Finemap-MiXeR: A variational Bayesian approach for genetic finemapping

Bayram Cevdet Akdeniz, Oleksandr Frei, Alexey Shadrin, Dmitry Vetrov, Dmitry Kropotov, Eivind Hovig, Ole A. Andreassen, Anders M. DalebioRxiv 2022 · DOI

TL;DR

This paper introduces Finemap-MiXeR, a novel variational Bayesian method for pinpointing causal genetic variants within complex genomic regions. It achieves higher accuracy and computational efficiency compared to existing methods, making it a powerful tool for interpreting genetic association studies.

What Problem Does This Paper Address?

Genome-wide association studies (GWAS) are excellent at identifying regions of the genome associated with traits or diseases. However, these regions often contain many genetic variants (SNPs) that are highly correlated due to linkage disequilibrium (LD). This makes it challenging to determine which specific SNP(s) are truly responsible for the observed association. This problem, known as genetic finemapping, is crucial for understanding the biological mechanisms underlying complex traits and for developing targeted interventions. Existing finemapping methods can be computationally intensive, especially when dealing with a large number of SNPs or multiple causal variants within a locus.

Methods

The core of Finemap-MiXeR is a variational Bayesian approach applied to the MiXeR model. The MiXeR model is designed to infer causal variants by considering a biologically plausible prior distribution of genetic effect sizes. Finemap-MiXeR optimizes the Evidence Lower Bound (ELBO) of the MiXeR model's likelihood function. This optimization is performed using the Adaptive Moment Estimation (ADAM) algorithm, which efficiently finds the posterior probabilities of each SNP being causal. A key innovation is that the computational complexity of Finemap-MiXeR scales linearly with the number of SNPs (M) and is independent of the number of causal SNPs (k) or heritability (h²), unlike some other methods. The authors also propose a Finemap-MiXeR PCA variant that further reduces computational complexity by using Principal Component Analysis.

Datasets

  • Synthetic genomic data: Generated using the Hapgen2 tool to simulate various scenarios with controlled parameters (number of SNPs, causal variants, heritability).
  • UK Biobank (UKB) data: Used for real-world application, specifically focusing on standing height as the phenotype. The data was split into 80% for training and 20% for testing.

Key Insights and Ideas

  • Variational Bayesian approach for MiXeR: Adapting variational inference to the MiXeR model provides a computationally tractable way to estimate posterior probabilities of causality.
  • Computational efficiency independent of k and h²: This is a significant advantage, as it means the method remains efficient even in scenarios with many causal variants or high heritability, which can be bottlenecks for other methods.
  • MiXeR model's flexibility: The underlying MiXeR model's ability to infer Bayesian inference for finemapping allows it to detect more causal variants, leading to better accuracy in real-world applications.
  • PCA for further optimization: The Finemap-MiXeR PCA variant demonstrates that computational complexity can be further reduced without sacrificing accuracy, making it even more scalable.

Key Results

  • Accuracy: Finemap-MiXeR consistently outperformed or matched existing methods (FINEMAP and SuSiE) in terms of accuracy (measured by Area Under the Curve - AUC) across various synthetic scenarios. For example, in one scenario, Finemap-MiXeR achieved a mean AUC of 0.932, compared to SuSiE's 0.875 and FINEMAP's 0.920.
  • Real Data Performance: When applied to UK Biobank height data, Finemap-MiXeR also showed superior performance. In one specific locus, Finemap-MiXeR detected 24 causal SNPs for prediction, while other methods detected only 7, resulting in a significantly higher R² value for phenotype prediction.
  • Computational Complexity: The runtime of Finemap-MiXeR scales quadratically with the number of SNPs (M) but is independent of the number of causal variants (k) and heritability (h²). The Finemap-MiXeR PCA variant further reduces this scaling to be linear with M.

Limitations and Future Work

  • Equal Priors: The current model assumes equal priors for all SNPs. The authors suggest that using individual priors for each SNP could lead to significant improvements and is a direction for future work.
  • Hyperparameter Optimization: While the authors used fixed or pre-defined hyperparameters for the MiXeR model, optimizing these parameters could also enhance performance.
  • Gene Set Enrichment Analysis: The authors propose that the MiXeR model's framework could be extended to applications like gene set enrichment analysis, opening up new avenues for functional genomics research.
  • Optimization of PCA: Further optimization of the PCA computation is planned for future work.

Summary updated

Abstract

Abstract Discoveries from genome-wide association studies often contain large clusters of highly correlated genetic variants, which makes them hard to interpret. In such cases, finemapping the underlying causal variants become important. Here we present a new method, the Finemap-MiXeR, based on a variational Bayesian approach for finemapping genomic data, i.e., determining the causal single nucleotide polymorphisms (SNPs) associated with a trait at a given locus after controlling for correlation among genetic variants due to linkage disequilibrium. Our approach is based on the optimization of Evidence Lower Bound of the likelihood function obtained from the MiXeR model. The optimization is done using Adaptive Moment Estimation Algorithm, allowing to obtain posterior probability of each SNP to be a causal variant. We tested Finemap-MiXeR in a range of different scenarios, using both synthetic and real data from the UK Biobank, using standing height phenotype as an example. In comparison to the existing finemapping methods FINEMAP and SuSiE methods, we observed that Finemap-MiXeR in most cases has better accuracy. Furthermore, it is computationally efficient, and unlike other methods the complexity is not increasing as the number of causal SNPs or the heritability increases. We show that our finemapping algorithm identifies a small number of genetic variants per locus which are informative for predicting the phenotype in an independent sample.
BibTeX
@misc{akdeniz2022finemap,
  author = {Bayram Cevdet Akdeniz and Oleksandr Frei and Alexey Shadrin and Dmitry Vetrov and Dmitry Kropotov and Eivind Hovig and Ole A. Andreassen and Anders M. Dale},
  title = {{Finemap-MiXeR: A variational Bayesian approach for genetic finemapping}},
  year = {2022},
  doi = {10.1101/2022.11.30.518509},
  howpublished = {bioRxiv},
}
Figure