Abstract
Gene regulatory network (GRN) inference serves as a cornerstone for deciphering cellular decision-making processes. Early approaches rely exclusively on gene expression data, thus their predictive power remain fundamentally constrained by the vast combinatorial space of potential gene-gene interactions. Subsequent methods integrate prior knowledge to mitigate this challenge by restricting the solution space to biologically plausible interactions. However, we argue that the effectiveness of these approaches is contingent upon the precision of prior information and the reduction in the search space will circumscribe the models' potential for novel biological discoveries. To address these limitations, we introduce KINDLE, a three-stage framework that decouples GRN inference from prior knowledge dependencies. KINDLE trains a teacher model that integrates prior knowledge with temporal gene expression dynamics and subsequently distills this encoded knowledge to a student model, enabling accurate GRN inference solely from expression data without access to any prior. KINDLE achieves state-of-the-art performance across four benchmark datasets. Notably, it successfully identifies key transcription factors governing mouse embryonic development and precisely characterizes their functional roles. In mouse hematopoietic stem cell data, KINDLE accurately predicts fate transition outcomes following knockout of two critical regulators (Gata1 and Spi1). These biological validations demonstrate our framework's dual capability in maintaining topological inference precision while preserving discovery potential for novel biological mechanisms.
Introduction
Mapping the "wiring diagram" of a cell—the complex interactions where transcription factors (TFs) regulate target genes (TGs)—is a fundamental challenge in systems biology. This task, known as Gene Regulatory Network (GRN) inference, is plagued by the "curse of dimensionality." With tens of thousands of genes, the space of potential interactions is vast. While modern methods use external biological priors (e.g., DNA motifs) to restrict this search space, they often become tethered to the quality of that prior knowledge. This prior-dependency prevents the discovery of novel regulatory mechanisms and makes models vulnerable to noise in external databases.
The KINDLE Framework
To overcome these limitations, we propose KINDLE (Knowledge-guIded Network DistiLlation for prior-free GRN infErence). KINDLE utilizes a "teacher-student" distillation architecture to internalize biological logic during training while remaining prior-free during inference.
The framework operates in three distinct stages:
- Teacher Stage: A hierarchical attention model is trained on temporal gene expression data. It utilizes a spatial mask derived from prior biological knowledge, forcing the model to learn the specific regulatory logic governing known interactions.
- Distillation Stage: The teacher’s learned insights are transferred to a "Student" model. Crucially, the student is denied access to the prior mask. By employing Gaussian Radial Basis Function (RBF) kernels, the student learns the complex, non-linear manifold of gene regulation without being restricted to the original prior list.
- Inference Stage: The trained student model is deployed on raw expression data to infer networks. Because it was taught by a model that understood biological priors but is not restricted by them, it can identify both known and novel interactions with high precision.
Experimental Results
KINDLE was evaluated using the BEELINE benchmarking protocol across four mouse differentiation datasets: embryonic stem cells (mESC) and three hematopoietic stem cell lineages (Erythrocyte, Granulocyte-Monocyte, and Lymphocyte).
KINDLE demonstrated a significant performance leap over existing baselines. On the mESC dataset, KINDLE-Gaussian improved the Area Under the Precision-Recall Curve (AUPRC) from 0.253 (the best baseline) to 0.646, representing a 155% improvement. Furthermore, in the mESC data, KINDLE identified 25 key transcription factors, 72% of which were independently verified in literature as critical for embryonic development.
Biological Validation and In Silico Knockouts
Beyond topological accuracy, KINDLE excels at capturing functional biological dynamics. By modeling temporal causality rather than static co-expression, the framework can predict the outcomes of genetic perturbations.
In mouse hematopoietic stem cell (mHSC) data, KINDLE successfully predicted the effects of knocking out critical regulators Gata1 and Spi1. The model accurately demonstrated that silencing Gata1—a high-expression marker in the Erythrocyte lineage—would arrest red blood cell development. Similarly, it characterized the role of Spi1 in myeloid lineage commitment, matching real-world biological observations.
Conclusion
KINDLE represents a paradigm shift in GRN inference by treating prior knowledge as "privileged information" useful for training but unnecessary for deployment. While the framework currently requires temporal or pseudotime-ordered data and focuses primarily on transcriptional regulation, its ability to discover novel interactions while maintaining state-of-the-art accuracy provides a robust tool for deciphering cellular decision-making.
Summary updated