Deciphering proteins in Alzheimer's disease: A new Mendelian randomization method integrated with AlphaFold3 for 3D structure prediction

Abstract

Hidden confounding biases hinder identifying causal protein biomarkers for Alzheimer's disease in non-randomized studies. While Mendelian randomization (MR) can mitigate these biases using protein quantitative trait loci (pQTLs) as instrumental variables, some pQTLs violate core assumptions, leading to biased conclusions. To address this, we propose MR-SPI, a novel MR method that selects valid pQTL instruments using Leo Tolstoy's Anna Karenina principle and performs robust post-selection inference. Integrating MR-SPI with AlphaFold3, we developed a computational pipeline to identify causal protein biomarkers and predict 3D structural changes. Applied to genome-wide proteomics data from 54,306 UK Biobank participants and 455,258 subjects (71,880 cases and 383,378 controls) for a genome-wide association study of Alzheimer's disease, we identified seven proteins (TREM2, PILRB, PILRA, EPHA1, CD33, RET, and CD55) with structural alterations due to missense mutations. These findings offer insights into the etiology and potential drug targets for Alzheimer's disease.

Keywords: 3D structure; AlphaFold3; Alzheimer's disease; Mendelian randomization; causal inference; genome-wide association study; instrumental variable; missense variants; pQTL; proteomics.

Graphical abstract

Figure 1

Overview of the pipeline First, we apply MR-SPI for each protein to (1) select valid pQTL IVs under the plurality condition and (2) estimate the causal effect on the outcome of interest. Second, we perform the Bonferroni correction procedure for causal protein identification. Third, for each causal protein biomarker, we apply AlphaFold3 to predict the 3D structural alterations due to missense pQTL IVs selected by MR-SPI.

Figure 2

The MR-SPI framework First, MR-SPI selects relevant IVs with strong pQTL-protein associations. Second, each relevant IV provides a ratio estimate of the causal effect and then receives votes on itself to be valid from the other relevant IVs whose degrees of violation of (A2) and (A3) are small under this ratio estimate of causal effect. For example, by assuming pQTL 1 is valid, the slope of the line connecting pQTL 1 and the origin represents the ratio estimate of pQTL 1, and pQTLs 2 and 3 vote for pQTL 1 to be valid because they are close to that line, while pQTLs 4, 5, and 6 vote against it since they are far away from that line. Third, MR-SPI estimates the causal effect by fitting a zero-intercept ordinary least squares regression of pQTL-outcome associations on pQTL-protein associations and construct the robust CI using selected valid pQTL IVs in the maximum clique of the voting matrix, which encodes whether two pQTLs mutually vote for each other to be valid IVs.

Figure 3

Empirical performance of MR-SPI and the other competing MR methods in simulated data with a sample size of 5,000 (A) Boxplot of the percent bias in causal effect estimates. (B) Empirical coverage of 95% CIs. The black dashed line represents the nominal level (95%). (C) Average lengths of 95% CIs.

Figure 4

Causal effect estimates, GO analysis, and structural alteration prediction for putatively causal proteins (A) Volcano plot of associations of plasma proteins with AD using MR-SPI. The horizontal axis represents the estimated effect size (on the log odds ratio scale), and the vertical axis represents the $-{\log }_{10}(p$ value). Positive and negative associations are represented by green and red points, respectively. The size of a point is proportional to $-{\log }_{10}(p$ value). The blue dashed line represents the significance threshold using Bonferroni correction ($p$$<5.48\times {10}^{-5}$). (B) 3D Structural alterations of CD33 predicted by AlphaFold3 due to missense genetic variation of pQTL rs2455069. The ribbon representation of 3D structures of CD33 with arginine and glycine at position 69 are colored in blue and red, respectively. The amino acids at position 69 are displayed in stick representation, with arginine and glycine colored in green and yellow, respectively. The predicted template modeling yields a score of 0.6 for both structures, which suggests that AlphaFold3 provides good predictions for these two 3D structures. (C) Forest plot of significant associations of proteins with AD identified by MR-SPI. The error bars represent the 95% CIs of MR-SPI and other competing MR methods, with CIs clipped to vertical axis limits. (D) Bubble plot of GO analysis results using the 7 significant proteins detected by MR-SPI. The horizontal axis represents the Z score of the enriched GO term, and the vertical axis represents the $-{\log }_{10}(p$ value) after Bonferroni correction. Each point represents one enriched GO term. The blue dashed line represents the significance threshold (adjusted $p$$<0.05$ after Bonferroni correction).