuals will have their genome sequenced in the foreseeable future. Then, it will be cost-effective to deploy PRS on a large scale. However, due to resource constraints, current WGS data is primarily stored as variant call format (VCF) files [65] with small data storage space requirements. One limitation of storing WGS data as VCF is that, in order to generate VCF files, raw sequences from a group of individuals are aligned and variants are then identified based on a reference genome. The sequence identity of loci which do not show variability between the group of individuals and the reference genome are not stored. As such, depending on the size of the group of individuals, the number of variations stored will be different, with more variation stored in VCF files of large and more diverse group and less variation stored in VCF files of smaller and more homogenous group of individuals. For locus without sequence identity assigned, it is not possible to accurately extract the genotype information at the individual level. One cannot assume that locus to be the homozygous genotype of the reference genome as the unassigned region could also be excluded due to low mapping quality or poor read coverage during sequencing. Hence, for WGS data to be clinically applicable at an individual level, an alternative VCF file, the genomic VCF (gVCF) file format which can be generated using the GATK suite [66], as done in this study, could be explored. Although the storage size required for gVCF files is larger than typical VCF files, it is still overall much smaller compared to the storage of BAM files containing the sequencing read alignments [67]. The advantage of the gVCF file is that it stores not only information of the genotypes of variant site, but it also compactly stores information of the invariant genomic regions, facilitating the more accurate assignment of genotype at invariant sites for clinical implementation of our prediction models.

In summary, PRS of the 9 ML-selected predictive SNPs is significantly associated (P < 1 × 10^–6) with and predictive (AUC > 0.9) of RA in all 3 independent, unseen test datasets. To facilitate individualized clinical applicability, RA ML-PRS calculator of these 9 SNPs (https://​xistance.​shinyapps.​io/​prs-ra/​) was developed. Majority of the predictive SNPs are protective and reside in non-coding regions and are either predicted potentially functional SNPs (pfSNPs) or in high linkage disequilibrium (r2 > 0.8) with un-interrogated pfSNPs. These data highlight the promise of this ML strategy to identify useful genetic features that can robustly predict disease with good potential for clinical application.