Estimating the genetic distance between subtypes of Major Depressive Disorder and their relationships with other traits using GDIS

This paper introduces GDIS, a novel method that estimates genetic distances and relationships between Major Depressive Disorder subtypes using only summary statistics from subtype-versus-control GWAS, thereby enabling direct subtype comparisons and revealing differential correlations with external traits to advance the understanding of genetic heterogeneity.

The Gist

The Big Problem: Depression is a "Mystery Box"

Imagine Major Depressive Disorder (MDD) as a giant, messy box of different fruits. Some people in the box have apples (sadness), some have oranges (anxiety), and some have bananas (sleep issues).

For a long time, scientists have been studying this box by comparing the entire box of fruit to a box of healthy vegetables (controls). They ask, "What makes this whole fruit box different from the vegetable box?"

While this helps, it misses a crucial detail: Not all fruit is the same. A person with "anxiety-depression" might have a completely different genetic recipe than a person with "sleep-depression." If we treat them all the same, we might miss the specific ingredients needed to cure them.

The Missing Comparison

Usually, scientists compare:

1. Fruit A vs. Vegetables
2. Fruit B vs. Vegetables

But they rarely compare Fruit A directly to Fruit B. Why? Because it's mathematically tricky to measure the "distance" between two types of fruit without a common ruler. It's like trying to measure the distance between two cities when one map uses miles and the other uses kilometers.

The Solution: GDIS (The Genetic GPS)

The authors of this paper invented a new tool called GDIS (Genetic DIstance of disorder Subtypes). Think of GDIS as a Genetic GPS that draws a map of how different types of depression relate to each other.

Instead of just listing numbers, GDIS turns complex genetic data into a geometric shape (a triangle or a 3D structure). Here is how it works:

1. The Ruler (Genetic Distance)

Imagine genetic "distance" is like physical distance.

• If two groups of people are genetically very similar, they stand close together.
• If they are very different, they stand far apart.
• GDIS creates a ruler that works for everyone, regardless of how common the depression is in the population. It standardizes the map so we can compare apples to oranges fairly.

2. The Angles (Genetic Correlation)

Imagine the groups are people standing in a field holding hands.

• If two groups are genetically identical, they stand in a straight line (0 degrees apart).
• If they are completely unrelated, they stand at a right angle (90 degrees apart).
• GDIS measures the angle between them. A sharp angle means they share very different genetic causes. A flat angle means they share the same causes.

What Did They Discover?

The researchers applied this GPS to 7 different types of depression (like depression with childhood trauma, depression with anxiety, depression with sleep issues, etc.).

Here are the big takeaways:

• The "Childhood Trauma" Divide: The biggest genetic gap was found between people with depression who had childhood trauma and those who didn't. It's as if these two groups are standing on opposite sides of a canyon. They share very little genetic DNA related to their depression. This suggests they might need very different treatments.
• The "Recurrent Episodes" Clue: For people who get depressed over and over again (recurrent), the genetic difference was mostly about intensity, not type. It's like the same recipe, but with double the sugar. The genes are the same, but they are turned up louder.
• The "Sleep/Weight" Connection: Depression involving sleep or weight gain was genetically very close to BMI (Body Mass Index). This makes sense intuitively, but GDIS proved it with hard numbers, showing these two traits are genetically linked.

Why This Matters (The "Aha!" Moment)

Imagine you are a doctor trying to treat a patient.

• Old Way: You look at the patient's depression and say, "Okay, you have depression. Here is a standard pill."
• New Way (with GDIS): You look at the patient's specific "subtype" on the GDIS map. You see, "Ah, this patient's depression is genetically close to 'Childhood Trauma' and far from 'Recurrent Episodes'."

This tells you that this patient's depression might be rooted in how their brain handles stress from the past, rather than just a chemical imbalance that happens repeatedly. This could lead to personalized medicine—giving the right treatment to the right person based on their genetic "map."

The Bottom Line

This paper didn't just find new genes; it built a new way to visualize how depression works. By turning complex math into a simple geometric map, GDIS helps scientists see that depression isn't just one big lump of "sadness." It's a collection of different genetic stories, and understanding the distance between those stories is the key to better, more personalized care.

Technical Summary

1. Problem Statement

Major Depressive Disorder (MDD) is highly heterogeneous, with patients presenting diverse symptom constellations. While Genome-Wide Association Studies (GWAS) have successfully identified genetic variants associated with MDD, standard approaches typically contrast broad MDD cases against healthy controls. This masks genetic heterogeneity.

• Current Limitation: Existing research investigates subtypes (e.g., MDD with vs. without anxiety) by comparing each subtype separately against controls. However, the most critical comparison for understanding genetic distinctiveness—directly comparing Subtype 1 cases against Subtype 2 cases—is often neglected.
• The Barrier: Direct subtype-vs-subtype comparisons are difficult to interpret because standard heritability metrics (observed or liability scale) are not directly comparable across different ascertainment schemes or population prevalences. Furthermore, collecting individual-level data for direct case-vs-case GWAS is often impractical.
• Goal: To develop a method that benchmarks the genetic distance between disorder subtypes and visualizes their relationships with external traits using only summary statistics.

2. Methodology: The GDIS Framework

The authors introduce GDIS (Genetic DIstance of disorder Subtypes), a novel computational framework that transforms genetic statistics into a geometric representation.

