Abstract
Graphical models find numerous applications in biology, chemistry, sociology, and neuroscience. While substantial progress has been made in graph estimation, it remains largely unexplored how to select significant graph signals with uncertainty assessment, especially those graph features related to topological structures including cycles (i.e., wreaths), cliques, and hubs. These features play a vital role in protein substructure analysis, drug molecular design, and brain network connectivity analysis. To fill this gap, we propose a novel inferential framework for general high-dimensional graphical models to select graph features with false discovery rate (FDR) controlled. Our method is based on the maximum of $p$-values from single edges that comprise the topological feature of interest, allowing for the detection of weak signals. Moreover, we introduce the $K$-dimensional persistent Homology Adaptive selectioN (KHAN) algorithm to select all homological features within $K$ dimensions with uniform control of the FDR over continuous filtration levels. The KHAN method applies a novel discrete Gram-Schmidt algorithm to select statistically significant generators from the homology group. We apply this structural screening method to identify important residues of the SARS-CoV-2 spike protein during the binding process to ACE2 receptors, identifying residues crucial for protein conformational changes and potential targets for inhibition.
Introduction
In fields like neuroscience and proteomics, data is frequently modeled as graphs. While existing literature provides robust methods for estimating whether an individual edge exists between two nodes, the field lacks rigorous "uncertainty assessment" for larger, more complex structures. Identifying a cycle (a "wreath") or a tightly knit group (a clique) in a brain network is one thing; determining if that structure is statistically significant or merely an artifact of noise is another.
The Challenge of Topological Inference
Selecting significant graph features involves overcoming two primary hurdles:
- Probabilistic Dependency: Edges in a graphical model are rarely independent; a signal in one often correlates with another, complicating standard statistical tests.
- Algebraic Dependency: In topology, features are often redundant. For instance, a large cycle might simply be the algebraic sum of two smaller ones.
Existing methods often rely on computationally expensive bootstrapping or conservative maximal statistics that struggle to detect weak signals. The "Wreaths of KHAN" framework bridges this gap with a method that is both computationally efficient and statistically sensitive.
Methodology
The proposed framework introduces three primary innovations to handle high-dimensional graphical models, such as Gaussian Graphical Models (GGM) for continuous data and Ising Models for discrete data.
Feature-wise p-values and the "Weakest Link"
To test the reality of a complex feature, the framework evaluates the "weakest link"—the maximum $p$-value among all edges comprising that feature. If the least significant edge in a structure remains statistically significant, the entire structure is likely real. This approach allows for the detection of weaker signals than methods relying on average or total structure strength.
The KHAN Algorithm and Persistent Homology
The $K$-dimensional persistent Homology Adaptive selectioN (KHAN) algorithm handles features across "filtration levels," which represent varying thresholds of connectivity. By utilizing persistent homology, the method identifies which features "persist" as the graph's connectivity threshold changes, providing a multi-scale view of the data.
Discrete Gram-Schmidt (DGS)
The "secret sauce" of the KHAN algorithm is the Discrete Gram-Schmidt process. Similar to the Gram-Schmidt process in linear algebra that finds an orthogonal basis, DGS iteratively selects the most significant (lowest $p$-value) homological generators that are linearly independent of those already chosen. This prevents the "double counting" of topological features and ensures algebraic independence.
Experimental Results
Statistical Rigor in Simulations
In synthetic simulations using graphs with ground-truth structures (triangles, four-cycles, etc.), the empirical FDR was consistently maintained below nominal levels (e.g., 0.05 or 0.1). Furthermore, the statistical "power"—the ability to find true features—remained high even as the number of nodes in the network increased.
Biological Discovery: SARS-CoV-2 Spike Protein
The method was applied to trajectory data of the SARS-CoV-2 spike protein in its closed, partially open, and fully open states. The model successfully identified known critical residues (such as A570 and I569) that facilitate binding to human ACE2 receptors.
More significantly, the method identified novel residues in the "linker" and "S2" domains. These residues appear to act as a pivotal hinge, facilitating the transition of the receptor-binding domain (RBD) from a "down" to an "up" position. These findings highlight potential new targets for drug inhibition to prevent viral entry.
Conclusion and Future Work
The Wreaths of KHAN provides a rigorous framework for identifying significant topological features with uniform FDR control. While the method is significantly faster than bootstrap-based approaches, calculating high-dimensional homology in graphs with millions of nodes remains a computational challenge. Future work may involve extending these theoretical guarantees to non-parametric graphical models and exploring even higher-dimensional polytopes in biological manifolds to further unlock the secrets of protein folding and dynamics.
Summary updated