🧬 biology

Topology-based Physical Reconstruction of Biological Networks

This paper proposes a reproducible, topology-based method that reconstructs the physical spatial coordinates of biological networks by minimizing effective potential energy, thereby transforming topological data from sources like STRING into a metric space suitable for quantitative physical and mathematical analysis.

Original authors: Iryna Oliynyk

Published 2026-08-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Iryna Oliynyk

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

In the vast, invisible machinery of life, cells do not operate as isolated islands. Instead, they function through intricate webs of communication, where proteins, genes, and molecules constantly reach out to one another to coordinate complex tasks. Scientists have long mapped these connections, creating digital lists that tell us which parts interact with which. These maps, known as biological networks, are powerful tools for understanding how life works, from the spread of a virus to the regulation of a heartbeat. However, for decades, these maps have been flat. They show the links between the parts, but they lack a third dimension: space. They tell us that two proteins are connected, but they do not tell us how far apart they sit or how they are arranged relative to one another in a physical sense. Without this spatial information, researchers have been unable to apply the laws of physics and geometry to these systems, treating them as abstract lists rather than tangible structures that occupy a coordinate space.

A researcher at Lviv Medical University has now developed a way to fill this gap, turning these flat lists into three-dimensional, physical models without needing any prior knowledge of where the parts actually sit. The work focuses on a specific network of proteins involved in the body's response to COVID-19 and pneumonia, a system containing 51 distinct proteins and 412 connections between them. The challenge was to take a database that only knew which proteins were friends and to figure out where those friends should stand in a room to make the most sense of their relationships. The researcher treated the network not as a drawing, but as a mechanical system. Imagine each protein as a small ball and every connection between them as a spring. In this model, proteins that are directly connected pull toward each other, while all proteins, regardless of whether they are connected, push away from one another to avoid crowding.

By running a computer simulation that mimics the laws of physics, the researcher let this system of balls and springs settle into a state of rest. The computer calculated the forces acting on every single protein, moving them slightly in each step until the entire system stopped jiggling and found a stable, balanced shape. This process, which took about 250 steps to stabilize, resulted in a unique arrangement where every protein was assigned a specific set of coordinates, effectively giving the network a physical address. The result was a geometric map that preserved the exact same connections as the original data but added a layer of spatial reality. In this new map, the most important proteins, those with the most connections to others, naturally settled near the center of the cluster, while the less connected proteins drifted to the edges. This arrangement was not random; it was the direct mathematical consequence of the system seeking the lowest possible energy state.

To ensure this method was reliable and not just a lucky accident of the computer's starting point, the researcher ran the simulation thirty separate times, each time beginning with the proteins in a completely different, random arrangement. Despite these different starting points, the final shapes were remarkably consistent. When the researcher compared the results, the relative positions of the proteins remained the same, with a high degree of agreement in how far each protein sat from the center. This proved that the spatial arrangement was a stable property of the network itself, not a fluke of the calculation. The study confirmed that the new map was a perfect mirror of the original data in terms of connections, with no links added or lost, but it added something entirely new: a reproducible, quantitative description of the network's geometry.

This approach changes how scientists can study biological systems. Previously, the spatial layout of such a network was merely a visual aid, a way to make a chart look nice on a screen. Now, the coordinates generated by this method are treated as a real, independent dataset. Researchers can measure the distance between any two proteins, calculate the density of connections in a specific area, or apply mathematical tools that require a defined space. The study makes it clear that these coordinates do not represent the actual physical location of proteins inside a human cell, which is a much more complex and crowded environment. Instead, they represent a mathematically consistent structure derived purely from the pattern of interactions. By turning a list of relationships into a physical model, this work opens the door to analyzing biological networks with the same rigorous tools used in physics and engineering, allowing scientists to ask new questions about the shape and structure of life's most complex systems.

Technical Summary: Topology-Based Physical Reconstruction of Biological Networks

Problem Statement
Biological networks, such as protein–protein association networks derived from databases like STRING, are fundamental tools in systems biology and biophysics. However, raw data from these sources typically consists only of topological information (nodes, edges, and interaction scores) without spatial coordinates. This absence of a metric space prevents the direct application of physical and mathematical methods that rely on geometric distances, such as analyzing inter-node distances, local density, or equilibrium configurations. While spatial embedding algorithms exist, they are predominantly used for visualization rather than as a source of reproducible, quantitative physical data. Consequently, there is a gap in the ability to treat the spatial arrangement of biological networks as an independent, mathematically rigorous dataset derived solely from topology.

