Data: preserve cellular context
Link expression matrices, cell labels, interventions, doses, time points and batches. Document sample-level relationships between modalities.
Explore validation methods ↗AI VIRTUAL CELL · TECHNOLOGY
From cellular data to testable predictions.
Organize data and models around a research question, with explicit inputs, outputs and evaluation criteria that support traceable analysis.
RESEARCH ARCHITECTURE
Four stages. One evidence chain.
This is a research framework. Model selection, data modalities and deliverables are defined after project-specific assessment.
Representation: create comparable state spaces
Choose state representations appropriate to the task and examine cell types, individual variation and batch effects to distinguish technical signals from biological differences.
Explore validation methods ↗Prediction: define intervention-specific outputs
Study expression changes, state transitions or phenotypic responses for a specified cell background and perturbation, and assess the model's scope of applicability.
Explore validation methods ↗Validation: return predictions to experiments
Test predictions with held-out data, baseline comparisons and independent experiments. Record model versions, data provenance and failure conditions.
Explore validation methods ↗INPUT → QUESTION → OUTPUT
Define the question before choosing the model.
Research data
Expression matrices / Images / Metadata / Controls
Biological context
Cell type / Origin / Culture conditions / Time points
Computational task
State comparison / Perturbation response / Candidate prioritization
Validation outputs
Evaluation / Limitations / Experimental recommendations
Data preparation checklist
| Data category | What to prepare | Key checks |
|---|---|---|
| Single-cell data | Expression matrices, gene identifiers, cell and sample annotations | QC criteria, sequencing platform, batch and origin |
| Intervention metadata | Compound or gene identifiers, dose, time and controls | Consistent condition names and identifiable replicates |
| Phenotypes and imaging | Raw readouts, acquisition parameters and label definitions | Explicit endpoints and traceable sample matching |
| Permissions and governance | Permitted uses, sharing restrictions and de-identification status | Confirm usage rights and avoid uploading sensitive data |
MODEL EVALUATION PROTOCOL
Define the task and its evaluation
Align prediction targets, data splits and biological readouts in one study plan. Model selection, evaluation metrics and acceptance criteria are agreed before analysis.
| Task | Data split | Evaluation focus |
|---|---|---|
| Cell state comparison | Hold out donors or biological samples and inspect batch confounding | Cell-type preservation, state stability and independent marker validation |
| Drug perturbation response | Evaluate seen contexts, unseen compounds and unseen cell contexts separately | Changes relative to controls, differential genes and pathway direction, compared with simple baselines |
| Dose and time analysis | Distinguish interpolation within observed conditions from extrapolation | Dose trends, temporal consistency and independent phenotype endpoints |
Record input versions, QC rules, data splits, baseline settings, seeds and evaluation scripts so each research result can be reproduced and reviewed.
PROJECT SERVICES
Define deliverables and acceptance criteria
Agree on the work package before starting analysis.
Traceable inputs
Agree on the sample manifest, data dictionary, QC decisions and permitted uses before analysis.
Reviewable results
Specify analysis tables, figures, methods, baseline comparisons and limitations in the deliverable list.
Reproducibility and handover
Agree on script availability, software and model versions, environment records, intellectual property and handover format. Third-party licenses remain applicable.
START WITH A QUESTION
Begin with a cellular question.
Clarify the objective, data and validation pathway so predictions can inform the next experiment.