Kolmogorov-Arnold Network for Gene Regulatory Network Inference

Tsz Pan Tong, Aoran Wang, George Panagopoulos, Jun Pang

Introduction

Understanding Gene Regulatory Networks (GRNs) is a central goal in systems biology, as these networks dictate how cells differentiate and respond to their environment. While single-cell RNA sequencing (scRNA-seq) has provided a wealth of data for this task, inferring the precise interactions between genes remains computationally and theoretically challenging. Current state-of-the-art methods, such as GENIE3 and GRNBOOST2, typically rely on decision tree ensembles. However, these models face two primary limitations: they represent biological processes as discontinuous "steps" and often fail to provide the directionality or sign (activation vs. inhibition) of regulatory relationships.

The scKAN Framework

To address these limitations, researchers have developed scKAN, a novel approach that utilizes Kolmogorov-Arnold Networks (KANs) for GRN inference. Unlike traditional Multi-Layer Perceptrons (MLPs) that use fixed activation functions at nodes, KANs employ learnable activation functions—specifically B-splines—on the edges of the network.

This architectural shift provides several key advantages:

  1. Differentiability: The model is third-order differentiable, allowing it to represent the "Waddington landscape" of gene expression as a smooth, continuous surface rather than a series of discrete boundaries.
  2. Explainability: By calculating the gradients of the learned functions, scKAN can determine the nature of a regulatory link. A positive gradient indicates gene activation, while a negative gradient indicates inhibition.
  3. Efficiency: KANs are often more parameter-efficient than MLPs, which helps prevent overfitting when dealing with the high-dimensional but often sparse data found in genomics.

scKAN operates using a One-vs-Rest (OvR) formulation, where the model is trained to predict the expression level of a single target gene based on the expression of all other genes in the dataset.

Experimental Validation

The performance of scKAN was evaluated using the BEELINE benchmark, a standardized framework for GRN inference. The testing suite included:

  • Synthetic Networks: Simulated dynamics including linear, cyclic, and bifurcating pathways generated via BoolODE.
  • Curated Networks: Real-world biological networks (e.g., HSC, GSD) derived from literature.
  • Scalability Tests: Larger networks with over 100 nodes to assess computational efficiency.

Key Results and Insights

scKAN demonstrated significant improvements over existing signed GRN inference models. In benchmarking tests, it outperformed the nearest competitors by 5.40% to 28.37% in AUROC and showed even more substantial gains in AUPRC, with improvements ranging from 1.97% to 40.45%. The model was particularly adept at capturing long-range dependencies in linear pathways.

Beyond raw accuracy, the model provides unique biological insights:

  • Cell-Type Specificity: Because scKAN generates a gradient for every individual cell, it can identify how regulatory logic shifts across different cell subtypes or during cellular maturation.
  • The Shadow Problem: The authors identified a fundamental challenge in OvR modeling termed the "Shadow Problem." In branching (furcating) biological lineages, the mapping of gene expressions can become ambiguous, similar to how a 3D shape's shadow can overlap itself when projected onto a 2D surface. This mathematical ambiguity remains a hurdle for function-based models in complex differentiation trajectories.

Conclusion and Future Directions

scKAN represents a major step forward in making GRN inference both more accurate and more interpretable. By moving away from discrete decision trees toward smooth, differentiable functions, it better captures the underlying biochemistry of gene regulation. Future iterations of the model aim to address the challenges of branching logic, incorporate temporal "pseudotime" data, and improve robustness against the "zero-inflation" noise common in experimental scRNA-seq data.

Summary updated

BibTeX
@inproceedings{tong2025kolmogorov,
  author = {Tsz Pan Tong and Aoran Wang and George Panagopoulos and Jun Pang},
  title = {{Kolmogorov-Arnold Network for Gene Regulatory Network Inference}},
  booktitle = {CMSB},
  year = {2025},
  doi = {10.1007/978-3-032-01436-8\_12},
}
Figure