Core Concepts

GDIS maps genetic parameters onto a Euclidean space:

1. Genetic Distance ($D$): Defined as the square root of the SNP-heritability ($h^2$) on the observed scale with 50/50 ascertainment.
   • Rationale: This scale provides a unified metric independent of population prevalence, allowing direct comparison between subtype-vs-control and subtype-vs-subtype comparisons.
   • Formula: $D = \sqrt{h^2_{50/50}}$
2. Genetic Angle ($\theta$): Defined as the inverse cosine of the genetic correlation ($r_g$) between two GWAS comparisons.
   • Rationale: A correlation of 1 implies an angle of 0° (identical genetic architecture), while 0 implies 90° (orthogonal/independent).
   • Formula: $\theta = \arccos(r_g)$

Workflow Steps

GDIS requires only two input GWAS summary statistics: Subtype 1 vs. Controls and Subtype 2 vs. Controls. It performs the following steps:

1. Input Processing: Uses LD Score Regression (LDSC) to estimate heritabilities ($h^2$) and genetic correlations ($r_g$) for the two input comparisons.
2. Geometric Construction (Triangle):
   • Constructs a triangle where vertices represent Subtype 1, Subtype 2, and Controls.
   • Side lengths are the genetic distances ($\sqrt{h^2}$) from each subtype to controls.
   • The angle at the "Controls" vertex is derived from the $r_g$ between the two subtype-vs-control GWASs.
3. Derivation of Missing Metrics: Using the Law of Cosines, GDIS calculates:
   • The genetic distance between Subtype 1 vs. Subtype 2 (the side opposite the control vertex).
   • The genetic correlations between the direct subtype comparison and the controls.
   • The position of the "All-Cases" group (the union of both subtypes) along the line connecting the two subtypes, weighted by their prevalence.
4. Extension to External Traits:
   • GDIS extends to 3D geometric space to include an external trait (e.g., BMI, Schizophrenia).
   • It calculates the angles between the external trait vector and the subtype vectors, allowing for the derivation of how the external trait differentially correlates with the two subtypes.
5. Validation & Filtering:
   • The method includes filters to ensure geometric validity (e.g., checking if triangle inequalities hold for correlations).
   • It validates derived values against empirical LDSC estimates from direct GWASs (where available).

3. Key Contributions

• Novel Metric: Introduces a generalizable distance metric (square root of 50/50 heritability) that allows for the direct comparison of genetic architectures across different ascertainment schemes.
• Geometric Visualization: Transforms complex statistical outputs (heritabilities and correlations) into intuitive 2D and 3D geometric plots, making it easier to visualize the "distance" and "angle" between psychiatric subtypes.
• Efficiency: Eliminates the need for individual-level data or direct case-vs-case GWAS. It infers the genetic relationship between subtypes solely from two standard case-vs-control GWASs.
• Differentiation of Effects: Distinguishes between quantitative differences (same genetic variants, different effect sizes) and qualitative differences (different sets of genetic variants) between subtypes.

4. Results

The authors applied GDIS to seven MDD subtypes using UK Biobank data ($N \approx 113,000$):

1. Validation: GDIS-derived values for genetic distances and correlations matched closely with empirical LDSC estimates (average absolute difference for heritability: 0.002; for correlations: 0.013).
2. Subtype Distinctiveness:
   • Childhood Trauma: Showed the greatest genetic distance between subtypes ($D^2 = 0.105$), even exceeding the distance between all MDD cases and controls. This suggests a qualitative genetic difference (distinct genetic architecture) between those with and without childhood trauma.
   • Recurrent Episodes: Showed a small genetic distance ($D^2 = 0.018$) and a genetic correlation near 1. This indicates a quantitative difference (same risk alleles, but larger effect sizes in the recurrent group).
   • High Severity: Demonstrated significant genetic differentiation, driven by both quantitative and qualitative factors.
3. External Trait Analysis:
   • GDIS revealed that external traits could differentiate subtypes even when the trait's correlation with each subtype individually appeared similar.
   • Example: The "Internalizing Factor" showed high correlations with both high-severity and non-high-severity MDD individually. However, the direct comparison (Subtype 1 vs. Subtype 2) showed a strong correlation with the internalizing factor ($r_g = 0.854$), proving it is a distinguishing trait.
   • Example: BMI was strongly correlated with "MDD with sleep/weight increase" but not with the "without" group, a distinction clearly visualized in the 3D geometry.
4. Subtype Definition Comparison: When comparing different definitions of MDD (e.g., anxiety-based vs. trauma-based), GDIS showed they are genetically divergent, forming cross-shaped geometric representations rather than overlapping clusters.

5. Significance and Implications

• Clinical Stratification: GDIS provides a tool to identify which MDD subtypes are genetically distinct enough to warrant separate clinical management or targeted therapies.
• Etiological Insight: By distinguishing between quantitative and qualitative genetic differences, GDIS helps researchers understand whether a subtype represents a more severe form of the same disorder or a fundamentally different biological entity.
• Personalized Medicine: The ability to benchmark subtypes against external traits (like BMI or educational attainment) helps in predicting comorbidities and tailoring treatment plans based on genetic profiles.
• Accessibility: The method is implemented as a freely available tool (Shiny app and R package) that requires only summary statistics, making it accessible to the broader research community without the need for massive individual-level datasets.

Limitations Noted:

• Requires 50/50 ascertainment heritability, which is less common in literature than liability scale.
• Currently limited to binary splits of cases (cannot handle three or more subgroups simultaneously).
• Visualizations are restricted to 3D space, limiting the number of traits/subtypes that can be plotted at once.
• Based primarily on European ancestry data.