Specify the transition
Represent the cellular source state and the desired target state without assuming a single correct intervention identity.
source → targetVCDesign / finite-budget intervention design
Design the experiment, not just the prediction.
Rank a feasible intervention set toward a desired cellular state under a finite experimental budget.
VCDesign evaluates models by the utility of the experiments they prioritize. VCDesign-CED uses historical perturbation measurements and biological knowledge to infer effects for candidates whose own responses are unavailable to its effect atlas.
Framework overviewFigure 1
Core principle
Useful outcomes, not identity recovery.
01 / Framework
Make the downstream experimental decision explicit.
A design problem starts with a source state, a target state, the candidates that can actually be tested, and the number of experiments the budget permits. VCDesign turns those ingredients into one shared ranking interface.
Rank feasible candidates
Apply one shared scorer to a variable candidate set, then return the top-B prefix allowed by the experimental budget.
candidates → rankingGrade the decision
Exclude the query identity and evaluate selected candidates using independently measured held-out outcomes.
outcomes → utility
02 / Information regime
Choose the solver that matches what is known at decision time.
Measured responses provide direct evidence when they exist. CED addresses the prospective setting in which a candidate must be prioritized before its own perturbational response is available.
Measured-response regime
Use the observed candidate effect directly.
When candidate responses have already been profiled, response-profile retrieval is the more direct source of evidence for alignment with the requested transition.
candidate response→target alignment→ranked candidates
Candidate Effect Distillation
response-unseen deployment
Historical atlasmeasured perturbation responses
+
Candidate knowledgeSTRING and MAP-KG
→
Predicted effectridge map in a response basis
→
Design scorebase compatibility + effect alignment
03 / Evaluation
Separate context transfer from response-unseen transfer.
VCDesign is evaluated across four CRISPRi Perturb-seq contexts. The protocol keeps the candidate pool, response availability, held-out outcomes, budget, and identity exclusions explicit.
B
One ranking, multiple experimental budgets.
The top-B prefix makes the budget part of the decision rather than an afterthought. Outcome-based evaluation then measures the utility of the experiments selected under that budget.
04 / Software
Run the complete synthetic path without downloading an atlas.
The bundled demo checks response-basis construction, ridge effect prediction, missing-knowledge handling, and score fusion. Its outputs are software checks, not biological results.
git clone https://github.com/Boom5426/VCDesign-CED.git
cd VCDesign-CED
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[model,data,dev]"
python examples/ced_demo.py
python -m pytest -qExpected repository test result: 162 passed.
05 / Data & artifacts
Frozen inputs and run records live on Hugging Face.
The companion dataset contains processed feature matrices, four-context packs, comparator scores, the selected checkpoint, evaluation records, and a SHA-256 manifest.
HF Hugging Face datasetBoom5426/VCDesign ↗01
Processed inputsFrozen response, STRING, MAP-KG, and candidate-knowledge arrays.
02
Reference checkpointThe selected epoch-8 final-clean model checkpoint.
03
Run recordsSelection, evaluation, external-baseline, and verification manifests.
04
Integrity manifestByte sizes and SHA-256 digests for the released snapshot.
VPublic manuscript
06 / Paper
VCDesign: Finite-Budget Intervention Design for Virtual Cells
The manuscript defines the design task, response-availability regimes, evaluation protocol, comparator settings, and controlled diagnosis of transferable design.