Language-encoded network topology enables large language models to reason about complex networks

Large language models (LLMs) are powerful reasoning engines, yet they struggle to interpret network structures when provided as raw edge lists or numerical tables. This paper introduces BioGlyph, a method that compiles network topology into a human- and model-readable vocabulary of structural roles. Instead of forcing the LLM to infer structure from raw data, BioGlyph uses graph-partitioning algorithms to identify roles such as hubs, community cores, and cross-community connectors. These roles are then translated into a universal, semantically interpretable language that includes supporting evidence and the structural consequences of removing specific nodes or edges.

BioGlyph consistently outperforms alternative network representations across twenty networks spanning biological, social, and information domains. In benchmarks, BioGlyph improved system accuracy by up to 26 percentage points compared to edge-based or numerical representations. The method is particularly effective in dense, community-structured networks where direct textual representations often exceed the model's context window or are too complex to interpret. Ablation studies reveal that the performance gains are driven by two factors: the selective compression of network information and the explicit communication of structural consequences, which helps the model reason about perturbations.

Beyond improving LLM reasoning, BioGlyph identifies biologically meaningful organization in protein-interaction networks without prior annotation. In the budding-yeast interactome, proteins identified as cross-community connectors were significantly more likely to be essential genes than peripheral proteins. Similarly, in human functional interaction networks, cut nodes identified by BioGlyph were enriched for selective cancer dependencies. These results suggest that the structural vocabulary captures genuine biological importance, providing a bridge between quantitative network analysis and scientific interpretation.