🧬 biology

A Multi-Omics Framework for Survival Mediation Analysis of High-Dimensional Proteogenomic Data

This article introduces SMAHP, a novel causal mediation framework for multi-omics data based on the accelerated failure time (AFT) model that effectively processes high-dimensional proteogenomic data to identify causal pathways influencing survival outcomes while demonstrating superior statistical power and better control of the false discovery rate compared to existing methods.

Original authors: Seungjun Ahn, Weijia Fu, Maaike van Gerwen, Lei Liu, Zhigang Li

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Seungjun Ahn, Weijia Fu, Maaike van Gerwen, Lei Liu, Zhigang Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to understand why a specific car (a patient) breaks down at a specific time (survival outcome). You have a massive garage full of data: thousands of blueprints (genes) and thousands of mechanical parts (proteins).

For a long time, scientists tried to understand the breakdown by looking either only at the blueprints or only at the parts. They might have asked: "Does this blueprint cause the car to break down?" or "Does this part cause the breakdown?" Yet in doing so, they overlooked the big picture: how the blueprint actually manufactures the part, which then causes the breakdown.

This work introduces a new tool called SMAHP (Survival Mediation Analysis of High-dimensional Proteogenomic data) to solve this puzzle. Here is how it works, simply explained:

1. The Problem: Too Much Noise, Too Many Connections

In the past, methods for analyzing this data were like trying to find a needle in a haystack while blindfolded.

The "Haystack": Modern technologies provide us with too much data (high-dimensional). We have thousands of genes and proteins to check.
The "Blindfold": Old methods usually looked at only one thing at a time (a gene or a protein) or assumed a very rigid relationship between them. They often missed the "mediator"—the protein that a gene produces, which then actually influences the patient's survival.
The "Broken Clock": Many old tools rely on a rule called the "Cox model." This rule assumes that the risk of breakdown remains constant over time. However, in real life, risks change. When this rule is violated, the old tools deliver false answers.

2. The Solution: SMAHP (The Clever Detective)

The authors developed SMAHP, a new statistical "detective" capable of handling the entire garage at once. It uses a three-step process to find the true culprits:

Step 1: The Big Sweep (Penalization)
Imagine you have 10,000 suspects (genes and proteins). You cannot interrogate them all. SMAHP uses a "filter" (called penalization) to quickly sweep the room and keep only the top suspects who seem most likely to be involved. It throws out the noise.
Step 2: The Connection Check (Screening)
Now it examines the remaining suspects to see who is actually talking to whom. It asks: "Does Gene A create Protein B, and does Protein B actually influence the car's breakdown?" It uses a technique called "Sure Independence Screening" to find the strongest connections between the blueprints and the parts.
Step 3: The Final Verdict (Testing)
Finally, it brings the best candidates to court. It uses a strict test to ensure these are not just lucky guesses. It controls for "false alarms" (False Discovery Rate) and ensures that if it says, "This gene causes this protein, which shortens lifespan," that statement is actually true.

The Secret Weapon: Unlike the old tools, SMAHP does not use the rule of the "broken clock" (Cox model). Instead, it uses an Accelerated Failure Time (AFT) model. Imagine this as a tool that measures how much time is cut off from a patient's life, rather than just estimating the risk at any arbitrary point in time. This is more flexible and accurate for complex diseases.

3. The Test Drive (Simulations)

Before applying this to real patients, the authors conducted a massive simulation. They created fake data with thousands of genes and proteins, along with known "truths" about who caused the breakdown.

The Result: SMAHP was like a high-performance sports car. It found the true causes (high power) and rarely made false accusations (low False Discovery Rate), even when the data was chaotic or the "breakdowns" were hard to detect (high censoring rates).
The Competition: Other methods either missed the true causes or cried "wolf" too often (marking innocent genes as guilty).

4. The Real-World Case: Head and Neck Cancer

The authors brought SMAHP into the real world by using data from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) regarding squamous cell carcinomas of the head and neck (HNSCC).

The Puzzle: They wanted to know how genes influence survival in patients who are negative for the HPV virus (a common cause of this cancer).
The Discovery: SMAHP found a specific chain reaction:
1. A gene named HMGB1P23 (a blueprint).
2. This gene influences a protein named LCE3E (a mechanical part).
3. This protein then influences how long the patient survives.
The Twist: Interestingly, the gene seemed to directly promote survival, while the protein it produced actually impaired survival. This is the kind of "hidden mediator" story that old methods would have completely overlooked.

Summary

Imagine SMAHP as a combination of a highly advanced translator and a detective. It takes the chaotic, massive language of thousands of genes and proteins, filters out the noise, and tells a clear story: "Gene A builds Protein B, and Protein B is the reason why the patient's survival time changes."

It is the first tool of its kind specifically designed to handle this massive amount of data while accurately measuring time-to-event outcomes without relying on outdated assumptions.

Technical Summary: SMAHP – A Multi-Omics Framework for Survival Mediation Analysis

Problem Statement
Survival analysis is critical for understanding time-to-event outcomes, such as disease progression and patient mortality. While recent advancements have enabled high-throughput profiling of proteomics and genomics data, existing causal mediation frameworks for survival outcomes face significant limitations. Current methods, such as High-dimensional Mediation Analysis in Survival Models (HIMAsurvival) and MASH, primarily rely on Cox proportional hazards (PH) regression. These approaches typically focus on a single exposure or individual omics layers, failing to capture the complex interplay between high-dimensional exposures (e.g., gene expressions) and high-dimensional mediators (e.g., protein levels) within a multi-omics context. Furthermore, reliance on the Cox model imposes the PH assumption, which may be violated in certain biological contexts. There is a lack of unified frameworks capable of simultaneously handling high-dimensional exposures and mediators while integrating multi-omics data for survival outcomes.

