How CellDAO Works: A Walkthrough of the Stack

CellDAO is designed to standardize, align, and continuously learn from high-value cellular datasets. Below is an end-to-end view of how raw experimental output becomes part of an evolving Virtual Cell.

1. Data Generation & Local Packaging

Labs perform experiments (imaging, perturbation screens, single-cell omics). Alongside raw files, they export a preliminary metadata bundle: plate maps, channel descriptions, perturbation annotations, acquisition logs.

2. OMS Manifest Construction

A local tooling CLI ingests the preliminary bundle and produces an OMS manifest:

• Sample provenance (cell line, passage, culture conditions)
• Perturbations (compound structures, doses, genetic edits, exposure times)
• Acquisition parameters (instrument ID, objective, channels w/ roles + stains)
• Processing placeholders (to be filled post-feature extraction)
• Licensing & attribution metadata

Validation runs, flagging missing or ambiguous fields for user correction.

3. Feature Extraction & Processing Trace

Containerized pipelines perform segmentation and feature extraction (e.g., cell morphology vectors). Each step appends to the processing trace:

• Segmentation model + version
• Feature extractor + parameter digest
• Normalization and QC filters applied

Outputs: per-cell or per-well feature matrices, QC flags, lineage links back to raw assets.

4. Submission & Integrity Anchoring

The dataset (raw assets optional, processed features mandatory) plus manifest is prepared:

1. Content hashes computed (raw subset, features, manifest).
2. An integrity anchor (hash of hash set) optionally registered on-chain as a contribution record.
3. Dataset registered in a discovery index with search facets.

5. Ingestion & Standardization Layer

On the network side:

• Manifests parsed; schema version compatibility checked.
• Controlled vocabulary normalization (channel roles, perturbation units).
• Batch/site effect variables extracted for downstream modeling.
• Derived dataset lineage incorporated into a global graph.

6. Virtual Cell Training Pipeline

A scheduled (or continuous) training job:

• Loads new OMS-compliant datasets.
• Updates multimodal encoders (imaging, expression) to align latent spaces.
• Trains perturbation-aware dynamics (dose, time, sequential interventions).
• Calibrates uncertainty via likelihood heads + OOD detectors.
• Evaluates on benchmark suite (held-out cell types, perturbations, timepoints).

Model checkpoints record dataset contribution proportions (for attribution or incentives).

7. Inference & Simulation Services

APIs expose:

• State embedding of new observations.
• Predictive simulation: given a baseline state + perturbation plan (compound X @ dose D over T hours), return predicted morphology / expression trajectories.
• Counterfactual queries: difference between interventions A vs. B.
• Uncertainty + OOD indicators.

8. Active Learning Loop

The system ranks candidate experiments by expected information gain:

• High uncertainty regions in latent space.
• Poorly disentangled batch vs. biological signals.
• Underrepresented perturbation-time combinations.

Suggested experiments feed back to labs, closing the loop.

9. Incentive & Governance Layer (Optional / Progressive)

If a DAO / token layer is enabled:

• Usage-based reward distribution to contributing datasets.
• Reputation accrues via validation accuracy, reproducibility audits.
• Community voting on OMS schema evolution and model benchmark updates.

10. Observability & Audit

Dashboards track:

• Dataset ingestion velocity & compliance scores.
• Model performance drift over time.
• Uncertainty calibration metrics.
• Lineage queries (which datasets influenced prediction X?).

Summary Flow Diagram (Conceptual)

Raw Data -> Local OMS Packaging -> Feature Extraction + Trace -> Submission & Integrity Anchor -> Ingestion Standardization -> Virtual Cell Training -> Inference & Simulation -> Active Learning Suggestions -> (Optional) Incentive Distribution & Governance

Conclusion

CellDAO operationalizes a virtuous cycle: standardized data fuels better models; better models propose informative experiments; new data refines the system. The stack is intentionally modular so labs can adopt components incrementally while moving toward a fully integrated Virtual Cell ecosystem.