NeuroSC: Leveraging Pretrained Single-Cell Models for Brain Cell Classification

1. Overview

We propose to fine-tune large-scale pretrained foundation models of single-cell data, such as scGPT[1], GeneFormer[2], SCimilarity[3],… to accurately predict brain cell types using the Allen Brain Cell Atlas dataset[4]. This project is motivated by the growing importance of efficient, accurate cell type annotation in neuroscience and single-cell transcriptomics. The foundation models are pretrained on large-scale biological corpora and represent a promising approach to transfer learning in single-cell analysis[1, 2, 3].

We confirm access to required computational resources (GPU clusters from AWS services), data (publicly available Allen Brain Cell Atlas), and ethics exemption (no new data collection involving humans/animals).

The project will run over 6 months, including model tuning (2 months), evaluation/benchmarking (2 months), and final packaging/output (2 months). We commit to sharing all model weights, benchmarking results, and a user-friendly interface to promote transparency and community uptake.

2. Introduction

Background and Rationale:
Single-cell RNA-seq has revolutionized our understanding of cellular diversity, particularly in the brain. Annotating cell types, however, remains a major bottleneck, typically requiring manual curation or classical machine learning models with limited generalizability. Recently, foundation models such as scGPT and GeneFormer have emerged as powerful tools trained on massive scRNAseq datasets, demonstrating strong transferability and representational learning across tasks.

Research Questions:
(1) Can foundation models be fine-tuned to improve brain cell type annotation? (2) How do they perform relative to conventional classifiers? (3) Can we provide practical tools to accelerate adoption in neuroscience?

Hypotheses:

H1: Fine-tuned foundation models will outperform conventional methods (e.g., logistic regression, random forest, Seurat-based classifiers) in predicting cell types in the Allen Brain Cell Atlas.
H2: Model performance will generalize across held-out brain regions and developmental stages.
H3: A streamlined interface will reduce annotation time and technical entry barriers for end users.

Impact Statement:
Regardless of performance, this study will provide a systematic benchmark of foundation models for brain transcriptomic data. It will inform both the scientific community and tool developers, enabling more scalable and reproducible cell annotation pipelines.

3. Methods

Participants/Sample

We will use the Allen Brain Cell Atlas, which contains annotated scRNA-seq data spanning diverse human brain regions and developmental timepoints. This dataset includes three million cells that were clustered into hierarchical groups of 31 superclusters, 461 clusters, and 3313 subclusters.

Inclusion Criteria: Annotated cell types with high-confidence labels; high-quality cells (e.g., minimum gene count, mitochondrial threshold).

Exclusion Criteria: Ambiguous labels, low-quality cells.

Sample Size Justification: >3 million cells available. We will subsample to ensure computational tractability while preserving diversity.

Materials and Procedures

We will use:

Foundation Models: Pretrained versions of scGPT and GeneFormer.
Task: Multiclass classification of cell types.
Data Preparation: Normalize and preprocess scRNA-seq expression matrices. Map gene symbols to model vocabularies. Matching the annotated labels.
Fine-tuning: Implement task-specific heads and train using cross-entropy loss on annotated subsets.
Validation: Hold out 20% of the data (already subsampled) for testing generalization.
Infrastructure: GPU clusters (A100 or equivalent).

Planned Analyses

Preprocessing Pipeline:

Quality filtering
Normalization (e.g., log CPM)
Feature selection (highly variable genes).

Analysis Pipeline:

Fine-tune each foundation model.
Compare performance to baseline classifiers (e.g., logistic regression, SVM, Seurat's label transfer).
Metrics: accuracy, F1 score, and confusion matrices across major classes.

Hypothesis-Specific Tests:

H1: Use repeated k-fold cross-validation. Test improvement using paired t-tests of model accuracy.
H2: Evaluate on unseen regions/stages. Use ANOVA to test performance variation.
H3: Conduct informal user testing of prototype interface.

Contingent Analyses:

If overfitting occurs, then apply dropout regularization and early stopping.

Ethics and Data Management

Ethical Approval: Not required (no human or animal subject research).
Confidentiality: Not applicable (public dataset).
Data Sharing: All processed datasets, model weights, and code will be shared on GitHub, HuggingFace, and Zenodo under open licenses. The user interface will be published as a pip-installable package and hosted demo.

Ethics and Data Management

State ethical approval status or plan, confidentiality measures, and data sharing/archiving plan.

4. Budget

	Item Estimated Cost (USD)
GPU Compute (Cloud credits)	$2,000
HPC access	$1,000
Data sharing and hosting	$1,800
Miscellaneous/Contingency	$200
Total	$5,000

6. References

[1] Cui, H., Wang, C., Maan, H., Pang, K., Luo, F., Duan, N. and Wang, B., 2024. scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature methods, 21(8), pp.1470-1480.

[2] Cui, Z., Xu, T., Wang, J., Liao, Y. and Wang, Y., 2024, April. Geneformer: Learned gene compression using transformer-based context modeling. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 8035-8039). IEEE.

[3] Heimberg, G., Kuo, T., DePianto, D., Heigl, T., Diamant, N., Salem, O., Scalia, G., Biancalani, T., Turley, S., Rock, J. and Bravo, H.C., 2023. Scalable querying of human cell atlases via a foundational model reveals commonalities across fibrosis-associated macrophages. BioRxiv, pp.2023-07.

[4] Siletti, K., Hodge, R., Mossi Albiach, A., Lee, K.W., Ding, S.L., Hu, L., Lönnerberg, P., Bakken, T., Casper, T., Clark, M. and Dee, N., 2023. Transcriptomic diversity of cell types across the adult human brain. Science, 382(6667), p.eadd7046.

Funding

Funding

NeuroSC: Leveraging Pretrained Single-Cell Models for Brain Cell Classification

Funding Progress

Nonprofit Recipient

Funders

Author Updates

Topics

DOI