Methodology: SMAHP
The authors propose SMAHP (Survival Mediation Analysis of High-dimensional Proteogenomic data), a novel framework grounded in the counterfactual paradigm. The method is designed to identify causal pathways where genomic exposures influence survival outcomes through protein mediators. The procedure consists of three main steps:

Step 1: Penalized Variable Selection (Pre-screening):
- Outcome Model: A penalized Accelerated Failure Time (AFT) model is used to assess the direct effects of high-dimensional exposures ( $X$ ) and mediators ( $M$ ) on the survival outcome ( $T$ ). The AFT model is chosen to avoid the PH assumption inherent in Cox regression. The authors employ the Minimax Concave Penalty (MCP) via an alternating direction method of multipliers algorithm to identify active sets of significant genes ( $S_1$ ) and proteins ( $T$ ).
- Mediation Model: A penalized linear regression model (also using MCP) is fitted to model the association between exposures and each mediator ( $X \to M_k$ ), identifying active sets of genes ( $J^k_1$ ) associated with specific mediators.
Step 2: Screening Important Mediators (SIS):
- To reduce dimensionality and computational burden, the Sure Independence Screening (SIS) approach is applied. The method calculates the product of the estimated coefficients from the exposure-mediator and mediator-outcome paths ( $|\hat{\alpha} \cdot \hat{\beta}|$ ).
- Pairs of mediators and exposures with the top $n/\log(n)$ (or a multiple thereof) largest values are retained. This step identifies "important" proteomic mediators that account for exposures, effectively filtering the candidate sets to $S_2$ (mediators) and $J^s_2$ (exposures).
Step 3: Hypothesis Testing and FDR Control:
- A Joint Significance (JS) test is performed on the reduced sets to determine if a specific mediator lies in the causal pathway ( $H_0: \alpha \beta = 0$ ).
- The method calculates elementwise p-values for the product of coefficients using the maximum of the p-values from the individual paths.
- False Discovery Rate (FDR) is controlled using the Benjamini-Hochberg (BH) procedure on the resulting p-value matrix.

Key Contributions

Novel Framework: SMAHP is presented as one of the first attempts to integrate the Accelerated Failure Time (AFT) model into a multi-omics causal mediation framework for survival outcomes.
High-Dimensional Handling: Unlike previous methods that often treat exposures as single variables, SMAHP simultaneously handles high-dimensional vectors of both exposures (genes) and mediators (proteins).
Multi-Omics Integration: The framework explicitly models the synergy between genomic exposures and proteomic mediators, moving beyond single-omics layer analysis.
Robustness to PH Violation: By utilizing the AFT model, the method offers a robust alternative to Cox-based approaches when the proportional hazards assumption does not hold.
Implementation: The method is implemented as an R package (SMAHP) freely available on CRAN.

Results

Simulation Studies: The authors conducted extensive simulations across various scenarios (varying sample sizes $n$ $n$ , number of genes $p$ $p$ , and proteins $k$ $k$ ) and censoring rates (25%, 50%, 75%).
- Performance: SMAHP demonstrated high statistical power while effectively controlling the False Discovery Rate (FDR) at the 5% level.
- Comparison: Compared to a "Naïve" approach (no penalization/SIS) and a "SIS + SIS" approach (univariate marginal models with SIS), SMAHP maintained superior FDR control, particularly as censoring rates increased. The Naïve method exhibited inflated FDR, while the SIS + SIS method struggled to control FDR at high censoring rates (50% and 75%).
- Computation: While SMAHP required more computational time than the other methods due to penalization, it remained feasible (e.g., ~2-4 minutes for $n=400$ ).
Real-Data Application: The method was applied to Clinical Proteomic Tumor Analysis Consortium (CPTAC) data for Head and Neck Squamous Cell Carcinoma (HNSCC), specifically focusing on HPV-negative cases.
- Finding: The analysis identified a significant indirect effect where the protein LCE3E mediates the association between the gene HMGB1P23 and overall survival.
- Biological Context: While HMGB1P23 is not well-characterized in disease literature, LCE3E is associated with plantar warts (HPV-related), offering a potential biological link in an HPV-negative cohort where immune control of HPV might be relevant. The direct effect of the gene on survival was positive, while the indirect effect through the protein was negative.

Significance and Claims
The paper claims that SMAHP addresses critical gaps in survival mediation analysis by enabling the simultaneous analysis of high-dimensional multi-omics data without relying on the proportional hazards assumption. The authors position SMAHP as a tool that provides a more comprehensive view of molecular mechanisms (gene-protein-outcome pathways) compared to analyzing gene-outcome and protein-outcome relationships separately. The study concludes that the method offers a robust statistical framework for identifying causal pathways in complex biological systems, with successful application to real-world proteogenomic data. The authors acknowledge limitations, such as the need for larger sample sizes under very high censoring rates (75%) to maintain power and FDR control, and suggest future work could incorporate group-level biological information (pathway-level mediation) and nonparametric AFT models.