Breakthrough in Alzheimer’s Disease Research! MR-SPI Crushes IVW with ‘Voting Rules’

1 Introduction

This introduction discusses the pathological complexity of Alzheimer’s Disease (AD) and the limitations of traditional research methods, gradually leading to the technical innovations and scientific significance of the study.

1.1 Disease Background and Clinical Needs

• Epidemiological Data: AD accounts for 60%-70% of global dementia cases, with an estimated 152 million patients by 2050, but existing drugs can only alleviate symptoms and cannot reverse disease progression.• Pathological Hypothesis Dilemma: The failure rate of phase III clinical trials targeting β-amyloid (Aβ) and Tau exceeds 99%, indicating the limitations of a single-target strategy.• Unmet Needs: There is an urgent need to systematically analyze the multi-pathway regulatory network of AD and develop causal relationship-based target discovery methods.

1.2 Challenges of Traditional Genetic Methods

• Limitations of Mendelian Randomization (MR): • Invalid Instrument Variable Interference: Among genetic instrument variables (pQTL) for plasma proteomics, about 30%-50% exhibit pleiotropy or linkage disequilibrium, leading to biased causal effect estimates.• Insufficient Statistical Power: Traditional MR methods (such as IVW, MR-Egger) require pre-specified instrument variable validity, significantly increasing error rates in scenarios with few pQTLs.• Disconnection in Mechanistic Research: Most MR studies only report statistical associations between proteins and diseases, lacking in-depth analysis of molecular mechanisms (such as structural effects of missense mutations).

1.3 Innovative Strategies of This Study

• Methodological Breakthrough: This study proposes MR-SPI (Mendelian Randomization via Stability-guided Protein Identification), which addresses the issue of invalid instrument variables through stability screening, specifically including:• Voting Matrix Construction: Conducting 10 heterogeneity tests (such as Cochran’s Q, MR-PRESSO) for each pQTL, retaining only loci with consistency ≥80%.• Non-parametric Statistical Framework: No need to pre-specify the distribution of instrument variables, significantly enhancing robustness in small sample scenarios (simulations show a 2.3-fold increase in statistical power).• Technical Integration: For the first time, AlphaFold3 is applied to protein target discovery in MR, predicting three-dimensional structural changes caused by missense mutations and associating them with clinical phenotypes.• Data Scale: Integrating UK Biobank (54,306 individuals’ proteomics) with AD GWAS (450,000 cases and controls), covering genetic-phenotypic associations of 1,463 plasma proteins.

1.4 Research Significance

• Scientific Value: Establishing a full-chain research paradigm of “genetic association-structural mechanism-drug development” provides a scalable methodological framework for AD and other complex diseases.• The Core Logical Chain of This Paper: Complexity of AD pathology → Defects of traditional MR methods → Innovations in MR-SPI algorithm → Multi-omics data validation → AI structural mechanism analysis → Targeted therapy translation

2 Research Overview

This study integrates Mendelian Randomization (MR) with protein structure prediction technology to systematically screen plasma proteins causally associated with Alzheimer’s Disease (AD) and analyze their molecular mechanisms. The research process is illustrated in Figure 1.

Figure 1. Research Design Framework: Step 1: Obtain genetic data from UK Biobank and ADGWASStep 2: Apply MR-SPI to screen for valid instrument variablesStep 3: Predict structural changes of missense mutations using AlphaFold3

3 Methods

Methodological Breakthrough: Technical Details of the MR-SPI Algorithm

3.1 Symbol Definition and Problem Modeling

1. Data Definition: • For the th protein, observe samples’ exposure (protein level ) and outcome (disease status )• Candidate instrument variable (IV) set , where is associated pQTL ()

2. Causal Effect Model: • : causal effect on (target estimate)• : direct effect on (pleiotropy), if , then is an invalid IV

3. Core Objective: Robust estimation of in the presence of unknown pleiotropy ().

3.2 Stability Screening: Voting Matrix Construction

MR-SPI selects valid IVs through multiple testing consistency, with the following steps:

1. Heterogeneity Test Set: Define , containing 10 heterogeneity testing methods:• Traditional methods: Cochran’s Q test, MR-Egger intercept test• Leave-one-out sensitivity analysis• Robust regression test (RANSAC)• Permutation test (Permutation-based threshold)• Other custom tests (e.g., instrument variable clustering consistency)

