PRS-Net: Interpretable polygenic risk scores via geometric learning

PRS-Net: Interpretable polygenic risk scores via geometric learning1 Han Li1, Jianyang Zeng2,∗, Michael P . Snyder3,∗, and Sai Zhang4,5,6,∗2 1 Institute for Interdisciplinary Information Sciences, Tsinghua University, Beijing, China3 2 School of Engineering, Westlake University, Hangzhou, Zhejiang, China4 3 Department of Genetics, Stanford University School of Medicine, Stanford, CA, USA5 4 Department of Epidemiology, University of Florida, Gainesville, FL, USA6 5 J. Crayton Pruitt Family Department of Biomedical Engineering, University of Florida, Gainesville, FL, USA7 6 The Genetics Institute, University of Florida, Gainesville, FL, USA8 ∗ Correspondence: [email protected]; [email protected]; [email protected] Abstract. Polygenic risk score (PRS) serves as a valuable tool for predicting the10 genetic risk of complex human diseases for individuals, playing a pivotal role in ad-11 vancing precision medicine. Traditional PRS methods, predominantly following a linear12 structure, often fall short in capturing the intricate relationships between genotype and13 phenotype. We present PRS-Net, an interpretable deep learning-based framework de-14 signed to effectively model the nonlinearity of biological systems for enhanced disease15 prediction and biological discovery. PRS-Net begins by deconvoluting the genome-16 wide PRS at the single-gene resolution, and then it encapsulates gene-gene interac-17 tions for genetic risk prediction leveraging a graph neural network, thereby enabling18 the characterization of biological nonlinearity underlying complex diseases. An atten-19 tive readout module is specifically introduced into the framework to facilitate model in-20 terpretation and biological discovery. Through extensive tests across multiple complex21 diseases, PRS-Net consistently outperforms baseline PRS methods, showcasing its22 superior performance on disease prediction. Moreover, the interpretability of PRS-Net23 has been demonstrated by the identification of genes and gene-gene interactions that24 significantly influence the risk of Alzheimer’s disease and multiple sclerosis. In sum-25 mary, PRS-Net provides a potent tool for parallel genetic risk prediction and biological26 discovery for complex diseases.27 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 1 1 Introduction28 Complex human diseases display polygenicity in their genetic architectures, characterized by a29 multitude of common genetic variants with minor individual effects accumulatively influencing the30 disease risk1–4. The polygenic risk scores (PRSs) are developed to quantitatively characterize the31 genetic susceptibility of individuals to specific traits or complex diseases based on the common32 genetic variants 5–7. This methodology empowers the early deployment of targeted therapeutic33 interventions and facilitates the practice of personalized medicine8–10.34 PRS is typically calculated using the summary statistics derived from genome-wide association35 studies (GWAS)11–17, a widely-used statistical method for identifying disease-associated genetic36 variants18–20. While GWAS can identify disease risk genetic variants, such as single nucleotide37 polymorphisms (SNPs), that exhibit significant differences in frequencies between cases and con-38 trols, these variants tend to have modest individual effects on the phenotype, resulting in limited39 prediction capability. In an effort to enhance predictive modeling, various statistical methods have40 been applied to aggregate the effects of individual SNPs. The widely adopted method for calculat-41 ing PRS, exemplified by PLINK21 and PRSice12, is known as clumping and thresholding (C+T)11,42 which involves summing allele counts weighted by effect sizes estimated from GWAS. More recent43 approaches like LDpred216 utilize Bayesian modeling to infer the posterior mean effect size of each44 marker by incorporating prior information on effect sizes and linkage disequilibrium (LD) data from45 an external reference panel. Similarly, lassosum217 estimates PRS using summary statistics and46 a reference panel within a penalized regression framework. With the notable increase in dataset47 sample sizes for GWAS, these methods have achieved enhanced predictive power 22. Nonethe-48 less, these techniques primarily rely on univariate effect sizes derived from linear GWAS models,49 thus often overlook potential non-linear associations between genetic factors and phenotypes,50 which can undermine their predictive performance.51 Efforts have also been made to construct models capable of capturing non-linear interac-52 tions in PRS calculation. These include tree-based methods like random forests 23,24 , gradient53 boosting25,26 , and AdaBoost 27,28 , as well as deep learning-based techniques such as multiple-54 layer perceptrons (MLP) 29 and convolutional neural networks 30. However, these methods only55 take a limited number of variants as their input, and lack the integration of versatile prior biological56 knowledge. Indeed, these approaches have demonstrated either comparable or, in many cases,57 less effective performance in predicting phenotypes when compared to linear models31,32 .58 In this study, we propose PRS-Net, a geometric deep learning-based approach designed to59 effectively model the intricate non-linear relationships among genetic factors such as genes in60 predicting the disease risk, thus delivering more accurate and robust PRSs. Based on the sum-61 mary statistics of GWAS, PRS-Net first maps PRS onto a gene-gene interaction (GGI) network62 through the derivation of gene-level PRSs using the C+T method. Subsequently, a graph neural63 network is employed to iteratively update the embedding of the genes via performing message64 passing on the GGI network, thus capturing the complex GGIs from the network. An attentive65 readout module is then introduced to provide interpretable PRS predictions. PRS-Net also inte-66 grates a mixture-of-expert module33 designed to enhance the accuracy of PRS predictions when67 dealing with multi-ancestry datasets. Our comprehensive evaluation encompasses six complex68 diseases extracted from the UK Biobank database 34, including Alzheimer’s disease, atrial fib-69 rillation, rheumatoid arthritis, multiple sclerosis, ulcerative colitis, and asthma. The results con-70 sistently demonstrated the superiority of PRS-Net over baseline methods, including PLINK 21,71 PRSice214, LDpred-2 16, and lassosum2 17 in PRS prediction. Notably, through case studies fo-72 cused on Alzheimer’s disease and multiple sclerosis, we illustrated that PRS-Net provided biolog-73 ically