About Cella

Decoding Cell-Type Identity Beyond Raw Expression

Cella (Latin for cell) is an open-source biological knowledge and decision-support platform built on the CellTypist Human Organ Atlases. It transforms machine learning classifier weights and cross-organ cellular networks into an interactive, interpretable knowledge engine.

🧬 12 Human Organs🔬 >3.6 Million Cells🏷️ 145 Harmonized Cell Types⚡ Machine Learning Weights🌐 Open Source & Reproducible

1. The Scientific Problem: Expression vs. Discriminative Identity

Single-cell RNA sequencing (scRNA-seq) has mapped millions of human cells across diverse organs. However, almost all existing atlas browsers (such as CELLxGENE or the Broad Single Cell Portal) only answer one fundamental question:

“Is gene X detected in cell type Y?” (Expression-based view)

While informative, raw gene expression is often noisy, plagued by technical dropouts, and dominated by ubiquitous metabolic or housekeeping genes that do not define what makes a cell type unique.

When machine learning models (such as CellTypist's regularized logistic regression classifiers) are trained to annotate cells, they learn discriminative classifier weights (β coefficients). These weights explicitly answer a far deeper biological question:

“Which genes distinguish cell type Y from every other cell type in this tissue, and how does that definition shift across the human body?”

Until Cella, these classifier weights remained trapped inside serialized Python .pkl model files and static journal figures. Cella extracts, harmonizes, and surfaces these weights as an interactive suite of analytical tools.

Traditional Atlas Viewers vs. Cella

Feature / Dimension	Traditional Atlas Portals	Cella Platform
Primary Data Asset	Raw / normalized expression matrices (X)	Learned discriminative weight matrices (W) + Cell Ontology graph
Marker Identification	Heuristic differential expression or curated lists	Regularized logistic regression classifier coefficients
Cross-Organ Comparison	Manual inspection of disjoint tissue datasets	Systematic split into Universal Core vs. Tissue-Adapted markers
Reverse Querying	Basic gene search across expression tables	Benchmarked signature ranking across 12 organs, with discriminative weights as evidence
Cross-Tissue Relationships	Static dendrograms or single-tissue UMAPs	Interactive ~145-node cross-organ similarity & lineage graph

2. Key Research & Methodological Contributions

This project represents an original integration of computational biology, machine learning interpretability, and scalable software systems engineering. Below are the primary methodological contributions:

Theoretical & Biological

Universal Core vs. Tissue-Adapted Identity

Formulated a systematic framework to decompose cellular identity across organs. By evaluating discriminative weights across all 12 organs where a cell type resides, Cella partitions markers into:

Universal Core Identity: Consensus genes with high positive weights (β > 0) in every organ (e.g., canonical lineage regulators like CD3D/E for T cells).
Tissue-Adapted Identity: Genes with strong discriminative power only in a single organ, capturing microenvironmental adaptations (e.g., MARCO in alveolar macrophages vs. C1QA/B in microglia).

Ontology & Harmonization

Cell Ontology (CL) Graph Alignment

Heterogeneous dataset annotations across 38 studies were mapped and standardized into a unified Cell Ontology directed acyclic graph (CL IDs). This provides a single join key across 145 canonical cell types, resolving synonymy and inconsistent nomenclature across differing lab annotations.

Network Biology

Cross-Tissue Relationship & Lineage Graph

Constructed an interactive cross-organ cell-type relationship graph based on cosine similarity between high-dimensional classifier weight vectors. Precomputed deterministic spring embeddings enable researchers to explore phenotypic convergence, developmental lineages, and shared regulatory programs across the human body without client-side simulation lag.

Algorithms & ML

Instant Reverse-Lookup Signature Classifier

A reverse-lookup engine that ranks every cell type across 12 organs by how widely it expresses a user-provided gene signature, chosen by benchmarking scoring rules against external marker sets. It provides:

A confidence call calibrated on held-out marker sets (expression level plus lead over the next cell type).
The organ in which each candidate cell type matches best, with per-gene discriminative weights.
Automated flagging of potential novel, rare, or uncharacterized cellular states.

Multi-Omics Inference

Downstream Functional & Regulatory Derivations

Integrated multi-modal analytical layers on top of atlas pseudobulk profiles:

Transcription Factor Activity: Univariate Linear Modeling (ULM) over CollecTRI regulons via Decoupler.
Pathway Activity: GSEA Hallmark enrichment per cell type.
Cell-Cell Communication: Ligand-receptor interactome reconstruction using LIANA consensus databases.

Systems Engineering

Zero-Overhead Data Pipeline & 60 FPS UI

Built an efficient, reproducible computational pipeline and high-performance web interface:

Remote HDF5 Range Reads: Extracts cell metadata from remote Sanger buckets in seconds via HTTP range requests without downloading 30+ GB raw files.
Vectorized In-Memory Serving: FastAPI backend with Parquet columnar caches, sub-millisecond query responses, and ETag cache revalidation.
Batched Canvas Rendering: Custom HTML5 Canvas rendering engine for 25,000 cells running at a locked 60 FPS.

3. The 5 Interactive Tools

🏛️

Organ Browser

Explore 12 organ atlases — donor demographics, cell composition, 25k UMAP embeddings, TF activity, and cell-cell communication networks.

🎯

Discriminative Marker Atlas

Query any cell type to inspect its universal core defining genes versus tissue-adapted organ-specific markers.

🔍

Fingerprint Explorer

Compare side-by-side heatmaps of positive and negative classifier weights for the same cell type across different organs.

🕸️

Relationship Graph

Navigate the cross-tissue network of 145 harmonized cell types, filtering by organ and inspecting shared discriminative features.

⚡

Reverse Lookup

Paste experimental gene lists to score them against all organ models, identify candidate cell types, and detect novel states.

4. Research Profile & Scientific Roadmap

Md. Jubayer Hossain

Founder & CEO, DeepBio · Multi-Omics Scientist

This project was conceived, architected, and developed by Md. Jubayer Hossain as an independent research initiative bridging large-scale single-cell transcriptomics, interpretable machine learning, and high-performance bioinformatics software engineering.

Future Research & Roadmap:

Spatial Transcriptomics Integration: Mapping discriminative classifier weights onto 10x Visium and CosMx spatial coordinates to study in situ cellular niches and neighborhood interactions.
Single-Cell Foundation Model Embeddings: Comparing linear discriminative weight manifolds against latent representations from generative foundation models (e.g., scGPT, Geneformer, scVI).
Disease State Perturbation Modeling: Extending the cross-organ framework to characterize pathological shifts in cell identity during autoimmune conditions, chronic inflammation, and cancer metastasis.

5. Data Sources & Citations

All biological data utilized in Cella is sourced from official, peer-reviewed reference resources:

CellTypist Organ Atlases & Models:
Domínguez Conde, C., Xu, C., Jarvis, L. B., Rainbow, D. B., et al. (2022). Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science, 376(6594), eabl5197. doi:10.1126/science.abl5197

CellTypist v2 Reference:
Xu, C., et al. (2024). CellTypist 2.0: A comprehensive reference atlas for automated cell-type annotation across human organs. celltypist.org

Cell Ontology (CL):
Diehl, A. D., et al. (2016). The Cell Ontology 2016: enhanced content, modularization, and ontology interoperability. Journal of Biomedical Semantics, 7(1), 44.