Smithery is now a part of Arcade.dev! Read more in our announcement here
Unified agent for leveraging single-cell foundation models (scGPT, scBERT, Geneformer, scFoundation) for cross-species annotation, perturbation prediction, and gene network inference.
name: 'scfoundation-model-agent' description: 'Unified agent for leveraging single-cell foundation models (scGPT, scBERT, Geneformer, scFoundation) for cross-species annotation, perturbation prediction, and gene network inference.' measurable_outcome: Execute skill workflow successfully with valid output within 15 minutes. allowed-tools:
The scFoundation Model Agent provides a unified interface to leverage state-of-the-art single-cell foundation models for diverse downstream tasks. It integrates scGPT, scBERT, Geneformer, scFoundation, and emerging models to enable cross-species cell annotation, in silico perturbation prediction, gene regulatory network inference, and batch integration.
Cross-Species Cell Annotation: Transfer cell type labels across species using unified embeddings.
In Silico Perturbation: Predict gene expression changes from knockouts/treatments.
Gene Regulatory Network Inference: Discover TF-target relationships from attention patterns.
Batch Integration: Remove technical variation while preserving biology.
Cell Embedding Generation: Generate universal cell representations for any downstream task.
Multi-Model Ensemble: Combine predictions from multiple foundation models.
LLM-Based BGC Discovery (Reference): DeepSeMS (Xu et al., Nat Comput Sci, Apr 2026) demonstrates a large language model approach for revealing hidden biosynthetic potential of the global ocean microbiome. Complementary reference for applying LLM/foundation-model methods to biosynthetic gene cluster (BGC) discovery from microbiome metagenomic data, extending foundation-model coverage beyond single-cell transcriptomics. PMID 42062603.
scPRINT-2 (Next-Gen Single-Cell Foundation Model): scPRINT-2 (cantinilab, GitHub repo, Apr 2026) is a next-generation single-cell foundation model alongside scGPT/scBERT/Geneformer/scFoundation. Intended for large-scale single-cell perturbation prediction, cell annotation, and gene regulatory network inference. Distributed as a Jupyter Notebook reference implementation; consult the upstream repository for licensing terms before use. https://github.com/cantinilab/scPRINT-2
Stack (In-Context Learning Single-Cell Foundation Model): Stack (Arc Institute, GitHub repo, Apr 2026) is a single-cell foundation model that supports in-context learning at inference time, removing the need for task-specific fine-tuning when adapting to new datasets. Important pattern shift versus prior fine-tuning-based foundation models — enables few-shot cell type annotation and perturbation prediction across previously unseen datasets by conditioning on context examples at inference. https://github.com/ArcInstitute/stack
Phenotype-Associated Subpopulation Identification: Zhao et al. (BMC Biol, Apr 2026, PMID 42050525) describe a workflow combining transfer learning from single-cell foundation models with statistical ensemble learning to identify phenotype-associated cell subpopulations. Use foundation-model embeddings as the transfer substrate when phenotype signal may reflect cell states within or across standard clusters; validate candidates through ensemble stability/resampling and explicit phenotype-association checks rather than treating Leiden/UMAP clusters alone as phenotype-linked discoveries.
| Model | Parameters | Training Data | Strengths |
|---|---|---|---|
| scGPT | 50M | 33M human cells | General purpose, perturbations |
| Geneformer | 10M | 30M cells | Chromatin, gene networks |
| scBERT | 20M | 1.2M cells | Cell type annotation |
| scFoundation | 100M | 50M cells | Large-scale, multi-species |
| scTab | 15M | 22M cells | Tabular prediction |
| UCE (Universal Cell Embeddings) | 100M | 36M cells | Cross-species transfer |
Input: Single-cell RNA-seq data (AnnData format).
Model Selection: Choose appropriate model(s) for task.
Preprocessing: Tokenize genes, normalize expression.
Inference: Generate embeddings or predictions.
Task Execution: Annotation, perturbation, or network inference.
Ensemble (Optional): Combine multi-model predictions.
Output: Annotated data, predictions, networks.
User: "Use scGPT to predict the effect of CRISPR knockout of TP53 on these cancer cells."
Agent Action:
python3 Skills/Genomics/scFoundation_Model_Agent/foundation_predict.py \
--input cancer_cells.h5ad \
--model scgpt \
--task perturbation \
--perturbation "TP53 knockout" \
--model_checkpoint scgpt_human_gene_v1.pt \
--output tp53_ko_predictions.h5ad
python3 foundation_predict.py \
--input query_cells.h5ad \
--model geneformer \
--task annotation \
--reference tabula_sapiens.h5ad \
--output annotated_cells.h5ad
python3 foundation_predict.py \
--input cells.h5ad \
--model scgpt \
--task grn_inference \
--transcription_factors tf_list.txt \
--output gene_network.csv
python3 foundation_predict.py \
--input multi_batch.h5ad \
--model scfoundation \
--task integration \
--batch_key batch \
--output integrated.h5ad
Transfer-learn a single-cell foundation model on labeled reference cells, project phenotype labels (disease status, treatment response, etc.) onto query embeddings, then run statistical ensemble learning over multiple resamples/models to flag robust phenotype-linked subpopulations. Use this mode when the goal is phenotype association rather than only cell-type annotation: confirm that selected subpopulations remain stable across ensemble runs and show phenotype association beyond standard clustering boundaries. Pattern follows Zhao et al. (BMC Biol, 2026, PMID 42050525).
python3 foundation_predict.py \
--input query_cells.h5ad \
--model scfoundation \
--task phenotype_subpopulation \
--reference labeled_reference.h5ad \
--phenotype_key disease_status \
--transfer_mode finetune \
--ensemble_runs 25 \
--ensemble_method bootstrap_vote \
--output phenotype_subpopulations.h5ad
| Task | Output | Format |
|---|---|---|
| Annotation | Cell type labels | .h5ad obs column |
| Perturbation | Predicted expression | .h5ad layer |
| GRN | TF-target edges | .csv, .graphml |
| Integration | Corrected embeddings | .h5ad obsm |
| Embeddings | Cell representations | .h5ad obsm |
| Task | Model | Dataset | Performance |
|---|---|---|---|
| Annotation | scGPT | Tabula Sapiens | 93% accuracy |
| Annotation | Geneformer | HLCA | 91% accuracy |
| Perturbation (R²) | scGPT | Norman 2019 | 0.87 |
| Integration (kBET) | scFoundation | Multi-atlas | 0.92 |
| Cross-species | UCE | Human→Mouse | 85% F1 |
Transformer Backbone:
Perturbation Module:
Transfer Learning:
| Use Case | Recommended Model | Reason |
|---|---|---|
| General annotation | scGPT | Broad training, robust |
| Cross-species | UCE | Species-agnostic embeddings |
| Perturbation | scGPT | Best perturbation performance |
| GRN inference | Geneformer | Attention → regulatory links |
| Large-scale | scFoundation | Efficient, scalable |
| Tabular prediction | scTab | Optimized for classification |
| Strategy | Method | Benefit |
|---|---|---|
| Majority Vote | Mode of predictions | Robust to outliers |
| Weighted Average | Confidence-weighted | Leverages uncertainty |
| Stacking | Meta-model | Learns model strengths |
| Attention Fusion | Cross-model attention | Deep integration |
AI Group - Biomedical AI Platform