meaningful interpretability by identifying specific genes and GGIs that significantly influence74 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 2 H. Li et al. disease risk. In summary, PRS-Net stands as a potent and innovative tool for precise PRS pre-75 diction, addressing the limitations of current linear models and offering a more comprehensive76 approach to unraveling the genetic underpinnings of complex traits and diseases.77 2 Method78 GWAS Population Summarystatistics-Variant-Riskallele-P-value-Effectsize DiseaseaGWASanalysisbGene-geneinteractionnetwork Graphneuralnetwork Attentivereadoutmodule cPRS-NetDiseaseprediction Disease-relatedgene/GGIidentification dApplications Genomesequence Protein-proteininteractionnetwork C+T GeneAGeneBGeneC Messagepassingxklayers SNPA1SNPA2…SNPAnSNPB1SNPB2…SNPBnSNPC1SNPC2…SNPCn Geneembeddings Phenotype Attentionscores Predictor Global-levelembedding … … -log10(P-value)10 0 5 ControlCase 0 8 4 135791113151719212 6 Mixture-of-expertAncestryqueryPhenotypequeryAttentivereadout CasesControls GeneAGeneBGeneCGeneDGeneEGeneFGeneGGeneHGeneIGeneJ GeneAGeneC GeneD GeneF GeneI Fig. 1: An illustrative diagram of PRS-Net.a The proposed framework is based on summary statis- tics, including variants, risk alleles, P-values, and effect sizes derived from GWAS.b A gene-gene interaction network is constructed based on the protein-protein interaction network. Gene-level PRSs are calculated with the C+T method to serve as the node features for the nodes within the network. c A graph neural network is employed to update node features via message passing and subsequently an attentive readout module is applied to provide interpretable PRS predictions. d The PRS-Net can be applied for disease prediction and disease-related gene/GGI identification. In this section, we present our proposed framework for PRS estimation (Fig. 1), covering the79 establishment of the GGI network, the derivation of gene-level PRS, and the architecture of PRS-80 Net.81 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 3 2.1 GGI network82 It is widely recognized that the disease phenotype is not solely determined by individual genes83 but rather involves the intricate interactions among multiple genes, which can exhibit additive or84 non-additive genetic relationships35–37. Additive genetic interactions manifest when the cumulative85 effects of genes jointly contribute to a specific phenotype. Furthermore, there are increasing stud-86 ies highlighting the significance of non-additive genetic interactions38–40. Epistasis is a prominent87 example of non-additive genetic interaction, which occurs when the impact of a gene mutation88 depends on the presence or absence of mutations in one or more other genes 41–43. We estab-89 lish a GGI network that empowers PRS-Net to capture the intricate genetic relationships that are90 potentially associated with the target phenotypes (Fig. 1b).91 We construct our GGI network based on the protein-protein interactions derived from the92 STRING database44, as protein-protein interactions represent potent indicators of functional rela-93 tionships between genes. Formally, we construct a GGI network, denoted as G = (V , E), where94 V stands for the set of nodes and E stands for the set of edges. Each node vi ∈ V stands for a95 coding gene and each edge (vi, vj) ∈ E stands for an interaction between nodes vi and vj derived96 from the STRING database 44. Note that, we add a self-loop (vi, vi) for each node vi ∈ V. This97 network construction results in a GGI network encompassing 19,836 coding genes and 250,23698 interactions.99 Upon deriving the GGI network, we proceed to compute gene-level PRSs for the genes within100 the network using a C+T approach 11,21 . More precisely, for each gene in the network, we focus101 on the SNPs falling within a designated range, spanning from its transcription start site - L to102 its transcription end site + L. In our tests, we set L to 10 kilobases (KB), thereby encompassing103 the SNPs situated in non-coding regions, such as the promoters of the genes. Subsequently, for104 each gene, we perform LD clumping on the associated SNPs from the GWAS data, utilizing the105 LD information estimated in the target data. Following this, we filter the SNPs based on a specific106 P-value threshold. The gene-level PRSs are then derived by multiplying the genotype matrix by107 the effect sizes obtained from the GWAS data, and then dividing this by the number of allele108 observations for each gene. For the LD clumping process, we set the LD threshold R2 to 0.5 and109 the physical distance threshold to 250 KB. As for the thresholding step, we set the P-values to110 1e−5, 1e−4, 1e−3, 1e−2, 5e−2, 0.1, 0.2, 0.3, 0.5, and 1, respectively. This process results in the111 computation of eleven PRSs for each gene, which serves as their initial features. We denote the112 initial feature of vi ∈ V as hi ∈ H , where H ∈ R|V|×11 and |V| stands for the number of genes in113 G.114 2.2 PRS-Net115 Graph neural network116 We harness the power of a graph neural network to capture the complex interactions among genes117 within our established GGI network (Fig. 1c). In our tests, we specifically opt for a graph isomor-118 phism network (GIN) 45 due to its proven theoretical and experimental expressiveness. Formally,119 we first encode the initial feature of nodes, denoted as H, by employing an MLP in the following120 manner:121 H 0 = MLP 0(H), (1) where H 0 ∈ R|V|×D and D is the dimension of hidden states. Subsequently, we apply multiple GIN122 layers to iteratively update the representation of each node by aggregating the representations of123 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 4 H. Li et al. its neighbors, as depicted below:124 hk i = MLP k((1 + ϵk) · hk−1 i + X vj ∈N (vi) hk−1 j ), (2) where hk−1 i is the hidden states of vi at the (k − 1)-th layer, N (vi) stands for the neighbors of vi in125 the GGI network, hk i stands for the updated hidden states of vi at the k-th layer, MLPk is the MLP126 at the k-th layer, and ϵ stands for a learnable variable. Following k iterations of aggregation, each127 gene effectively encapsulates the interaction information within its k-hop neighborhood.128 Attentive readout module129 To make predictions for each data sample, we derive the global-level representation for each130 sample through an attentive readout module, illustrated as follows:131 hG = Attentive readout(Q, K, V ), hG = A · V , A = Sigmoid(Q · K), K = H k · WK, V = H k · WV , (3) where WK ∈ RD×D and WV ∈ RD×D stand for trainable projection matrices to derive the key132 (i.e., K) and value (i.e., V ) matrices, respectively. Q ∈ R1×D stands for a trainable query vector.133 Sigmoid stands for the sigmoid function. A ∈ R1×|V| stands for the attention scores, with elevated134 scores signifying a greater significance of the associated genes. hG ∈ R1×D stands for the global-135 level representation.136 After deriving the global-level representation, we employ an MLP to derive the final prediction,137 denoted as ˆPRS, as follows:138 ˆPRS = MLP( hG). (4) Additionally, we implement a mixture-of-expert module 33 to effectively handle datasets that139 encompass data samples from multiple ancestries. More specifically, we introduce a specialized140 attentive readout module for each distinct ancestry. These dedicated attentive readout modules141 are activated when processing data from individuals with specific ancestral origins. To illustrate,142 when dealing with input samples of Western European ancestry, we derive the ancestry-specific143 global-level representation as follows:144 hEUR G = Attentive readout(Q EUR, KEUR, V EUR). (5) The ancestry-specific readout module is designed to capture the unique knowledge pertaining to145 each ancestry in relation to the disease. In addition, we introduce another shared readout module146 to capture disease-related knowledge that holds general applicability across all ancestries:147 hPH G = Attentive readout(Q PH, KPH, V PH). (6) Then, we derive the final global-level representation by combining the aforementioned two repre-148 sentations:149 hG = hEUR G + hPH G . (7) The process for deriving global-level representations of individuals from other ancestries follows a150 similar approach. The final PRS prediction can be computed with equation 4, utilizing the derived151 global-level representation. We refer to the single-ancestry variation as PRS-Net and the multiple-152 ancestry variation as PRS-NetMA.153 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 5 3 Results154 3.1 PRS-Net outperforms baseline methods in PRS prediction155 Alzheimer’sdiseaseAtrial fibrillationUlcerative colitis AsthmaRheumatoid arthritisMultiple sclerosis Fig. 2: The PRS prediction performance of PRS-Net compared to baseline methods across a range of complex diseases, including Alzheimer’s disease, atrial fibrillation, ulcerative colitis, asthma, rheumatoid arthritis, and multiple sclerosis, measured in terms of the area under the receiver operating characteristic curve (AUROC). The bars are the estimated standard errors. We extracted genotype-phenotype data from the UK Biobank database34 for six different com-156 plex diseases, which encompassed Alzheimer’s disease, atrial fibrillation, rheumatoid arthritis,157 multiple sclerosis, ulcerative colitis, and asthma. ICD-10 codes46 were employed to define the dis-158 ease phenotypes (Supplementary Table 1). For our primary experiments, we focused exclusively159 on individuals of Western European ancestry due to the insufficient size of the non-European160 ancestry population, which did not provide an adequate amount of training data (Supplementary161 Table 2). Following a quality control process, each disease dataset consisted of roughly 411,000 in-162 dividuals (Supplementary Note 1.1). To prevent data leakage, we ensured that none of the GWAS163 were conducted on samples from the UK Biobank database (see Data availability). For each dis-164 ease dataset, we randomly partitioned it into training, validation, and test sets with a ratio of 8:1:1.165 To evaluate the performance of PRS-Net, we compared it against several previously proposed166 methods, such as C+T -based methods (PLINK21 and PRSice214), lassosum217, and three vari-167 ations of LDpred2 16 (LDpred2-auto, LDpred2-grid, and LDpred2-inf), utilizing the area under the168 receiver operating characteristic curve (AUROC) as the metric. To ensure a rigorous and equi-169 table comparison, we utilized LD matrices estimated from European populations within the 1000170 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 6 H. Li et al. Genomes Project47 across all methods in our study. Our results were based on three indepen-171 dent runs with different random seeds to ensure robustness and reliability. The results revealed172 that PRS-Net consistently outperformed all baseline methods on all disease datasets, resulting in173 relative improvements ranging from 0.5% to 3.7%. Interestingly, the largest improvements were174 obtained for two autoimmune diseases, i.e., ulcerative colitis (with a relative improvement of 3.0%)175 and multiple sclerosis (with a relative improvement of 3.7%), reinforcing the observed nonadditiv-176 ity of genomic factors underlying these diseases 38,48–50 . Altogether, our data demonstrates that177 PRS-Net possesses the capacity to capture more intricate associations between genotypes and178 phenotypes that are beyond the reach of previously proposed linear models.179 We utilized the Aalen-Johansen estimator 51 to estimate the disease occurrence over a life-180 time for individuals categorized into high-risk and low-risk groups, as determined by the PRSs181 estimated by PRS-Net and baseline methods. High-risk individuals were defined as those with182 the highest 5% of PRSs, while low-risk individuals were identified as those with the lowest 5%183 of PRSs. The cumulative incidence plots revealed that individuals classified as high-risk by PRS-184 Net generally exhibited a heightened risk of disease throughout their lifetime compared to base-185 line methods, especially for ulcerative colitis, asthma, rheumatoid arthritis, and multiple sclerosis186 (Fig. 3a). Conversely, those categorized as low-risk by PRS-Net tended to maintain a lower risk of187 all diseases over their lifetime in comparison to baseline methods (Supplementary Fig. 1). These188 findings underscore the potential of PRS-Net as a powerful tool for individual risk stratification.189 Next, we assessed the performance of PRS-Net and our multiple-ancestry model, PRS-NetMA,190 on a dataset comprising individuals from diverse ancestral backgrounds. Specifically, we curated a191 mixed-ancestry dataset encompassing Western European, South Asian, and African for asthma,192 which provides a reasonable number of asthma cases (over 1,000) for each ancestry (Supplemen-193 tary Table 2). The results revealed that PRS-Net outperformed baseline methods on the mixed194 ancestry and South Asian ancestry test sets, indicating that the PRS-Net trained solely on the195 Western European ancestry dataset captured the underlying disease biology independent of dif-196 ferent ancestries (Fig. 3b). Additionally, PRS-Net MA demonstrated superior performance when197 compared to PRS-Net on the mixed ancestry, Western European