Abstract
Alzheimer’s disease (AD) is the leading cause of death among individuals over 65. Despite many AD genetic variants detected by large genome-wide association studies (GWAS), a limited number of causal genes have been confirmed. Conventional machine learning techniques integrate functional annotation data and GWAS signals to assign variants functional relevance probabilities. Yet, a large proportion of genetic variation lies in the non-coding genome, where unsupervised and semi-supervised techniques have demonstrated greater advantage. Furthermore, cell-type specific approaches are needed to better understand disease etiology. Studying AD from a microglia-specific lens is more likely to reveal causal variants involved in immune pathways. Therefore, in this study, we developed S-BEAM: a semi-supervised ensemble approach using microglia-specific data to prioritize non-coding variants and their target genes that play roles in immune-related AD mechanisms. We designed a transductive positive-unlabeled and negative-unlabeled learning model that employs a bagging technique to learn from unlabeled variants, generating multiple predicted probabilities of variant risk. Using a combined homogeneous-heterogeneous ensemble framework, we aggregated the predictions. We applied our model to AD variant data, identifying 11 risk variants acting in well-known AD genes, such as TSPAN14, INPP5D, and MS4A2. These results validated our model’s performance and demonstrated a need to study these genes in the context of microglial pathways. We also proposed further experimental study for 37 potential causal variants associated with less-known genes. Our work has utility in predicting AD relevant genes and variants functioning in microglia and can be generalized for application to other complex diseases or cell types.
Introduction
Genome-wide association studies (GWAS) have successfully identified numerous loci associated with Alzheimer’s Disease (AD). However, pinpointing the specific causal variants and their target genes remains a significant challenge. Over 90% of disease-associated variants reside in non-coding regions, functioning as regulatory elements rather than directly altering protein sequences. Additionally, genetic variants often exhibit cell-type-specific activity. Given the emerging role of microglia in AD neuroinflammation, general models often fail to capture critical signals specific to these immune cells.
Methodology
To address these challenges, we developed S-BEAM (Semi-supervised Bagging Ensemble Approach for Microglia). The framework is specifically designed to handle the "missing label" problem in genomics, where only a few "risk" variants are confirmed.
The S-BEAM pipeline utilizes a transductive positive-unlabeled (PU) and negative-unlabeled (NU) learning approach. This method treats the vast majority of variants as "unlabeled" rather than assuming they are non-causal, thereby reducing bias and false negatives. The architecture employs a homogeneous-heterogeneous ensemble combining five base classifiers: Decision Trees, Random Forests, Multi-layer Perceptrons (MLP), XGBoost, and Logistic Regression. Bagging techniques are applied to repeatedly train these models on subsets of unlabeled data, aggregating the results into a robust risk score.
Data Integration and Features
The model integrates high-dimensional datasets, including a large-scale AD GWAS meta-analysis and nine microglia-specific functional annotation features.
Key features include ATAC-seq for chromatin accessibility, H3K27ac ChIP-seq for histone modifications, and pcHi-C for 3D chromatin interactions. These features allow the model to identify variants that may be physically distant from a gene on a linear DNA strand but regulate it through 3D chromatin folding. Training was conducted using a high-confidence set of 14 positive and 14 negative controls derived from CRISPRi experiments.
Results and Discussion
S-BEAM prioritized 155 top-performing variants from an initial pool of tens of thousands. The model successfully identified 11 variants linked to established AD genes, such as INPP5D, MS4A2, and TSPAN14, validating its predictive power.
Furthermore, the model identified 37 potential causal variants associated with less-established AD genes. A notable discovery is the strong link between variant rs11039131 and the CELF1 locus, which is involved in microglial phagocytosis—the process of clearing toxic amyloid-beta plaques. Performance metrics indicated that the ensemble approach significantly outperformed single-model methods, with tree-based algorithms (Random Forest and XGBoost) achieving sensitivity levels exceeding 90%.
Conclusion and Future Directions
S-BEAM demonstrates the power of semi-supervised learning in leveraging large unlabeled genomic datasets to overcome the limitations of small labeled training sets. By focusing on microglia-specific data, the model provides a specialized lens for understanding AD etiology. Future work will involve experimental validation of the 37 newly identified variants using CRISPR-based cellular assays and the application of the S-BEAM framework to other cell types and complex neurodegenerative diseases.
Summary updated