2. Voting Matrix Construction: • Rows: candidate IVs

   • Columns: testing methods

   • Elements , indicating a test result of 1, otherwise 0

3. Valid IV Selection Criteria: • Threshold (based on 80% tests)• Plurality Rule Essence: Valid IVs must perform consistently across the majority of tests, while invalid IVs show dispersed results due to pleiotropy

3.3 Permutation Test to Determine Significance Threshold

To avoid subjective setting of , a permutation test calibration is used:

1. Zero Distribution Generation: • Randomly permute times, disrupting the causal association between and • For each permuted dataset, calculate the voting scores for all

Family-wise Error Rate Control: • Calculate the maximum voting score • Take • If the actual data satisfies , retain these IVs

3.4 Causal Effect Estimation and Variance Calculation

1. Robust Weighted Estimator: For the selected IV set , use an improved form of inverse variance weighting (IVW):

   • : single IV causal effect estimate

   • : precision weight

2. Variance Estimation:

   • Use bias-corrected variance estimator to address selection bias introduced by screening

3.5 Theoretical Properties Proof

1. Consistency Condition: If the following conditions are met, $ ilde{\beta} j$ is a consistent estimator:• Plurality Rule: Valid IVs must be the majority (even if less than 50%)• Stability Condition: The voting score of valid IVs $ extstyle extsum V{ks} $ is probabilistically greater than that of invalid IVs

2. Asymptotic Normality: When , follows an asymptotic normal distribution:• Variance can be calculated using Bootstrap or Sandwich estimators

3.6 Summary of Innovations

1. No Need for Pre-specified Distribution Assumptions: • Traditional methods require assuming pleiotropy follows a specific distribution (e.g., normal mixture), while MR-SPI avoids this limitation through a data-driven voting mechanism• Avoids NOME assumption (No Measurement Error), allowing for measurement errors in exposure variables

2. Computational Efficiency Optimization: • The complexity of constructing the voting matrix is , significantly lower than the full likelihood-based MRMix ()• Parallel design: 10 tests can be computed in a distributed manner

3. Enhanced Interpretability: • Voting scores provide a quantifiable reliability indicator for IVs (e.g., score 8 vs. 5)• Permutation test calibration avoids subjective threshold settings

3.7 Mathematical Appendix: Key Formula Derivation

Bias-Corrected Variance Estimator for IVW Estimator: Traditional IVW variance underestimates the true variance, as it ignores the “winner’s curse” effect caused by IV screening. MR-SPI adopts the following correction:where is estimated through leave-one-out cross-validation.

This table systematically compares 8 mainstream MR methods in handling invalid instrument variables (Invalid IVs) statistical assumptions, providing a theoretical basis for method selection. The following interpretations are based on three major assumption dimensions:

Core Assumption Definitions

1. InSIDE Assumption (Instrument Strength Independent of Direct Effect)• Meaning: The strength of the association between the instrument variable (IV) and the exposure factor (X) must be independent of its pleiotropic effects (i.e., the IV’s effect on outcome Y is mediated only through X).• Importance: If this assumption is violated (e.g., directional pleiotropy exists), MR estimates will be biased.

2. Majority/Plurality Rule Condition• Majority Rule: >50% of IVs are valid (no pleiotropy).• Plurality Rule: The proportion of valid IVs is maximized (but may be less than 50%).

3. Other Distribution Assumptions• Balanced Pleiotropy: The mean of pleiotropic effects is zero (but individual deviations are allowed).• NOME Assumption (No Measurement Error): Measurement of exposure X is error-free (or errors are negligible).

Key Conclusions and Selection Recommendations

1. Methods Dependent on InSIDE Assumption (IVW, MR-Egger, MR-RAPS, MR-PRESSO)• Advantages: More efficient estimates when assumptions are met.• Limitations: Pleiotropy is common in real scenarios, and assumptions are easily violated.• Applicable Scenarios: Research pre-screening high-confidence IVs (e.g., protein-coding SNPs).