ancestry, and African ancestry198 test sets (Fig. 3b). These findings underscored the ability of PRS-Net MA to leverage the multi-199 ancestry dataset effectively, enhancing its portability in estimating PRS for individuals from diverse200 ancestral backgrounds.201 3.2 PRS-Net identifies disease-related genes and GGIs for Alzheimer’s disease and202 multiple sclerosis203 Following the demonstration of the superior performance of PRS-Net in predicting PRS, we sought204 to explore its capability to identify risk genes and GGIs underlying complex diseases. Alzheimer’s205 disease, a progressively degenerative condition, has been the subject of extensive research for206 many years, leading to the identification of numerous genes associated with the disease52–56. We207 employed PRS-Net to identify disease-related genes and GGIs, with the expectation that our find-208 ings would align with prior research outcomes. Specifically, we first applied the Mann–Whitney U209 test57 to each gene within our constructed GGI network, assessing whether the attention scores210 associated with the gene for individuals with Alzheimer’s disease were notably higher than those211 of the control group. This analysis yielded a gene set comprising 309 genes with compelling sta-212 tistical significance (P-value <0.001). Please refer to Supplementary Data 1 for the complete list213 of the genes. Subsequently, we conducted gene set enrichment analyses (GSEA) 58 utilizing the214 gene ontology (GO) 59 and Kyoto Encyclopedia of Genes and Genomes (KEGG) 60 datasets on215 the identified gene set. Notably, the GO terms related to lipoprotein particles emerged as sig-216 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 7 Alzheimer’sdiseaseAtrial fibrillationUlcerative colitis AsthmaRheumatoid arthritisMultiple sclerosis a b Asthma PRS-NetMAPRS-NetPLINKPRSice-2lassosum2LDpred2-autoLDpred2-inf LDpred2-gridPRS-NetMAPRS-NetPLINKPRSice-2lassosum2LDpred2-autoLDpred2-inf LDpred2-gridPRS-NetMAPRS-NetPLINKPRSice-2lassosum2LDpred2-autoLDpred2-inf LDpred2-gridPRS-NetMAPRS-NetPLINKPRSice-2lassosum2LDpred2-autoLDpred2-inf LDpred2-grid PRS-NetPLINKLDpred2-autoReferenceMixedEURSASAFR Fig. 3: a The cumulative incidence plots of high-risk individuals (with the highest 5% PRSs) identi- fied by PRS-Net and baseline methods. Each plot illustrates the estimated percentage of individ- uals diagnosed with a specific disease at different ages. We provide cumulative incidence plots for the original datasets as a reference.b The PRS prediction performance of PRS-Net compared to baseline methods on an asthma dataset encompassing multiple ancestries, including Western European (EUR), South Asian (SAS), and African (AFR) ancestry, measured in terms of the area under the receiver operating characteristic curve (AUROC). The results on the mixed ancestry test set are also reported. The bars are the estimated standard errors. nificantly enriched within the gene set (Supplementary Fig. 2a). This observation is in line with217 prior studies that have implicated lipoprotein particles as significantly potential risk factors for218 Alzheimer’s diseasee61–63 and have highlighted the role of metabolic dysregulation in the pro-219 gression of Alzheimer’s disease64,65 . Notably, the exploration of high-density lipoprotein-inspired220 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 8 H. Li et al. c dHLA-E HLA-DPB1 HLA-DPA1 HLA-DQB1HLA-DOA HLA-F LCK HLA-DRB1 CD247 HLA-DMB HLA-DQA1 HLA-DRAPTPRC CD48GRB7 EPHB2 IL2RA a b BIN1DNM1 APOC4-APOC2 APOC2 APOEAPOC4 APOC1CLU SORL1 CFB ************* ***** * * * Fig. 4: PRS-Net identifies disease-related genes and GGIs for Alzheimer’s disease and multiple sclerosis. a Top 20 genes with the highest statistical significance in the Mann-Whitney U test for Alzheimer’s disease. The Mann–Whitney U test was utilized to assess whether the attention scores for a particular gene among the cases were significantly higher than those observed in the control group. An asterisk preceding the gene name signifies that the gene has been reported to be as- sociated with Alzheimer’s disease in previous studies.b Examples of interactions within the gene set with statistical significance (P-value <0.001) from the Mann-Whitney U test for Alzheimer’s disease. c Top 20 genes with the highest statistical significance in the Mann-Whitney U test for multiple sclerosis. d Examples of interactions within the gene set with statistical significance (P- value <0.001) from the Mann-Whitney U test for multiple sclerosis. treatments for Alzheimer’s disease has been a well-documented area of study 62,63 . Fig. 4a il-221 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 9 lustrates the top 20 genes with the utmost statistical significance in the Mann-Whitney U test.222 Remarkably, 15 out of these 20 genes have been identified as potential risk factors for Alzheimer’s223 disease in previous studies. One notable example is APOE, which is the most prevalent high-224 density lipoprotein in the central nervous system and has been consistently linked to Alzheimer’s225 disease in numerous studies 66–71. Fig. 4b illustrates examples of the interactions from the GGI226 network between genes within the identified gene set. Please refer to Supplementary Data 2 for227 the complete list of the GGIs. Interestingly, aside from APOE, other genes within the APOE gene228 cluster, including APOC1, APOC2, and APOC4, were also identified as disease-related genes.229 This finding aligns with previous studies that have shown interdependent or independent associ-230 ations of genes within the APOE gene cluster with Alzheimer’s disease72–76. For instance, it has231 been shown that the variant APOE and APOC2 exhibit interactive effects on metabolic pathways,232 potentially contributing to the risk of Alzheimer’s disease 72. APOC1 also has been reported to233 serve as a risk factor for Alzheimer’s disease in combination withAPOE74. Furthermore, the com-234 bined effect of APOE and CLU on Alzheimer’s disease has been observed77. SORL1 is an APOE235 receptor gene, which has been recognized as a genetic risk factor in Alzheimer’s disease. Recent236 research has elucidated the mechanistic connection between these two significant genetic factors237 in Alzheimer’s disease78. A neuron-specific interaction between Alzheimer’s disease risk factors238 SORL1, APOE, and CLU have also been shown in a recent study79. These observations highlight239 the proficiency of PRS-Net in not only identifying disease-related genes but also uncovering gene240 clusters that exhibit interactions contributing to the risk of Alzheimer’s disease.241 We also utilized PRS-Net to uncover genes and GGIs