Methodology
The paper proposes a reproducible approach to physically reconstruct biological networks by treating the graph $G=(V, E)$ as a many-body mechanical system. The reconstruction is framed as a physical-mathematical problem: finding the equilibrium configuration of interacting elements by minimizing the system's effective potential energy.

Physical Model: The network is modeled using force-directed principles. Vertices are treated as material points, and edges as interactions. The system evolves under two competing forces:
1. Repulsive Force ( $F_r$ ): A mutual repulsion acting between all pairs of vertices (analogous to Coulomb repulsion) to prevent overlap.
2. Attractive Force ( $F_a$ ): An elastic attractive force acting only between connected vertices, proportional to the edge weight ( $\omega_{ij}$ ).
Algorithm: The study utilizes the Fruchterman–Reingold algorithm to numerically relax the system. The process involves iteratively calculating the resultant force on each vertex and updating coordinates until the system reaches a state of mechanical equilibrium where the total potential energy is minimized ( $U \to min$ ).
Normalization: To ensure the coordinates are independent of arbitrary initial conditions, the resulting configuration undergoes geometric normalization. This involves shifting the center of mass to the origin and scaling coordinates relative to the maximum distance from the center, rendering the representation invariant to translation, rotation, and scaling.
Validation Protocol: The approach was tested on a protein–protein interaction network (51 nodes, 412 edges) derived from the STRING database for the query 'COVID-19 + pneumonia'. Validation included:
- Convergence Analysis: Monitoring the Spearman correlation of pairwise distance matrices over 500 iterations.
- Robustness Testing: Running 30 independent reconstructions with different random seeds to assess the reproducibility of the radial ranking of nodes.
- Topological Verification: Confirming graph isomorphism between the reconstructed network and the original STRING data.

Key Results

Topological Preservation: The reconstruction process fully preserved the original network topology. The reconstructed graph was isomorphic to the source data, with complete agreement in the number of nodes (51), edges (412), and local connection structures.
Convergence: The relaxation process reached a stable equilibrium configuration after approximately 250 iterations, with the correlation between intermediate distance matrices and the final configuration exceeding 0.98.
Reproducibility: Thirty independent reconstructions demonstrated high reproducibility of geometric characteristics. The median Spearman correlation coefficient for the radial ranking of nodes (distance from the geometric center) was $R = 0.944$ .
Structural Correlations: A strong negative correlation was observed between a node's degree (connectivity) and its distance from the geometric center ( $r = -0.8949$ ). Similarly, a significant negative correlation existed between node degree and the average length of incident edges ( $r = -0.7848$ ). This indicates that highly connected "hub" proteins naturally cluster near the center of the equilibrium configuration, while less connected nodes localize to the periphery.

Significance and Claims
The paper claims that the primary contribution is not the specific algorithm used (Fruchterman–Reingold), but the formulation of the reconstruction problem as the minimization of effective potential energy. This approach transforms the spatial coordinates of network nodes from a mere visualization aid into an independent, reproducible quantitative dataset.

The author emphasizes that:

Metric Space Generation: The method generates a metric Euclidean space where inter-node distances are defined, enabling physical and mathematical analyses (e.g., density, distance-based metrics) that are impossible with purely topological data.
Stability of Geometric Properties: While the absolute coordinates depend on the specific force model and random initialization, stable geometric characteristics—specifically the radial ordering of nodes and the correlation between connectivity and centrality—are reproducible and invariant.
Distinction from Visualization: Unlike traditional spatial embeddings used for illustration, this approach treats the resulting coordinates as a valid physical characteristic of the network, suitable for further mathematical modeling.

The paper maintains a modest scope, acknowledging that the reconstructed coordinates do not represent the actual physical arrangement of proteins within a cell. Instead, they represent a mathematically consistent spatial representation derived strictly from the network's topological structure. The study validates this approach on a single, moderate-sized network, noting that generalization to significantly larger or sparser networks requires further verification.

Technical Summary: Topology-Based Physical Reconstruction of Biological Networks

More like this