scACORN: Context-engineered agent orchestration of specialized small language models for single-cell transcriptomic interpretation

Single-cell atlases have surpassed 66 million cells, yet converting expression profiles and biological queries into accurate, evidence-based responses remains unresolved. Scaling a single model proves inadequate because single-cell interpretation encompasses diverse tasks, each with answers contingent on tissue, cohort, perturbation, and annotation resolution.

Introducing scACORN, an alternative to monolithic single-cell language models that employs agentic orchestration of specialized small language models. Each expert undergoes two stages: domain-aligned contrastive adaptation aligns a pretrained cell-to-text backbone with the transcriptomic geometry of a specific dataset, followed by geometry-preserving specialization that learns question-conditioned biological completions without losing that geometry.

A language model agent orchestrates the selection and combination of experts based on a natural-language playbook, which is optimized using textual feedback without modifying the orchestrator itself. Across 10 Tabula Sapiens tissues, domain alignment increased transfer macro-F1 from 0.36 to 0.64 and Recall@5 from 0.87 to 0.97; specialized experts achieved 0.89 mean exact-match annotation accuracy; and playbook optimization lowered unsupported gene citations from 14.5% to 3.5%.

The results underscore the complementary roles of specialization and orchestration in addressing the heterogeneity and evidentiary challenges of single-cell analysis.