associated with multiple sclerosis. The242 Mann-Whitney U test identified a gene set with 456 potential risk genes (P-value <0.001). Please243 refer to Supplementary Data 3 for the complete list of the genes. The GSEA 58 using the KEGG60244 dataset on this gene set highlighted numerous immune-related pathways of statistical significance,245 such as antigen processing and presentation, allograft rejection, and graft-versus-host disease246 (Supplementary Fig. 4b). This finding aligns with the well-established understanding of multiple247 sclerosis as an autoimmune inflammatory disorder. The GSEA using the GO 59 dataset, unveiled248 significant enrichment of GO terms related to the major histocompatibility complex (MHC) protein249 complex within the identified gene set (Supplementary Fig. 4a), which can be supported by pre-250 vious studies that underscore the substantial genetic impact of MHCs on multiple sclerosis 80–84.251 HLA-DRA, a subunit of HLA-DR which is a human MHC, was identified as the most significant252 gene in our analysis (Fig. 4c). Moreover, substantial HLA genes were identified as risk genes in253 our analysis (Fig. 4d). Please refer to Supplementary Data 4 for the complete list of the GGIs.254 This finding is in line with a previous study indicating that HLA interactions modulate genetic risk255 for multiple sclerosis 85. Additionally, non-additive interactions between HLAs have been widely256 reported to significantly affect the risk of autoimmune diseases 38,48–50 . These discoveries col-257 lectively provide compelling evidence of the potential of PRS-Net to offer valuable insights that258 advance our understanding of diseases.259 3.3 Ablation studies260 To assess the effectiveness of specific design choices in PRS-Net, we conducted comprehensive261 ablation studies. We introduced various modified frameworks derived from PRS-Net, each with262 distinct constraints: PRS-Net-GGI (omitting the GGI network), PRS-Net-Att+Sum (replacing the263 attentive readout module with a sum readout module, which summarized the node feature to de-264 rive the global-level representations), PRS-Net-Att+Mean (replacing the attentive readout module265 with a mean readout module, which computes the average of node features to derive global-level266 representations), and PRS-Net-Att+Max (replacing the attentive readout module with a max read-267 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 10 H. Li et al. b ADMSUC a PRS-NetPRS-Net-GGIPRS-Net-Att+MaxPRS-Net-Att+SumPRS-Net-Att+Mean Fig. 5: The results of ablation studies on PRS-Net. a The comparison results of PRS-Net and its variations, including PRS-Net-GGI (omitting the GGI network), PRS-Net-Att+Sum (replacing the attentive readout module with a sum readout module, which summarized the node feature to de- rive the global-level representations), PRS-Net-Att+Mean (replacing the attentive readout module with a mean readout module, which computes the average of node features to derive global- level representations), and PRS-Net-Att+Max (replacing the attentive readout module with a max readout module, which extracts maximum values across node features to derive the global-level representations), conducted on the datasets of Alzheimer’s disease (AD), multiple sclerosis (MS), and ulcerative colitis (UC). The bars are the estimated standard errors.b The PRS prediction per- formance of PRS-Net versus the extension lengths upstream and downstream of the transcription start and end sites. out module, which extracts maximum values across node features to derive the global-level rep-268 resentations). We compared the performance of PRS-Net against these variants using datasets269 related to Alzheimer’s disease, multiple sclerosis, and ulcerative colitis. The results showcased270 that PRS-Net surpassed PRS-Net-GGI by an average relative improvement of 11.6%, underscor-271 ing the significance of incorporating the GGI network to capture the intricate genetic interactions272 associated with diseases (Fig. 5a). Furthermore, PRS-Net outperformed PRS-Net-Att+Sum, PRS-273 Net-Att+Mean, and PRS-Net-Att+Max with average relative improvements of 33.0%, 2.2%, and274 8.4%, respectively, highlighting the effectiveness of the attentive readout module in summarizing275 node features (Fig. 5a).276 Additionally, we explored the impact of varying extension lengths both upstream and down-277 stream of the transcription start and end sites when calculating gene-level PRSs. We assessed278 different length values, including 0, 5, 10, 20, and 50 KB, and subsequently evaluated their predic-279 tion performance. The results demonstrated that PRS-Net is generally robust to different extension280 lengths (Fig. 5b). However, it is noteworthy that the performance of PRS-Net on the multiple scle-281 rosis dataset significantly declined when the extension length was set to 0 KB (Fig. 5b). This282 observation suggested that including SNPs from non-coding regions can indeed enhance the ac-283 curacy of PRS prediction.284 Discussion285 In this study, we develop PRS-Net, a deep-learning framework that offers interpretable and im-286 proved PRS predictions. By constructing a GGI network and incorporating a graph neural net-287 work, PRS-Net fully takes advantage of the power of non-linear associations between genetic288 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 11 factors and phenotypes. Additionally, the integration of an attentive readout module empowers289 PRS-Net to deliver interpretable predictions. Through comprehensive testing across six complex290 diseases, PRS-Net consistently achieved superior performance in comparison with baseline meth-291 ods in PRS prediction. Furthermore, we demonstrated the interpretability of PRS-Net by using it292 to identify specific genes and GGIs that significantly impact the risk of Alzheimer’s disease and293 multiple sclerosis. In summary, PRS-Net provides a potent tool for accurate PRS prediction and294 biological discovery for complex diseases.295 Data availability296 The GWAS data for Alzheimer’s disease can be accessed at https://ctg.cncr.nl/software/summarys297 tatistics/. The GWAS data for atrial fibrillation can be accessed at https://cvd.hugeamp.org/download298 s.html#summary/. The GWAS data for ulcerative colitis can be accessed at ftp://ftp.sanger.ac.uk/pub299 /project/humgen/summary statistics/human/2016-11-07/. The GWAS data for asthma can be ac-300 cessed at https://www.globalbiobankmeta.org/resources/. The GWAS data for rheumatoid arthritis301 can be accessed at https://data.cyverse.org/dav-anon/iplant/home/kazuyoshiishigaki/ragwas/ra g302 was-10-28-2021.tar/. The GWAS data for multiple sclerosis can be