RegNetAgents: A Multi-Agent LangGraph Built on Top of Pre-trained Networks for Identifying Cancer Drivers

The main focus of RegNetAgents is not to invent yet another method for constructing gene regulatory networks, but to build a multi-agent layer on top of existing ones. The authors candidly refer to the system as a “downstream analytical layer”: it takes pre-computed networks from TCGA (bulk tumors) and the GREmLN project (single-cell), and then, via a DAG workflow on LangGraph, classifies candidates, filters them using OncoKB, and assigns a mode of action to the regulatory links from the tumor networks.

Architecturally, this is more interesting than “just another LLM pipeline.” Agents operate via a unified Python API and the Model Context Protocol (MCP) client—meaning the system can be integrated as a tool into broader agent chains, rather than just being run via a script. For those building agents in biomedicine, this is a sign: MCP has made its way into genomic analysis tasks.

Candidates are ranked by the consistency of evidence across three categories: Both (both types of networks), TCGA-only, GREmLN-only. For 11 focus genes in breast cancer and 12 in colorectal cancer, the system yields significant enrichment for annotated oncogenes from OncoKB. Stouffer Z ranges from 5.51 to 7.06, with all p < 0.0001. Control sets (housekeeping, non-driver) show no enrichment—this confirms the specificity of the signal rather than a ranking artifact.

The advanced module adds assessments of oncogenic potential, druggability, clinical relevance, and network vulnerability. Essentially, the authors are attempting to close the loop from “regulatory candidate” to “hypothesis for the biologist”—and this is where multi-agent orchestration is appropriate: each step requires different sources of knowledge and different modes of reasoning.

The catch is obvious from the setup itself. The quality of the candidates depends on the quality of the pre-computed networks (ARACNe for bulk data, GREmLN for single-cell data). If the upstream inference is weak for a specific tumor type, the multi-agent layer won’t be able to extract what isn’t there. This isn’t a generative model that “infer” regulatory connections—it’s a structured query on top of third-party data. For tasks where the networks are already reliable, this is an advantage; for rare cancer types with poor single-cell coverage, it is a limitation.

Source: cs.AI updates on arXiv.org