2. Methods Not Dependent on InSIDE Assumption (Weighted Median, Mode Estimation, MRMix, Contamination Methods, MR-SPI)• Advantages: More robust against invalid IVs, especially suitable for complex traits (e.g., plasma proteins).• Unique Value of MR-SPI: ◦ Only requires plurality rule (valid IVs must be maximized, no requirement for >50%)◦ Zero Distribution Assumption: No need to pre-specify pleiotropy distribution or measurement error models◦ Computational Efficiency: Automatically screens consistent IVs through voting matrix (as shown in Figure 2A)

3. Method Selection Trade-offs• When the proportion of valid IVs is high (>50%): Prefer MR-PRESSO or Weighted Median• In the presence of unknown pleiotropy: MR-SPI or Mode Estimation is more robust• Small Sample Scenarios: MR-RAPS or MRMix (must meet distribution assumptions)

• Reasons for Choosing MR-SPI in This Study:

1. High proportion of invalid IVs among genetic instrument variables (pQTL) for plasma proteins (about 30%-50%)
2. Need to address potential non-balanced pleiotropy (e.g., confounding by the inflammatory pathway of CD33)
3. Avoid NOME assumption (technical variation in proteomics measurements)

Instrument Variable Screening Criteria

• Primary Screening: Extract loci significantly associated with protein expression (p<5×10⁻⁸) from 214,857 pQTLs
• Consistency Testing: Each candidate pQTL must pass 10 heterogeneity tests (Cochran’s Q test, MR-Egger intercept test, etc.), retaining only loci that pass ≥8 tests (green area in Figure 2A)
• Sensitivity Analysis: Validate the impact weight of individual SNPs on results through leave-one-out analysis

Comparison of Effectiveness with Traditional MR

In simulations with 50% invalid instrument variables, the false positive rate of traditional MR reached 32.7% (95% CI 28.4-37.1%), while MR-SPI controlled the false positive rate at 4.8% (95% CI 3.1-6.5%) (Figure 2B). Computational efficiency improved: processing proteomics data from 54,306 samples took only 2.3 hours (traditional MR required 5.7 days).

4 Core Process Analysis of the MR-SPI Framework

This diagram systematically illustrates the workflow of the MR-SPI method, with the core goal of addressing the issue of invalid IVs in traditional Mendelian Randomization (MR) through stability screening of instrument variables (IVs).

4.1 Selecting Relevant IVs

1. Input Data and Screening Criteria• pQTL-Protein Associations (left half): Screen genetic loci (pQTL) significantly associated with target protein expression, which must meet:◦ Statistical significance (p < 5 × 10^{-8})◦ Consistency in genetic effect direction (to avoid linkage disequilibrium confounding)• pQTL-Outcome Associations (right half): Extract the association effect of pQTL with disease outcomes (e.g., AD) (\hat{\gamma}_Y)

2. Threshold Setting and Classification• Relevant IVs (blue nodes): Strong instrument variables that meet both pQTL-protein and pQTL-outcome associations• Weak IVs (gray nodes): Loci that pass only a single association test, excluded in subsequent steps

3. Scientific Significance: • Avoid weak instrument bias (Weak Instrument Bias)• Reduce interference of invalid IVs on causal effect estimates

4.2 Selecting Valid IVs by Voting

1. Voting Mechanism Design (dense network in the middle of the figure)• Ratio Estimate: Each pQTL calculates (X is protein level)• Consistency Testing: Evaluate the stability of each pQTL estimate through permutation testing (Permutation Test)◦ Support Votes (green lines): If the confidence interval of a pQTL estimate overlaps with another pQTL, it counts as a vote◦ Opposition Votes (red crosses): No overlap in confidence intervals counts as opposition votes

2. Valid IV Selection Criteria• Valid IVs (green nodes): Receive support votes exceeding a threshold (e.g., 80%)• Invalid IVs (red nodes): Insufficient support votes or excluded due to pleiotropy (horizontal pleiotropy)

3. Algorithm Innovations: • Non-parametric Screening: No pre-specified pleiotropy distribution, identifying stable IVs through data-driven methods• Plurality Rule Implementation: Valid IVs only need to be the largest consistent group, without needing to exceed 50%

4.3 Causal Effect Estimation

1. Robust Weighted Estimation (lower half of the figure)• Maximum Clique: Select the pQTL subset with the highest support votes (e.g., pQTL4 and pQTL6)• Inverse Variance Weighting (IVW): Calculate the weighted average causal effect

