How important are the genes to explain the outcome - the asymmetric Shapley value as an honest importance metric for high-dimensional features

This paper proposes using asymmetric Shapley values as a superior metric for quantifying the importance of high-dimensional genomic features in clinical prediction models, addressing limitations of traditional approaches by accounting for collinearity and known causal directions, and provides efficient algorithms validated through a colorectal cancer progression study.

The Gist

Here is an explanation of the paper, translated into everyday language using analogies to make the complex math feel like a story.

The Big Problem: Who Gets the Credit?

Imagine you are trying to predict how long a patient will survive after a cancer diagnosis. You have two types of information:

1. The "Big Picture" (Clinical Data): Things like the patient's age, gender, and how advanced the cancer is (Stage I, II, III, or IV).
2. The "Deep Dive" (Genomics): Thousands of tiny genetic signals inside the patient's cells.

The Old Way (The "Leave-One-Out" Mistake):
Traditionally, doctors and scientists tried to figure out which information was more important by playing a game of "What if?"

• Step 1: Build a model using Age, Gender, and Cancer Stage. It works okay.
• Step 2: Add the thousands of genes. Does the model get much better?
• Result: Often, the answer is "Not really." The model was already pretty good with just the clinical data.
• The Flaw: The scientists concluded, "Genes aren't important."

Why this is wrong: Imagine you are baking a cake. You have flour (Genes) and eggs (Clinical Data). If you already have eggs, adding flour doesn't make the cake twice as good; it just completes it. But if you take the eggs away, the cake falls apart. The old method was like saying, "Since the cake was already good with eggs, the flour didn't matter." It failed to realize that the eggs and flour were working together, and the eggs were hiding the flour's true value.

The New Solution: The "Asymmetric Shapley Value"

The authors propose a new way to measure importance called Asymmetric Shapley Values. To understand this, we need two concepts: The Team and The Order of Arrival.

1. The Team (Shapley Values)

In game theory, the "Shapley Value" is a way to fairly split the prize money among team members based on how much they contributed.

• If you have a team of 3 people, you don't just look at them individually. You look at every possible combination: Person A alone, Person A+B, Person A+C, Person A+B+C, etc.
• You calculate how much the team's score improves when a specific person joins the group.
• This solves the "collinearity" problem (where variables are correlated). It spreads the credit fairly between the eggs and the flour.

2. The Twist: Asymmetry (The "Root Cause" Rule)

The standard Shapley value treats everyone as equals. But in biology, order matters.

• Genes (G) are the root cause. They happen first.
• Disease State (D) (like the cancer stage) happens because of the genes.
• Outcome (Y) (survival) happens because of the disease.

Think of it like a Rube Goldberg machine (a complex chain reaction).

• The first ball (Genes) hits the second ball (Disease), which hits the third ball (Outcome).
• If you remove the first ball, the whole machine stops.
• If you remove the second ball, the machine stops, but only because the first ball had nowhere to go.

The "Asymmetric" Rule:
The authors say: "When we calculate credit, we must respect the order."

• You can't have the Disease (D) in the team without the Genes (G) being there first.
• If you try to calculate the importance of Disease without Genes, the math says, "That's impossible in this universe."

By forcing the math to respect this order, the Genes get more credit because they are the "root cause" that enables the Disease to exist. The Disease gets less credit because it's just a messenger carrying the Genes' message.

The Real-World Test: Colorectal Cancer

The authors tested this on real data from 845 colorectal cancer patients.

• The Setup: They tried to predict "Progression-Free Survival" (how long before the cancer comes back).
• The Variables:
  • Genes: 500 specific genes (High-dimensional).
  • Disease State: The tumor stage (Low-dimensional).
  • Confounders: Age, Gender, Tumor location.

The Results:

1. The Old Way: When they removed the genes, the model barely got worse. They thought, "Genes are useless."
2. The New Way (Asymmetric Shapley): When they used their new math, the story changed completely.
   • The Genes were revealed to be much more important than the Disease State.
   • The Disease State actually "stole" some of the Genes' credit in the old method.
   • The new method showed that the Genes are the true drivers, and the Disease State is just the middleman.

Why This Matters (The "Mediator" Metaphor)

Imagine a detective trying to solve a crime.

• The Genes are the Mastermind.
• The Disease State is the Messenger who delivers the order.
• The Outcome is the Crime.

If you only look at the Messenger, you might think the Messenger is the criminal. But if you use "Asymmetric Shapley," you realize the Messenger is just doing what the Mastermind told them to do. The Mastermind (Genes) is the one who deserves the "importance" rating.

The Technical Magic (How they did it)

Calculating this for thousands of genes is a nightmare for computers. It's like trying to count every possible way to arrange a deck of cards.

• The Problem: There are too many combinations.
• The Fix: The authors invented a "smart sampling" trick. Instead of checking every single combination, they used a statistical shortcut (Importance Sampling) to guess the answer with high accuracy, but much faster.
• The Dependency: They also figured out how to handle the fact that genes and disease states are mathematically linked, ensuring the "messengers" don't get blamed for the "mastermind's" actions.

The Bottom Line

This paper gives us a new, fairer ruler to measure what matters in medical predictions.

• Old Ruler: "If the model works without the genes, the genes don't matter." (False)
• New Ruler: "The genes are the root cause. Even if the model works without them because of other clues, the genes are still the most important piece of the puzzle."

This is crucial for precision medicine. It tells us that we shouldn't ignore our genes just because a doctor's quick glance at the tumor stage seems to explain everything. The genes are the silent engine driving the whole process.

Technical Summary

Here is a detailed technical summary of the paper "How important are the genes to explain the outcome - the asymmetric Shapley value as an honest importance metric for high-dimensional features" by van de Wiel et al.

