On why and how to encode probability distributions on graph representations of omics data: enhancing predictive tasks and knowledge discovery

This paper proposes a novel graph-based framework that integrates structured statistical distributions into omics data representations to achieve competitive predictive performance and enhanced biological interpretability for identifying regulatory modules across multiple cancer types.

The Gist

Imagine you are trying to understand a massive, chaotic city (the human body) by looking at its traffic patterns (biological data like genes and proteins).

For a long time, scientists have tried to predict what will happen in this city—like whether a traffic jam will lead to an accident (disease) or if a specific route will get you home safely (survival). Traditionally, they've used two main tools:

1. The "Average" Approach: They look at the average speed of cars on a road. If the average speed drops, they assume there's a problem. But this misses the nuance: maybe the road is fine for 90% of the day, but completely gridlocked for 10%. The average hides the danger.
2. The "Connection" Approach: They draw a map showing which roads connect to which. But often, they just draw a line and say, "Road A connects to Road B." They don't record how heavy the traffic is, or if that connection changes depending on the time of day or the weather.

The Problem

The authors of this paper argue that these traditional maps are missing the most important part of the story: the probability. In biology, things aren't just "on" or "off." They are messy, fluctuating, and full of chance. A gene might be active in a healthy person 20% of the time and in a sick person 80% of the time. Traditional graphs often ignore this "fuzziness" and just give a single number.

The Solution: The "Weather Map" Graph

The team proposes a new way to draw these maps. Instead of just drawing a line between two points (like two genes), they attach a mini weather forecast to every road and every intersection.

• The Nodes (Intersections): Instead of just saying "Gene A is here," the map says, "Gene A usually behaves like this in healthy people, but behaves like that in sick people." It stores the entire history of how that gene behaves, not just a single average.
• The Edges (Roads): Instead of just saying "Gene A talks to Gene B," the map records how their conversation changes. "When Gene A is loud, Gene B usually whispers in healthy people, but screams in sick people."

They call this "encoding probability distributions." Think of it like upgrading from a static, black-and-white street map to a dynamic, 3D hologram that shows traffic density, weather conditions, and accident risks for every single street, all at once.

How It Works (The Recipe)

1. Gather the Data: They took data from five different types of cancer (like colon, kidney, and lung cancer) from a massive public database (TCGA).
2. Build the "Weather" Map: They didn't just look at the numbers; they calculated the shape of the data. If a gene's activity looks like a bell curve in healthy people but a jagged spike in sick people, the graph captures that shape.
3. Prune the Noise: Just like a gardener trimming dead branches, they cut away the connections that didn't show a clear difference between healthy and sick people. This leaves a clean, sharp map of the most important biological relationships.
4. Predict the Future: When a new patient comes in, the system doesn't just compare their numbers to an average. It asks: "Does this patient's 'weather pattern' look more like the 'sick' forecast or the 'healthy' forecast?"

The Results: Why It Matters

The team tested this new "Weather Map" against the best standard computer programs (Machine Learning) used today.

• The Score: The new method performed just as well as the best existing tools at predicting if a patient would survive or what type of tumor they had.
• The Superpower: The real win wasn't just the score; it was understanding. Because the map kept all the statistical details, the scientists could look at it and say, "Ah! We found a specific group of proteins that act like a 'hub' in the city. When these hubs get clogged, the whole system crashes."

They found specific genes (like BRD4 and WEE1) that act as major traffic controllers. By looking at the "weather" of these specific genes, they could identify biological modules (groups of proteins working together) that drive the disease.

The Takeaway

Think of this paper as a shift from taking a snapshot of a city to watching a live, 4D simulation of it.

• Old Way: "The average speed on Main Street is 30 mph." (Good for a quick guess, bad for understanding why accidents happen).
• New Way: "Main Street has a 90% chance of gridlock between 5 PM and 6 PM, but only if it's raining, and this pattern is different in the North District than the South."

By keeping the full "story" of the data (the probabilities) inside the graph, this method helps doctors and scientists not only predict outcomes more accurately but also understand the "why" behind the disease, leading to better treatments and discoveries.

Technical Summary

1. Problem Statement

The rapid growth of multi-omics data (genomics, transcriptomics, proteomics) has created a need for algorithms that can model complex molecular systems. While graph-based learning methods are effective at representing biological interactions, traditional approaches often suffer from two main limitations:

1. Loss of Statistical Information: Conventional graphs typically treat nodes (biological entities like genes) and edges (interactions) as static values or simple weights, ignoring the inherent stochastic nature and statistical distributions of the underlying molecular features.
2. Oversimplification of Interdependencies: Standard methods often rely on aggregated statistics or correlation analyses, failing to capture the holistic, probabilistic relationships between biological entities and their impact on specific phenotypes (e.g., cancer survival or tumor subtype).

The authors argue that to improve both predictive accuracy and biological interpretability, graph representations must encode the full probability distributions of features and their relationships, conditioned on clinical outcomes.

