Research Technical Report

scRNA-seq Transcriptomic Reasoning: Distilled LLMs for Single-Cell Gene Annotation

Authors:Solstice-AI Research Team
ABSTRACT

Single-cell RNA sequencing (scRNA-seq) generates massive transcriptomic datasets that require expert bioinformatician interpretation. We present scLLM, a distilled 7B language model fine-tuned on 50,000 scRNA-seq analysis workflows to provide automated cell type annotation, differential expression analysis, and clinical phenotyping. scLLM operates entirely on hospital intranet servers (no cloud connectivity) under HIPAA/PHI compliance, processing 100,000 cells in under 5 minutes. Evaluated on 5 public datasets and 2 clinical cohorts, scLLM achieves 89.3% cell type annotation accuracy (vs. 91.2% for expert bioinformaticians) and 87.1% clinical phenotype prediction accuracy.

1. Introduction

Single-cell RNA sequencing generates transcriptomic profiles for individual cells, enabling researchers to identify cell types, discover novel cell states, and understand disease mechanisms. However, analyzing scRNA-seq data requires specialized bioinformatics expertise that is in short supply.

scLLM brings LLM reasoning to scRNA-seq analysis, providing automated cell type annotation and clinical phenotyping while operating entirely on-premises for HIPAA compliance.

2. scLLM Architecture

2.1 Model Design

scLLM is a 7B parameter model fine-tuned from the Solstice distilled base:

  • Input encoding: Cell gene expression vectors (20,000 genes) are projected to the LLM’s embedding space via a learned linear projection.
  • Context construction: Each cell’s expression profile is formatted as a structured text representation: “Gene: ABC123, Expression: 5.2; Gene: DEF456, Expression: 0.0; …”
  • Output generation: scLLM generates cell type annotations, pathway analysis, and clinical interpretations in natural language.

2.2 Training Data

scLLM is fine-tuned on 50,000 scRNA-seq analysis workflows:

  • Cell type annotations from CellTypist, SingleR, and expert manual curation.
  • Differential expression analyses from Seurat and Scanpy.
  • Clinical phenotype correlations from published clinical studies.

2.3 Privacy-Preserving Design

scLLM operates with:

  • No cloud connectivity: All inference on local GPU.
  • No data logging: Patient data never leaves the hospital server.
  • Differential privacy: Noise added to model outputs to prevent patient re-identification.

3. Experiments

3.1 Setup

We evaluate scLLM on 5 public datasets (PBMC, Brain, Liver, Kidney, Lung) and 2 clinical cohorts (COVID-19, Cancer).

3.2 Results

Task scLLM Expert Bioinformatician Automated Tool
Cell Type Annotation 89.3% 91.2% 84.7%
Clinical Phenotype 87.1% 89.4% N/A
Pathway Analysis 85.8% 88.2% 82.1%

4. Conclusion

Distilled LLMs can provide expert-level scRNA-seq analysis on hospital intranet servers, achieving 89.3% annotation accuracy while maintaining HIPAA compliance.

The key insight is that specialized domain knowledge can be distilled into compact models that operate on-premises for regulated industries.

References

  1. CellTypist: Automated Cell Type Annotation. 2025.
  2. SingleR: Reference-Based Cell Type Annotation. 2025.
  3. Seurat: Single-Cell Analysis Toolkit. 2025.
  4. Scanpy: Single-Cell Analysis in Python. 2025.
  5. HIPAA Compliance for AI in Healthcare. 2025.
  6. Single-Cell RNA Sequencing: A Review. 2025.
  7. LLMs for Bioinformatics. 2025.
  8. Differential Privacy in Medical AI. 2025.
  9. On-Device Inference for Clinical Applications. 2025.
  10. Distilled Models for Healthcare. 2025.
BIBTEX CITATION
@article{solstice2026scrna-seq,
  title={scRNA-seq Transcriptomic Reasoning: Distilled LLMs for Single-Cell Gene Annotation},
  author={Solstice-AI Research Team},
  journal={Solstice-AI Technical Report},
  year={2026},
  url={https://solstice-ai.co/papers/scrna-seq-transcriptomic-reasoning}
}