Overview
This project asks how virtual cell models can support biological understanding, rather than only produce strong benchmark performance.
Core Question
How can a virtual cell model help generate biological understanding instead of only producing impressive benchmark predictions?
Why This Matters
Prediction is useful when it reduces experimental ambiguity. It is weaker when accuracy becomes a substitute for explanation. Perturbation modeling is valuable only if it changes what experiment should be done next or clarifies why a response occurs.
Current Understanding
The current working distinction is between interpolation, extrapolation, and mechanistic counterfactual. A model may perform well in one mode while failing to support scientific explanation in another.
Evidence / Signals
- Performance under held-out perturbations.
- Ability to suggest interpretable experimental comparisons.
- Robustness across cell types, datasets, and perturbation classes.
- Cases where prediction points to a causal handle.
Tensions / Uncertainties
- Benchmark success may hide biological shortcuts.
- Perturbation labels can be too coarse to support mechanism.
- Explanation may require interventions that the model was not trained to represent.
Open Problems
- What would count as a mechanistically useful virtual cell prediction?
- How should model evaluation separate interpolation from true counterfactual reasoning?
- What experiment becomes cheaper or clearer because the model exists?
Related Notes
- Virtual Cell and Perturbation Prediction
- Mechanism-first vs Prediction-first Science
Future Directions
- Collect examples where predictions changed experiment design.
- Track evaluation frameworks that reward mechanistic usefulness.
Revision History
- 2026-07-07: Converted into the Research Project structure.