Virtual Cell Alignment Project

Overview

Align the dismech knowledge base with the CZI Virtual Cell initiative, enabling disease mechanism data to connect with single-cell foundation models (scGPT, UCE, TranscriptFormer, GREmLN, rBio) and the CELLxGENE Census data platform.

The CZI Virtual Cell project builds AI-based "digital twins" of human cells -- large neural-network models that simulate cellular behavior across molecular and structural scales. dismech's structured disease pathophysiology data (cell types, biological processes, gene products, perturbation targets) is a natural complement: it provides the disease-context layer that virtual cell models need for translational predictions.

Key CZI Resources

• CELLxGENE Discover: Largest standardized single-cell data aggregation (tens of millions of cells)
• CELLxGENE Census: Computational query layer with Python/R APIs (TileDB-SOMA)
• Virtual Cells Platform (VCP): Model hosting at virtualcellmodels.cziscience.com
• Models: scGPT, UCE, scGenePT, TranscriptFormer, GREmLN, rBio
• Billion Cells Project: 1B single-cell profiles mapping genetic perturbations

Reference Papers

Tasks

Tier 1: Low Effort (schema/config changes, data annotation)

• [x] Add cellxgene: prefix to schema for CELLxGENE dataset accessions
• [x] Add CELLxGENE dataset references to disorders that already have datasets sections
• [x] Annotate CELLxGENE datasets with publications, conditions, and sample_types
• [ ] Add CELLxGENE dataset references to disorders with cell_types but no datasets
• [ ] Document CL term coverage: which dismech cell types have CELLxGENE representation

Tier 2: Medium Effort (new data fields, cross-references)

• [ ] Add marker gene sets to CellTypeDescriptor entries (genes used by scGPT/UCE for cell identity)
• [ ] Cross-reference dismech biological_processes with GO terms used in scGenePT gene embeddings
• [ ] Link pathophysiology gene_products to perturbation datasets (e.g., Perturb-seq screens)
• [ ] Add computational_models section to Disease class for virtual cell model predictions
• [ ] Populate HCA project references (prefix exists but unused)

Tier 3: High Effort (schema extensions, integrations)

• [ ] Schema extension: PerturbationPrediction class linking gene perturbations to predicted cell state changes
• [ ] CELLxGENE Census API integration for automated cell type enrichment of disorder files
• [ ] Rendering: cell-level molecular data views in HTML pages (embedding visualizations, gene programs)
• [ ] Schema extension: GeneProgram class for co-expression modules identified by TranscriptFormer
• [ ] Integration with rBio for automated perturbation reasoning over dismech gene targets
• [ ] Add support for embedding-based cell similarity annotations (UCE latent space distances)

Tier 4: Aspirational (research directions)

• [ ] Map dismech treatment targets to virtual cell perturbation predictions (drug target -> predicted cell response)
• [ ] Use virtual cell models to predict novel cell types affected by disease mechanisms
• [ ] Connect comorbidity trajectories to shared cell-state perturbation signatures
• [ ] Validate dismech pathophysiology claims against scRNA-seq differential expression

STATUS

Tier 1 Progress

• [x] cellxgene: prefix added to schema (2026-03-07)
• [x] CELLxGENE dataset annotation (7 disorders: Alzheimer's, Parkinson's, Psoriasis, Ulcerative Colitis, Asthma, NSCLC, Cystic Fibrosis)
• [ ] CL term coverage audit (0/N)

Notes

2026-03-07 (Project Creation)

Project created to align dismech with CZI Virtual Cell initiative.

Low-effort wins identified: 1. Schema already has SINGLE_CELL_RNA_SEQ data type, hca: prefix, cell type descriptors with CL terms, and dataset infrastructure 2. Adding cellxgene: prefix is a one-line schema change 3. Several disorders already have datasets sections that could be augmented with CELLxGENE references 4. Cell type annotations in pathophysiology already use CL ontology terms that map directly to CELLxGENE cell type annotations

Key gap: No existing CELLxGENE references in any disorder file despite infrastructure support.

Schema observations: - DatasetTypeEnum already includes SINGLE_CELL_RNA_SEQ and SPATIAL_TRANSCRIPTOMICS - SampleTypeDescriptor already has cell_type_term (CL binding) and tissue_term (UBERON binding) - hca: prefix configured but unused across all disorder files - cellxgene: prefix missing entirely