2. Methodology

The proposed framework introduces a novel probabilistic graph representation where nodes and edges are annotated with class-conditional probability distributions. The methodology consists of three main phases:

A. Graph Generation

• Node Creation: Each feature (e.g., gene expression level) is mapped to a node. Instead of a single scalar value, the node stores an empirical probability distribution function (PDF for continuous data, PMF for categorical data).
  • In supervised settings, these are class-conditional distributions ($f_{X|c}$), estimated separately for each target class (e.g., "Alive" vs. "Deceased").
• Edge Creation: Edges represent relationships between pairs of features. The authors use a log-ratio transform ($\log(\frac{x_a + \delta}{x_b + \delta})$) to capture the associative strength between features.
  • Similar to nodes, edges are annotated with class-conditional distributions of these log-ratios.
  • Statistical Pruning: To ensure the graph captures meaningful biological signals, edges are filtered using statistical tests (Kolmogorov-Smirnov test). Only edges where the distributions between classes differ significantly (p-value < $\alpha$) are retained.

B. Graph Weighting

• Nodes and edges are weighted based on the dissimilarity between their class-conditional distributions.
• The Kolmogorov-Smirnov (KS) statistic is used as a score to quantify the discriminative power of a feature or feature-pair. Higher KS scores indicate greater ability to distinguish between classes.

C. Prediction (Instance-Specific Graphs)

• For a new testing instance, the model constructs a specific graph by evaluating the pre-computed class-conditional distributions at the instance's feature values.
• Likelihood Estimation: The model computes likelihood scores for each class by aggregating the weighted contributions of nodes and edges.
  • $Likelihood_c = \alpha \cdot g_{nodes}(W_V, c) + \beta \cdot g_{edges}(W_E, c)$
• Kernel Density Estimation (KDE): To prevent overfitting and handle sparse data, KDE (with Gaussian kernels) is applied to smooth the empirical distributions before likelihood calculation.
• Classification: Final class probabilities are derived using the Softmax function, and the class with the highest posterior probability is selected.

3. Key Contributions

1. Novel Graph Representation: A framework that encodes structured statistical distributions (PDFs/PMFs) directly onto graph nodes and edges, moving beyond simple scalar weights.
2. Robust Predictive Models: New algorithms that leverage these probabilistic graphs, demonstrating effectiveness even with limited sample sizes and highly imbalanced target distributions.
3. Enhanced Interpretability: The framework facilitates knowledge discovery by identifying regulatory modules (high-degree nodes, cliques, k-cores) that are statistically significant and biologically relevant to specific clinical outcomes.
4. Comprehensive Validation: Extensive empirical testing on The Cancer Genome Atlas (TCGA) data across five cancer types (COAD, KIRC, LGG, LUAD, OV) and three omic layers (mRNA, miRNA, Protein).

4. Results

The study evaluated the method against established machine learning baselines (Random Forest, Logistic Regression, Naive Bayes) using 5-fold cross-validation.

• Predictive Performance:
  • The graph-based approach achieved performance comparable to or superior to the best ML baselines across most tasks.
  • Vital Status Prediction: In the KIRC (Kidney) and LGG (Low-Grade Glioma) datasets, the graph-based method achieved statistically significant improvements in Accuracy and Precision, particularly when using miRNA data.
  • Tumor Site Prediction: The method showed strong performance in COAD and KIRC, often outperforming baselines in F1-score and Accuracy.
  • Imbalanced Data: The approach demonstrated robustness in handling imbalanced datasets (e.g., rare cancer subtypes), where traditional ML models often struggled.
• Knowledge Discovery (Case Study: LGG Proteomics):
  • Degree Analysis: Identified "hub" proteins (e.g., BRD4, WEE1, IGFBP2) with high connectivity. Gene enrichment analysis confirmed these hubs are involved in critical pathways like the MAPK cascade and Glioma development.
  • Discriminative Power: Statistical t-tests confirmed that proteins identified as high-degree nodes were highly discriminative between patient classes.
  • k-Core Analysis: The analysis of k-cores (densely connected subgraphs) revealed functional modules of proteins that consistently exhibit statistical differences across phenotypes, suggesting their role as regulatory units.

5. Significance

This work bridges the gap between statistical learning and graph theory in the context of omics data.

• Interpretability: Unlike "black-box" deep learning models, this approach provides a transparent mechanism to trace predictions back to specific statistical distributions and biological interactions, facilitating the discovery of biomarkers and regulatory modules.
• Handling Complexity: By modeling the full distribution of data rather than just means or correlations, the method captures the inherent stochasticity of biological systems, making it particularly suitable for complex, noisy omics datasets.
• Clinical Utility: The ability to perform competitive prediction on vital status and tumor subtyping, combined with the generation of biologically interpretable graphs, offers a powerful tool for precision medicine and hypothesis generation in cancer research.

In conclusion, the authors demonstrate that incorporating structured statistical information into graph representations creates a competitive, interpretable, and robust framework for predictive modeling and knowledge discovery in complex diseases.