1. Problem Statement

In clinical prediction models, researchers often combine high-dimensional genomic data (e.g., gene expression) with low-dimensional clinical variables (e.g., age, gender, disease state). A common approach to assess the importance of genomics is the "leave-one-out" method: removing genomic variables and measuring the drop in predictive performance.

The authors argue this approach is flawed for two main reasons:

1. Collinearity: If genomic variables are highly correlated with clinical variables, removing the genomics may result in a negligible performance drop because the clinical variables act as a "buffer." This leads to the premature conclusion that genomics are irrelevant, even if they are causally upstream.
2. Ignored Causality/Directionality: Standard feature importance metrics treat all variables symmetrically. However, in many biological contexts, there is a known causal ordering (e.g., Genes $\to$ Disease State $\to$ Outcome). Standard metrics fail to account for this mediation, often attributing the predictive power of the mediator (Disease State) back to the cause (Genes) or vice versa incorrectly.

2. Methodology

The paper proposes using Asymmetric Shapley Values to quantify feature importance in mixed-dimensional settings while respecting known causal structures.

Core Framework

• Asymmetric Shapley Values: Unlike standard (symmetric) Shapley values which consider all permutations of features, asymmetric values restrict the orderings to those respecting a known partial order (e.g., $G \to D$). This ensures that a feature (like Disease State) cannot be added to a coalition unless its upstream causes (Genes) are already present.
• Group Shapley Values: To handle high-dimensional genomic data ($G_1$) efficiently, the method extends to group Shapley values, treating the entire set of genes as a single feature block.
• Conditional Marginalization (SHAP): The authors utilize the SHAP (SHapley Additive Predictions) framework, specifically the conditional approach. Instead of assuming feature independence (marginal SHAP), they compute expectations conditional on the features already in the coalition ($E[f(x)|x_S]$). This is crucial for handling dependencies between genes and clinical variables.

Technical Innovations & Algorithms

To make this computationally feasible for high-dimensional data, the authors developed several extensions:

1. Efficient Weight Computation: They derived recursive formulas to calculate asymmetric Shapley weights ($w_h(S)$) for coalitions, avoiding the factorial complexity of enumerating all valid permutations.
2. Importance Sampling: For cases where the number of confounders makes exact enumeration impossible, they introduced a novel importance sampling scheme to approximate the weights proportional to the asymmetric distribution.
3. Dimension Reduction for Dependency Modeling: Computing conditional expectations for high-dimensional $G_1$ is difficult. The authors propose replacing $G_1$ with a low-dimensional summary (e.g., Principal Components) only for the dependency modeling step, while keeping the original $G_1$ for the prediction model.
4. D-Aware Summary: They introduce a specific summary score ($G_3$) derived by predicting the disease state ($D$) from genes ($G_1$). This captures the specific pathway of gene effects mediated by disease state.
5. Inference Framework: They propose two statistical tests to assess the global importance of features based on local Shapley values:
   • A semi-parametric likelihood-ratio test.
   • A non-parametric conditional independence test using matching and permutation (based on the coin R package).

3. Key Contributions

• Theoretical Extension: Adapting asymmetric Shapley values to mixed-dimensional settings (combining high-dimensional omics with low-dimensional clinical data) and grouped features.
• Computational Efficiency: Developing algorithms (recursive weights, importance sampling) that make asymmetric Shapley calculations tractable for large-scale genomic datasets.
• Mediation Handling: Demonstrating that asymmetric Shapley values correctly attribute the indirect effects of genes (via disease state) to the genes, whereas symmetric values dilute this importance.
• Software Implementation: The methods are implemented in the R package shapr and a dedicated GeneSHAP repository.

4. Results

The framework was applied to a colorectal cancer dataset ($N=845$) to predict progression-free survival.

• Comparison with Leave-One-Out: The traditional "leave-one-out" approach suggested genomic data had a very modest impact (C-index drop from 0.754 to 0.731).
• Symmetric vs. Asymmetric SAGE:
  • Symmetric Shapley Values: Assigned lower importance to Genes ($G$) and higher importance to Disease State ($D$), failing to capture the mediation.
  • Asymmetric Shapley Values: Significantly increased the importance of Genes ($G$) and decreased the importance of Disease State ($D$). The contribution of Genes rose from ~0.13 (symmetric) to ~0.17 (asymmetric) in the C-index decomposition.
• Mediation Effect: The analysis showed that for advanced disease stages (III and IV), the asymmetric Shapley values for genes were significantly higher than symmetric ones. This confirms that the model relies more heavily on genomic information when the disease state is advanced, a nuance captured only by the asymmetric approach.
• Inference: The proposed statistical tests confirmed that the increased importance of Genes in the asymmetric setting was statistically significant ($p < 0.001$).

5. Significance

• Honest Metric for Genomics: The paper provides a rigorous method to avoid underestimating the value of high-dimensional genomic data in clinical models, which is critical for justifying the cost of genomic sequencing.
• Causal Awareness without Full Causal Graphs: The method allows for "causally-aware" importance metrics using only a simple partial ordering (e.g., Genes precede Disease State), without requiring a fully specified causal graph or interventional data.
• Interpretability: By decomposing performance metrics (like C-index or $R^2$) into feature contributions that respect biological reality, the method offers a more interpretable view of how models make predictions.
• Generalizability: While focused on genomics, the framework is applicable to any mixed-dimensional prediction problem where feature dependencies and directional effects (mediation) are known.

In conclusion, the authors demonstrate that Asymmetric Shapley Values are a superior metric for feature importance in clinical genomics, effectively correcting the biases introduced by collinearity and ignoring causal directionality, thereby revealing the true predictive power of genomic features.