New Benchmark and Methods for Gene Regulatory Network Inference from Single-Cell Foundation Models
A new study on arXiv (2605.08128) addresses the challenge of Gene Regulatory Network (GRN) inference using single-cell Foundation Models (scFMs). While scFMs promise enhanced transcriptomic encoding, their performance in GRN inference remains poor because standard reconstruction-based pre-training fails to capture latent regulatory signals. The authors introduce a GRN generalization benchmark to evaluate zero-shot predictions on unseen genes and datasets, a task difficult for traditional methods. They also propose two novel methods—Virtual Value Perturbation and Gradient Trajectory—to distill regulatory knowledge from foundation models.
Key facts
- arXiv paper 2605.08128 addresses GRN inference from single-cell transcriptomic data.
- Standard pre-training objectives in scFMs fail to capture regulatory signals.
- A new GRN generalization benchmark evaluates zero-shot predictions on unseen genes and datasets.
- Two methods proposed: Virtual Value Perturbation and Gradient Trajectory.
- The benchmark is inherently challenging for traditional GRN inference methods.
- The study aims to unlock regulatory knowledge within foundation models.
- Single-cell Foundation Models (scFMs) are used for transcriptomic encoding.
- The paper is categorized as a cross-type announcement.
Entities
Institutions
- arXiv