accessed at https://imsgc.net/?303 page id=31/. The UKBB dataset is available at https://www.ukbiobank.ac.uk.304 Code availability305 The source code of PRS-Net can be downloaded from the Github repository at https://github.com/li306 han97/PRS-Net.307 Acknowledgments308 This work was supported in part by the National Natural Science Foundation of China (T2125007309 to J.Z.), the National Key Research and Development Program of China (2021YFF1201300 to310 J.Z.), the New Cornerstone Science Foundation through the XPLORER PRIZE (J.Z.), the Re-311 search Center for Industries of the Future (RCIF) at Westlake University (J.Z.), and the Westlake312 Education Foundation (J.Z.).313 Author contributions statement314 H.L. and S.Z. conceived the concept and designed the study. H.L. and S.Z. developed the method-315 ology and conducted data analysis. H.L., J.Z., M.S. and S.Z. are responsible for the data interpre-316 tation. S.Z., M.S. and J.Z. supervised the project. H.L. and S.Z. prepared the manuscript with the317 assistance from all other authors.318 Competing interests statement319 All authors declare no competing interests.320 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 12 H. Li et al. References321 1. Robert Clarke, John F Peden, Jemma C Hopewell, Theodosios Kyriakou, Anuj Goel, Simon C Heath, Sarah Parish,322 Simona Barlera, Maria Grazia Franzosi, Stephan Rust, et al. Genetic variants associated with lp (a) lipoprotein323 level and coronary disease. New England Journal of Medicine, 361(26):2518–2528, 2009.324 2. CKDGen Consortium, KidneyGen Consortium, EchoGen consortium, CHARGE-HF consortium, Thor Aspelund,325 Melissa Garcia, Y en-Pei C Chang, Jeffrey R O’Connell, Nanette I Steinle, et al. Genetic variants in novel pathways326 influence blood pressure and cardiovascular disease risk. Nature, 478(7367):103–109, 2011.327 3. Chris Gaiteri, Sara Mostafavi, Christopher J Honey, Philip L De Jager, and David A Bennett. Genetic variants in328 alzheimer disease—molecular and brain network approaches. Nature Reviews Neurology, 12(7):413–427, 2016.329 4. Valeria Orr `u, Maristella Steri, Gabriella Sole, Carlo Sidore, Francesca Virdis, Mariano Dei, Sandra Lai, Magdalena330 Zoledziewska, Fabio Busonero, Antonella Mulas, et al. Genetic variants regulating immune cell levels in health and331 disease. Cell, 155(1):242–256, 2013.332 5. Ali Torkamani, Nathan E Wineinger, and Eric J Topol. The personal and clinical utility of polygenic risk scores.333 Nature Reviews Genetics, 19(9):581–590, 2018.334 6. Cathryn M Lewis and Evangelos Vassos. Polygenic risk scores: from research tools to clinical instruments.335 Genome medicine, 12(1):1–11, 2020.336 7. Shing Wan Choi, Timothy Shin-Heng Mak, and Paul F O’Reilly. Tutorial: a guide to performing polygenic risk score337 analyses. Nature protocols, 15(9):2759–2772, 2020.338 8. Greg Gibson. On the utilization of polygenic risk scores for therapeutic targeting. PLoS genetics, 15(4):e1008060,339 2019.340 9. Takahiro Konuma and Yukinori Okada. Statistical genetics and polygenic risk score for precision medicine.341 Inflammation and regeneration, 41(1):1–5, 2021.342 10. Responsible use of polygenic risk scores in the clinic: potential benefits, risks and gaps. Nature Medicine,343 27(11):1876–1884, 2021.344 11. Bjarni J Vilhj ´almsson, Jian Y ang, Hilary K Finucane, Alexander Gusev, Sara Lindstr ¨om, Stephan Ripke, Giulio345 Genovese, Po-Ru Loh, Gaurav Bhatia, Ron Do, et al. Modeling linkage disequilibrium increases accuracy of346 polygenic risk scores. The american journal of human genetics, 97(4):576–592, 2015.347 12. Jack Euesden, Cathryn M Lewis, and Paul F O’reilly. Prsice: polygenic risk score software. Bioinformatics,348 31(9):1466–1468, 2015.349 13. Timothy Shin Heng Mak, Robert Milan Porsch, Shing Wan Choi, Xueya Zhou, and Pak Chung Sham. Polygenic350 scores via penalized regression on summary statistics. Genetic epidemiology, 41(6):469–480, 2017.351 14. Shing Wan Choi and Paul F O’Reilly. Prsice-2: Polygenic risk score software for biobank-scale data. Gigascience,352 8(7):giz082, 2019.353 15. Luke R Lloyd-Jones, Jian Zeng, Julia Sidorenko, Lo ¨ıc Y engo, Gerhard Moser, Kathryn E Kemper, Huanwei Wang,354 Zhili Zheng, Reedik Magi, T ˜onu Esko, et al. Improved polygenic prediction by bayesian multiple regression on355 summary statistics. Nature communications, 10(1):5086, 2019.356 16. Florian Priv ´e, Julyan Arbel, and Bjarni J Vilhj ´almsson. Ldpred2: better, faster, stronger. Bioinformatics, 36(22-357 23):5424–5431, 2020.358 17. Florian Priv ´e, Julyan Arbel, Hugues Aschard, and Bjarni J Vilhj ´almsson. Identifying and correcting for misspecifi-359 cations in gwas summary statistics and polygenic scores. Human Genetics and Genomics Advances, 3(4), 2022.360 18. William YS Wang, Bryan J Barratt, David G Clayton, and John A Todd. Genome-wide association studies: theoret-361 ical and practical concerns. Nature Reviews Genetics, 6(2):109–118, 2005.362 19. Arthur Korte and Ashley Farlow. The advantages and limitations of trait analysis with gwas: a review. Plant363 methods, 9(1):1–9, 2013.364 20. Emil Uffelmann, Qin Qin Huang, Nchangwi Syntia Munung, Jantina De Vries, Yukinori Okada, Alicia R Martin,365 Hilary C Martin, Tuuli Lappalainen, and Danielle Posthuma. Genome-wide association studies. Nature Reviews366 Methods Primers, 1(1):59, 2021.367 21. Shaun Purcell, Benjamin Neale, Kathe Todd-Brown, Lori Thomas, Manuel AR Ferreira, David Bender, Julian Maller,368 Pamela Sklar, Paul IW De Bakker, Mark J Daly, et al. Plink: a tool set for whole-genome association and population-369 based linkage analyses. The American journal of human genetics, 81(3):559–575, 2007.370 22. Ying Wang, Kristin Tsuo, Masahiro Kanai, Benjamin M Neale, and Alicia R Martin. Challenges and opportunities371 for developing more generalizable polygenic risk scores. Annual review of biomedical data science, 5:293–320,372 2022.373 23. Tin Kam Ho. Random decision forests. In Proceedings of 3rd international conference on document analysis and374 recognition, volume 1, pages 278–282. IEEE, 1995.375 24. Leo Breiman. Random forests. Machine learning, 45:5–32, 2001.376 25. Jerome H Friedman. Greedy function approximation: a gradient boosting machine. Annals of statistics, pages377 1189–1232, 2001.378 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 13 26. Jerome H Friedman. Stochastic gradient boosting. Computational statistics & data analysis, 38(4):367–378, 2002.379 27. Y oav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to380 boosting. Journal of computer and system sciences, 55(1):119–139, 1997.381 28. Trevor Hastie, Saharon Rosset, Ji Zhu, and Hui Zou. Multi-class