scACORN: Context-engineered agent orchestration of specialized small language models for single-cell transcriptomic interpretation
- 1York University; Vector Institute, Toronto, Ontario, Canada
Abstract
Single-cell atlases now exceed 66 million cells, but turning a ranked expression profile and a free-form biological question into a reliable, evidence-grounded answer remains unsolved. Scaling a single model does not resolve this, because single-cell interpretation is a heterogeneous family of tasks whose correct answer depends on tissue, cohort, perturbation and annotation resolution. Here we present scACORN, an agentic alternative to monolithic single-cell language models that combines specialized small language models with context-engineered agent orchestration for their selection and composition at inference time. Each expert is built in two stages: domain-aligned contrastive adaptation fits a pretrained cell-to-text backbone to the transcriptomic geometry of a target dataset, and geometry-preserving specialization learns question-conditioned biological completions without eroding that geometry. A fixed orchestrating language model agent then selects and combines experts under a natural-language playbook that is itself optimized from textual feedback, with no gradient updates to the orchestrator. Across 10 Tabula Sapiens tissues, domain alignment raised transfer macro-F1 from 0.36 to 0.64 and Recall@5 from 0.87 to 0.97; specialized experts reached 0.89 mean exact-match annotation accuracy; and playbook optimization reduced unsupported gene citations from 14.5% to 3.5%. Our findings support specialization and orchestration as complementary responses to the heterogeneity and evidentiary demands of single-cell analysis.
Related
- Capsid-specialized protein language models reveal higher-order viral architecture from sequence
- Deep generative embeddings of gene expression and splicing reposition the interpretation of single-cell transcriptomic signatures
- ScProteoAgent enables natural-language-driven single-cell proteomics analysis and interpretation
- Benchmarking Large Language Models for Transcriptomic Data Analysis and Visualization: Multi-Model Workflow Evaluation
- An immune world model with multi-agent orchestration for peptide vaccines
BibTeX
@misc{rastimeymandi2026scacorn,
title = {{scACORN: Context-engineered agent orchestration of specialized small language models for single-cell transcriptomic interpretation}},
author = {A. Rasti-Meymandi and S. Nahali and E. Paramithiotis and A. M. Cheung and E. Dolatabadi},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.10.750801},
url = {https://www.biorxiv.org/content/10.64898/2026.09.10.750801v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - scACORN: Context-engineered agent orchestration of specialized small language models for single-cell transcriptomic interpretation
AU - A. Rasti-Meymandi
AU - S. Nahali
AU - E. Paramithiotis
AU - A. M. Cheung
AU - E. Dolatabadi
PY - 2026
DA - 2026/09/10
PB - bioRxiv
DO - 10.64898/2026.09.10.750801
UR - https://www.biorxiv.org/content/10.64898/2026.09.10.750801v1
AB - Single-cell atlases now exceed 66 million cells, but turning a ranked expression profile and a free-form biological question into a reliable, evidence-grounded answer remains unsolved. Scaling a single model does not resolve this, because single-cell interpretation is a heterogeneous family of tasks whose correct answer depends on tissue, cohort, perturbation and annotation resolution. Here we present scACORN, an agentic alternative to monolithic single-cell language models that combines specialized small language models with context-engineered agent orchestration for their selection and composition at inference time. Each expert is built in two stages: domain-aligned contrastive adaptation fits a pretrained cell-to-text backbone to the transcriptomic geometry of a target dataset, and geometry-preserving specialization learns question-conditioned biological completions without eroding that geometry. A fixed orchestrating language model agent then selects and combines experts under a natural-language playbook that is itself optimized from textual feedback, with no gradient updates to the orchestrator. Across 10 Tabula Sapiens tissues, domain alignment raised transfer macro-F1 from 0.36 to 0.64 and Recall@5 from 0.87 to 0.97; specialized experts reached 0.89 mean exact-match annotation accuracy; and playbook optimization reduced unsupported gene citations from 14.5% to 3.5%. Our findings support specialization and orchestration as complementary responses to the heterogeneity and evidentiary demands of single-cell analysis.
ER -