2. Confidence Interval Construction: • Use Bootstrap method (1,000 resamples) to calculate 95% confidence intervals• Calibrate interval coverage probability through permutation testing to avoid overfitting

3. Output Validation: • Sensitivity Analysis: Leave-one-out test (stability of effects after excluding individual pQTL)• Direction Consistency: The effect direction of all valid IVs must be consistent (to avoid heterogeneity bias)

4.4 Core Value of the Framework

1. Addressing Traditional MR Pain Points: • Invalid IV Interference: Traditional methods (e.g., MR-Egger) fail when the proportion of invalid IVs is high, while MR-SPI filters through the voting mechanism• Dependence on Distribution Assumptions: Avoids pre-specifying pleiotropy distribution or measurement errors, enhancing result generalizability

2. Visualization Advantages: • Color Coding: Green/red distinguishes valid/invalid IVs, blue/gray distinguishes relevant/weak IVs, enhancing interpretability• Network Topology: Voting line density intuitively reflects the strength of consistency among IVs

3. Computational Efficiency: • Through pre-screening (exclusion of weak IVs) and parallel voting calculations, the time complexity of processing tens of thousands of pQTLs is reduced to (O(n log n))

4.4 Application Scenario Example (Using AD Research as an Example)

• Step 1: Screen pQTLs associated with CD33 protein (e.g., rs2455069)• Step 2: Voting excludes loci with pleiotropic associations with inflammatory pathways (e.g., rs123456)• Step 3: Estimate the causal effect of CD33 on AD based on remaining valid IVs (OR=1.24, 95%CI 1.12-1.38)

This flowchart not only clarifies the technical logic of the MR-SPI method but also reveals its core advantages of stability screening, non-parametric estimation, and efficient computation, providing reliable methodological tools for causal inference in complex diseases.

5 Simulation Studies

This figure systematically compares MR-SPI with traditional MR methods (IVW, MR-PRESSO, etc.) in causal effect estimation using simulated data (sample size 5,000), revealing the following core conclusions:

1. MR-SPI performs optimally in the presence of invalid instrument variables (Locally Invalid IV): • Minimal estimation bias (Figure A), coverage closest to the nominal level of 95% (Figure B), reasonable confidence interval lengths (Figure C).• Robustness significantly exceeds traditional methods (e.g., IVW, MR-PRESSO), which suffer severe bias and coverage deterioration in the presence of invalid IVs.
2. Traditional methods are highly sensitive to invalid IVs: • Even using correction methods like Weighted Median or Contamination Fixture, they cannot resolve bias and coverage issues caused by invalid IVs.

Method Comparison Highlights

• MR-SPI Advantage Mechanism: • Majority/Plurality Rule: Filters valid instrument variables through a statistical voting mechanism (marked S1-S4), reducing invalid IV interference.• Robust Confidence Intervals (robust CI): Adopts asymmetric error estimation, adapting to complex data distributions (see legend “MR-SPI (robust CI)”).• Limitations of Traditional Methods: • IVW assumes all instrument variables are valid, lacking correction capabilities for invalid IVs.• MR-PRESSO relies on outlier removal, easily affected by a high proportion of invalid IVs.

6 Case Analysis

This figure systematically integrates causal effect analysis, structural prediction, and functional annotation of proteins related to Alzheimer’s Disease (AD), revealing the following key scientific findings through four sub-figures (A-D):

5.1 Causal Effect Identification (Figure A & Figure C)

• Figure A (Volcano Plot): • Horizontal Axis: Causal effect values of proteins on AD (Log Odds Ratio),green points indicate positive associations (increased risk),red points indicate negative associations (protective effects).• Vertical Axis: Significance level (-log₁₀(p-value)), the size of the points reflects the strength of significance.• Key Results: ◦ Proteins such as CD33, TREM2, and PILRB are significantly associated with AD (p < 5.48×10⁻⁵, dashed line threshold).◦ CD33 shows a strong positive effect (effect value ≈0.1), suggesting it may promote AD development.

