Distribution-level virtual cell
Predict a population of possible cell states—not one average vector.
The scientific object is a perturbation-conditioned cell distribution: who responds, which states expand or disappear, how uncertain the transition is, and when the model must abstain.
Prediction object
One input population becomes five auditable outputs.
Current boundary: this page defines the future output object and evaluation rules. It does not generate synthetic DMD cells or a therapeutic prediction.
Model strategy
Build in three layers; never skip the baseline layer.
Statistical and hybrid baselines
Zero, train mean, pseudobulk, ridge and nearest neighbour establish whether a complex model adds information.
Current state: limited same-context ridge runConditional state transport
Predict movement from a control population to a perturbed population while conditioning on gene, cell state and context.
Current state: no qualifying DMD training truthDistribution-generative model
Flow or diffusion can represent heterogeneous responders, but only after cell-level truth and leakage-safe OOD splits exist.
Current state: architecture watch onlyMetric firewall
Seven metric families prevent one convenient score from defining success.
Mean fidelity
MAE · RMSE · correlationChecks average expression without treating it as the whole response.
Differential response
DES · signed LFC · AUPRCTests whether perturbation-responsive genes and direction are recovered.
Perturbation identity
PDS + scale diagnosticsTests discrimination while guarding against amplitude manipulation.
Population geometry
MMD · energy · Wasserstein / OTCompares predicted and observed cell distributions.
State composition
state-proportion error · divergenceTests responder fractions and shifts in cell-state occupancy.
Uncertainty
coverage · Brier · abstentionTests whether confidence is calibrated and refusal is useful.
DMD utility
hit rate · replication · toxicity miss rateConnects molecular prediction to disease-relevant function and safety.
PDS safeguardRaw PDS is never reported alone.
- Report raw, truth-norm-matched and prediction-norm-matched variants.
- Run amplitude scaling and normalization sensitivity checks.
- Publish DES, MAE and distribution metrics beside PDS.
- Freeze preprocessing and distance definition before scoring.
Arc Virtual Cell Challenge watch
Confirmed facts are separated from launch-day unknowns.
What the first challenge established
- More than 5,000 registrants and over 300 final submissions.
- Hybrid deep-learning plus classical statistical features led the final rankings.
- Models did not consistently beat naive baselines across every metric.
What is confirmed now
- Launch: Thursday, 20 August 2026.
- Round two will use a new prediction problem and a wider range of metrics.
- Grand prize: US$100,000.
What remains unknown
- Training and test datasets.
- Exact perturbation, context and output contract.
- Metric formulas, weights and submission format.
What NMD-VCell will do
- Freeze official files and checksums.
- Map task fields to the local contract.
- Run leakage and baseline audits.
- Register compatible and incompatible fields.
What can be used now
Adopt the evaluation architecture before training the generator.
Standardize future cell-level predictions and force uncertainty plus abstention.
Protect against scale-sensitive PDS and misleading single-metric gains.
Keep pseudobulk, ridge and nearest-neighbour comparisons permanent.
No released DMD candidate-perturbation cell distribution exists for calibration.
Primary-source ledger
Published facts, official announcements and preprints keep different labels.
- Arc metric explainer · official source
- Arc 2025 challenge wrap-up · official source
- PRiMeFlow · preprint, author-reported results
- Lingshu-Cell · preprint, author-reported results
- PDS sensitivity analysis · preprint