adaboost. Statistics and its Interface, 2(3):349–382 360, 2009.383 29. Hadeel Alzoubi, Raid Alzubi, and Naeem Ramzan. Deep learning framework for complex disease risk prediction384 using genomic variations. Sensors, 23(9):4439, 2023.385 30. Arn ´or I Sigurdsson, Ioannis Louloudis, Karina Banasik, David Westergaard, Ole Winther, Ole Lund, Sisse Rye386 Ostrowski, Christian Erikstrup, Ole Birger Vesterager Pedersen, Mette Nyegaard, et al. Deep integrative models387 for large-scale human genomics. Nucleic Acids Research, 51(12):e67–e67, 2023.388 31. Pau Bellot, Gustavo de Los Campos, and Miguel P ´erez-Enciso. Can deep learning improve genomic prediction of389 complex human traits? Genetics, 210(3):809–819, 2018.390 32. Yu Xu, Dragana Vuckovic, Scott C Ritchie, Parsa Akbari, Tao Jiang, Jason Grealey, Adam S Butterworth, Willem H391 Ouwehand, David J Roberts, Emanuele Di Angelantonio, et al. Machine learning optimized polygenic scores for392 blood cell traits identify sex-specific trajectories and genetic correlations with disease. Cell Genomics, 2(1), 2022.393 33. Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review,394 42:275–293, 2014.395 34. Cathie Sudlow, John Gallacher, Naomi Allen, Valerie Beral, Paul Burton, John Danesh, Paul Downey, Paul Elliott,396 Jane Green, Martin Landray, et al. Uk biobank: an open access resource for identifying the causes of a wide range397 of complex diseases of middle and old age. PLoS medicine, 12(3):e1001779, 2015.398 35. Heather J Cordell. Detecting gene–gene interactions that underlie human diseases. Nature Reviews Genetics,399 10(6):392–404, 2009.400 36. Or Zuk, Eliana Hechter, Shamil R Sunyaev, and Eric S Lander. The mystery of missing heritability: Genetic inter-401 actions create phantom heritability. Proceedings of the National Academy of Sciences, 109(4):1193–1198, 2012.402 37. Matthew B Taylor and Ian M Ehrenreich. Higher-order genetic interactions and their contribution to complex traits.403 Trends in genetics, 31(1):34–40, 2015.404 38. Tobias L Lenz, Aaron J Deutsch, Buhm Han, Xinli Hu, Yukinori Okada, Stephen Eyre, Michael Knapp, Alexandra405 Zhernakova, Tom WJ Huizinga, Goncalo Abecasis, et al. Widespread non-additive and interaction effects within406 hla loci modulate the risk of autoimmune diseases. Nature genetics, 47(9):1085–1090, 2015.407 39. Luis Varona, Andres Legarra, Miguel A Toro, and Zulma G Vitezica. Non-additive effects in genomic selection.408 Frontiers in genetics, 9:78, 2018.409 40. Marta Guindo-Mart ´ınez, Ramon Amela, Silvia Bon`as-Guarch, Montserrat Puiggr`os, Cecilia Salvoro, Irene Miguel-410 Escalada, Caitlin E Carey, Joanne B Cole, Sina R ¨ueger, Elizabeth Atkinson, et al. The impact of non-additive411 genetic associations on age-related complex diseases. Nature communications, 12(1):2436, 2021.412 41. James M Cheverud and Eric J Routman. Epistasis and its contribution to genetic variance components. Genetics,413 139(3):1455–1461, 1995.414 42. Jason H Moore and Scott M Williams. Epistasis and its implications for personal genetics. The American Journal415 of Human Genetics, 85(3):309–320, 2009.416 43. Ben Lehner. Molecular mechanisms of epistasis within and between genes. Trends in Genetics, 27(8):323–331,417 2011.418 44. Damian Szklarczyk, Rebecca Kirsch, Mikaela Koutrouli, Katerina Nastou, Farrokh Mehryary, Radja Hachilif, An-419 nika L Gable, Tao Fang, Nadezhda T Doncheva, Sampo Pyysalo, et al. The string database in 2023: protein–420 protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic421 acids research, 51(D1):D638–D646, 2023.422 45. Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In423 International Conference on Learning Representations, 2018.424 46. World Health Organization et al. Icd-10. international statistical classification of diseases and related health425 problems: Tenth revision 1992, volume 1= cim-10. classification statistique internationale des maladies et des426 probl`emes de sant´e connexes: Dixi`eme r´evision 1992, volume 1. 1992.427 47. 1000 Genomes Project Consortium et al. A global reference for human genetic variation. Nature, 526(7571):68,428 2015.429 48. Marc Lipsitch, Carl T Bergstrom, and Rustom Antia. Effect of human leukocyte antigen heterozygosity on infectious430 disease outcome: the need for allele-specific measures. BMC Medical Genetics, 4(1):1–9, 2003.431 49. Sue Tsai and Pere Santamaria. Mhc class ii polymorphisms, autoreactive t-cells, and autoimmunity. Frontiers in432 immunology, 4:321, 2013.433 50. Philippe Goyette, Gabrielle Boucher, Dermot Mallon, Eva Ellinghaus, Luke Jostins, Hailiang Huang, Stephan Ripke,434 Elena S Gusareva, Vito Annese, Stephen L Hauser, et al. High-density mapping of the mhc identifies a shared435 role for hla-drb1* 01: 03 in inflammatory bowel diseases and heterozygous advantage in ulcerative colitis. Nature436 genetics, 47(2):172–179, 2015.437 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint 14 H. Li et al. 51. Odd O Aalen and Søren Johansen. An empirical transition matrix for non-homogeneous markov chains based on438 censored observations. Scandinavian Journal of Statistics, pages 141–150, 1978.439 52. Dennis J Selkoe. Alzheimer’s disease: genes, proteins, and therapy. Physiological reviews, 2001.440 53. Rudolph E Tanzi and Lars Bertram. New frontiers in alzheimer’s disease genetics. Neuron, 32(2):181–184, 2001.441 54. Lars Bertram and Rudolph E Tanzi. Thirty years of alzheimer’s disease genetics: the implications of systematic442 meta-analyses. Nature Reviews Neuroscience, 9(10):768–778, 2008.443 55. Celeste M Karch, Carlos Cruchaga, and Alison M Goate. Alzheimer’s disease genetics: from the bench to the444 clinic. Neuron, 83(1):11–26, 2014.445 56. Y a-Ping Tang and Elliot S Gershon. Genetic studies in alzheimer’s disease. Dialogues in clinical neuroscience,446 2022.447 57. Henry B Mann and Donald R Whitney. On a test of whether one of two random variables is stochastically larger448 than the other. The annals of mathematical statistics, pages 50–60, 1947.449 58. Aravind Subramanian, Pablo Tamayo, Vamsi K Mootha, Sayan Mukherjee, Benjamin L Ebert, Michael A Gillette,450 Amanda Paulovich, Scott L Pomeroy, Todd R Golub, Eric S Lander, et al. Gene set enrichment analysis:451 a knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National452 Academy of Sciences, 102(43):15545–15550, 2005.453 59. Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P454 Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. Gene ontology: tool for the unification of biology.455 Nature genetics, 25(1):25–29, 2000.456 60. Minoru Kanehisa and Susumu Goto. Kegg: kyoto encyclopedia of genes and genomes. Nucleic acids research,457 28(1):27–30, 2000.458 61. Zhike Zhou, Yifan Liang, Xiaoqian Zhang, Junjie Xu, Jueying Lin, Rongwei Zhang, Kexin Kang, Chang Liu, Chuan-459 sheng Zhao, and Mei Zhao. Low-density lipoprotein cholesterol and alzheimer’s disease: a systematic review and460 meta-analysis. Frontiers in aging neuroscience, 12:5, 2020.461 62. Mirna Safieh, Amos D Korczyn, and Daniel M Michaelson. Apoe4: an emerging therapeutic target for alzheimer’s462 disease. BMC medicine, 17:1–17, 2019.463 63. Yi Jin, Kudzai Chifodya, Guochen Han, Wenxin Jiang, Yun Chen, Y ang Shi, Qiao Xu, Yilong Xi, Jun Wang, Jianping464 Zhou, et al. High-density lipoprotein in alzheimer’s disease: From potential biomarkers to therapeutics. Journal of465 Controlled Release, 338:56–70, 2021.466 64. Suzanne Craft. The role of metabolic disorders in alzheimer disease and vascular dementia: two roads converged.467 Archives of neurology, 66(3):300–305, 2009.468 65. Xu Y an, Yue Hu, Biyao Wang, Sijian Wang, and Xinwen Zhang. Metabolic dysregulation contributes to the progres-469 sion of alzheimer’s disease.Frontiers in neuroscience, 14:530219, 2020.470 66. MS Tsai, Eric G Tangalos, Ronald C Petersen, Glenn E Smith, Daniel J Schaid, Emre Kokmen, Robert J Ivnik, and471 Stephen N Thibodeau. Apolipoprotein e: risk factor for alzheimer disease. American journal of human genetics,472 54(4):643, 1994.473 67. Warren J Strittmatter and Allen D Roses. Apolipoprotein e and alzheimer disease. Proceedings of the National474 Academy of Sciences, 92(11):4725–4727, 1995.475 68. Marion R Meyer, JoAnn T Tschanz, Maria C Norton, Kathleen A Welsh-Bohmer, David C Steffens, Bonita W Wyse,476 and John Breitner. Apoe genotype predicts when—not whether—one is predisposed to develop alzheimer disease.477 Nature genetics, 19(4):321–322, 1998.478 69. Robert C Green, J Scott Roberts, L Adrienne Cupples, Norman R Relkin, Peter J Whitehouse, Tamsen Brown,479 Susan LaRusse Eckert, Melissa Butson, A Dessa Sadovnick, Kimberly A Quaid, et al. Disclosure of apoe genotype480 for risk of alzheimer’s disease.New England Journal of Medicine, 361(3):245–254, 2009.481 70. Emmanuelle Genin, Didier Hannequin, David Wallon, Kristel Sleegers, Mikko Hiltunen, Onofre Combarros,482 Marıa Jes´us Bullido, Sebastiaan Engelborghs, Peter De Deyn, Claudine Berr, et al. Apoe and alzheimer disease:483 a major gene with semi-dominant inheritance. Molecular psychiatry, 16(9):903–907, 2011.484 71. Yu Y amazaki, Na Zhao, Thomas R Caulfield, Chia-Chen Liu, and Guojun Bu. Apolipoprotein e and alzheimer485 disease: pathobiology and targeting strategies. Nature Reviews Neurology, 15(9):501–518, 2019.486 72. Robert A Hegele, W Carl Breckenridge, Diane W Cox, Graham F Maguire, J Alick Little, and Philip W Connelly. In-487 teraction between variant apolipoproteins c-ii and e that affects plasma lipoprotein concentrations.Arteriosclerosis488 and Thrombosis: A Journal of Vascular Biology, 11(5):1303–1309, 1991.489 73. Sebasti ´an Cervantes, Llu ´ıs Samaranch, Jos ´e Manuel Vidal-Taboada, Isabel Lamet, Mar ´ıa Jes´us Bullido, Ana490 Frank-Garc´ıa, Francisco Coria, Albert Lle´o, Jordi Clarim´on, Elena Lorenzo, et al. Genetic variation in apoe cluster491 region and alzheimer’s disease risk.Neurobiology of aging, 32(11):2107–e7, 2011.492 74. Qin Zhou, Fan Zhao, Ze-ping Lv, Chen-guang Zheng, Wei-dong Zheng, Liang Sun, Na-na Wang, Shenghang Pang,493 Fabiana Michelsen de Andrade, Mian Fu, et al. Association between apoc1 polymorphism and alzheimer’s disease:494 a case-control study and meta-analysis. PloS one, 9(1):e87017, 2014.495 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint PRS-Net 15 75. Yvonne Shao, McKenzie Shaw, Kaitlin Todd, Maria Khrestian, Giana D’Aleo, P John Barnard, Jeff Zahratka, Jagan496 Pillai, Chang-En Yu, C Dirk Keene, et al. Dna methylation of tomm40-apoe-apoc2 in alzheimer’s disease.Journal497 of human genetics, 63(4):459–471, 2018.498 76. Alexander M Kulminski, Ethan Jain-Washburn, Ian Philipp, Liang He, Yury Loika, Elena Loiko, Olivia Bagley, Svet-499 lana Ukraintseva, Anatoliy Y ashin, Konstantin Arbeev, et al. Apoe epsilon4 allele and tomm40-apoc1 variants jointly500 contribute to survival to older ages. Aging Cell, 21(12):e13730, 2022.501 77. Florence F Roussotte, Boris A Gutman, Sarah K Madsen, John B Colby, Paul M Thompson, Alzheimer’s Dis-502 ease Neuroimaging Initiative, et al. Combined effects of alzheimer risk variants in the clu and apoe genes on503 ventricular expansion patterns in the elderly. Journal of Neuroscience, 34(19):6537–6545, 2014.504 78. Hyo Lee, Aimee J Aylward, Richard V Pearse, Yi-Chen Hsieh, Zachary M Augur, Courtney R Benoit, Vicky Chou,505 Allison Knupp, Cheryl Pan, Srilakshmi Goberdhan, et al. Cell-type-specific regulation of apoe levels in human506 neurons by the alzheimer’s disease risk gene sorl1.bioRxiv, pages 2023–02, 2023.507 79. Pranav Preman and Amaia M Arranz. A neuron-specific interaction between alzheimer’s disease risk factors sorl1,508 apoe, and clu. Cell Reports, 42(9), 2023.509 80. Ute Traugott. Multiple sclerosis: relevance of class i and class ii mhc-expressing cells to lesion development.510 Journal of neuroimmunology, 16(2):283–302, 1987.511 81. Sunhee C Lee, GR Wayne Moore, George Golenwsky, and Cedric S Raine. Multiple sclerosis: a role for astroglia512 in active demyelination suggested by class ii mhc expression and ultrastructural study. Journal of neuropathology513 and experimental neurology, 49(2):122–136, 1990.514 82. Multiple Sclerosis Genetics Group, Jonathan L Haines, Henry A Terwedow, Katie Burgess, Margaret A Pericak-515 Vance, Jackie B Rimmler, Eden R Martin, Jorge R Oksenberg, Robin Lincoln, David Y Zhang, et al. Linkage of516 the mhc to familial multiple sclerosis suggests genetic heterogeneity.Human molecular genetics, 7(8):1229–1234,517 1998.518 83. Matthew R Lincoln, Alexandre Montpetit, M Zameel Cader, Janna Saarela, David A Dyment, Milvi Tiislar, Vincent519 Ferretti, Pentti J Tienari, A Dessa Sadovnick, Leena Peltonen, et al. A predominant role for the hla class ii region520 in the association of the mhc region with multiple sclerosis. Nature genetics, 37(10):1108–1112, 2005.521 84. David A Dyment, Blanca M Herrera, M Zameel Cader, Cristen J Willer, Matthew R Lincoln, A Dessa Sadovnick,522 Neil Risch, and George C Ebers. Complex interactions among mhc haplotypes in multiple sclerosis: susceptibility523 and resistance. Human molecular genetics, 14(14):2019–2026, 2005.524 85. Class ii hla interactions modulate genetic risk for multiple sclerosis. Nature genetics, 47(10):1107–1113, 2015.525 .CC-BY-NC 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 15, 2024. ; https://doi.org/10.1101/2024.02.13.580211doi: bioRxiv preprint