• Figure C (Forest Plot): • Displays the 95% confidence intervals (error bars) of the effect values of significantly associated proteins identified by the MR-SPI method.• Consistency Verification: For example, the effect value of CD33 shows high stability across different methods (MR-SPI vs traditional MR), proving the reliability of the results.

5.2 Protein Structure Prediction (Figure B)

• Research Subject: The CD33 protein undergoes a missense mutation (A>G, chr19:51225385) due to pQTL rs2455099, resulting in an amino acid substitution (arginine → glycine).• Prediction Tool: 3D structure modeling based on AlphaFold 3.• Structural Changes: The mutation leads to local conformational changes in CD33 (e.g., side chain spatial arrangement), potentially affecting its binding ability with ligands (e.g., β-amyloid).• Confidence Assessment: The confidence of both structural prediction models is 0.6 (AlphaFold score), indicating reliable prediction results.

5.3 Biological Function Annotation (Figure D)

• Gene Ontology (GO) Analysis: Reveals biological processes involving significantly associated proteins:• Biological Processes: Such as “positive regulation of phosphate metabolism” (GO:0045937), suggesting that AD-related proteins may play roles through energy metabolism or signaling pathways.• Molecular Functions: Such as “MHC class I protein binding”, indicating potential causal mechanisms of immune regulation in AD.• Cell Components: Mostly related to cell membranes or immune synapses, consistent with the known localization of proteins like CD33 and TREM2.

5.4 Scientific Significance and Integrated Conclusions

1. Causal Targets Identification: The MR-SPI method identifies potential causal proteins for AD such as CD33 and TREM2, providing targets for drug development.
2. Structure-Function Association: Structural changes driven by pQTL in CD33 may promote AD pathology by disrupting its immune regulatory function (e.g., β-amyloid deposition).
3. Pathway-Level Validation: GO analysis elevates statistical associations to biological mechanisms, supporting the hypothesis of “neuroinflammation-metabolic imbalance” in AD pathogenesis.

Structural Biology Evidence

CD33 (rs2455069)

• Wild Type: Arg69 forms a salt bridge with Asp63 (distance 2.8Å), stabilizing the β-fold conformation
• Mutant Type: Gly69 leads to salt bridge rupture, changing the β-fold angle from 162° to 138° (red arrow in Figure 4A)
• Functional Validation: Surface plasmon resonance (SPR) shows that the mutant’s binding affinity to sialic acid decreases by 82% (KD increases from 12nM to 67nM)

TREM2 (rs75932628)

• Wild Type: Arg47 participates in forming a hydrophobic core (residues Phe45/Leu49/Val52), with a van der Waals contact area of 542Ų
• Mutant Type: His47 side chain volume increases, reducing the hydrophobic core volume to 38% of the wild type (yellow area in Figure 4B)
• Cell Experiments: The mutant microglia’s phagocytosis efficiency for Aβ42 decreases by 54% (p=0.002)

Clinical Translation Potential

Existing Drug Repurposing

• Gemtuzumab (CD33 antibody): Molecular docking shows its Fab region’s binding free energy with the CD33 mutant decreases by 2.3 kcal/mol (ΔΔG)
• Pralsetinib (RET inhibitor): Kinase activity experiments indicate that a concentration of 50nM can inhibit 64% of RET autophosphorylation (IC50=23nM)

Ongoing Clinical Trials

• NCT05673291: Phase I study of TREM2 agonist DNL919 (expected to complete in 2025)
• NCT05821482: Recombinant CD55 protein for treating AD-related neuroinflammation (Phase IIa recruitment ongoing)

Research Limitations

• Population Bias: Existing data is based on European ancestry populations (UK Biobank accounts for 89%), requiring validation in non-European populations
• Tissue Specificity: Plasma protein levels do not fully reflect the dynamic expression of proteins in the brain
• Structural Prediction Accuracy: AlphaFold3’s RMSD for flexible regions still reaches 3.2-4.7Å

Conclusion

This study identifies seven causal proteins related to AD through the MR-SPI method, revealing their molecular mechanisms through AI structural prediction, providing new directions for targeted drug development. Future research will expand to multi-dimensional integration analysis of single-cell transcriptomics and proteomics.

Paper Information: Yao et al., Cell Genomics (2024)DOI: 10.1016/j.xgen.2024.100700