VCDDVirtual-cell-driven drug design

THE SCIENCE BEHIND THE DESIGN

Research

From a biological objective to a traceable molecular decision.

Explore the ideas behind state-first drug design, examine representative molecular outcomes, and follow the evidence through an actual feedback-guided search.

New design VCDD designs

RESEARCH APPROACH

Start with the cellular state, not a structure alone.

VCDD treats a desired transcriptional state as an objective for chemical-space search. A target expression profile, cell identity and treatment condition define what a molecular proposal should achieve. Virtual-cell queries connect these proposals to their predicted consequences, and feedback guides which candidate to evaluate next.

StateMol generates the feasible pool. PerturbCast predicts transcriptomic responses, while KillRead supplies source-specific sensitivity readouts from fixed cross-dose features. VCellGuide organizes the sequential search around both objectives.

CORE CONTRIBUTIONS

Three ideas shape the design loop.

01 / STATE AS OBJECTIVE

Condition and evaluate against the same biological aim.

The target state guides generation and also provides the reference for response alignment. A plausible structure is a starting point, not the final design objective.

02 / COMPLEMENTARY OBJECTIVES

Keep cellular alignment and sensitivity distinct.

Transcriptomic agreement does not establish killing activity. VCDD retains both objectives and source-specific evidence instead of treating a single scalar as sufficient.

03 / LOCAL ADAPTATION

Use feedback without retraining the virtual cell.

Observed queries correct local selection estimates within each condition. The molecular pool, generative model, virtual-cell predictors and global prior remain fixed; only query-local heads adapt.

REPRESENTATIVE FINDINGS

Evidence at the molecular and search levels.

GENE-STATE CONCORDANCE

0.908

VCDD-01 · VCAP · 5.01 µM · 24 h
Pearson r across 978 genes · Figure 6

PATHWAY CONCORDANCE

0.949

VCDD-02 · VCAP · 4.75 µM · 24 h
Pearson r across 101 pathways · Figure 6

QUERY BUDGET

8 / 128

Eight evaluations per readout source
Fixed 128-candidate pool · Figure 5 protocol

Molecule-specific concordances are computational predictions under different conditions, not an average drug-efficacy claim. The query budget is not a measure of total computational savings.

MOLECULES & CELLULAR CONSEQUENCES

VCDD molecular records

Figure 6 · computational records
10 molecular records978 gene-level readouts101 pathway-level readoutsVCAP & HT29 · 24 h

VCDD-01

Loading record
VCDD-01 molecular structure
Inspect in VCDD

Target & predicted cellular state

Gene-state concordance—Pearson r · 978 genes
Pathway concordance—Pearson r · 101 pathways
Target statePredicted state

Source-specific sensitivity percentiles

Archived model predictions, not measured drug activity. Sensitivity integrates fixed dose features; it is not a single-dose killing probability.

THE PAPER EXPERIMENT

Detailed computational replay

One cellular condition. A fixed molecular pool. Eight traceable decisions.

Archived replay · Figure 5j
  1. 01Target state
  2. 02Generated pool
  3. 03Virtual evaluation
  4. 04Feedback selection
  5. 05Recommendation

Generated pool

Loading the archived experiment…

Recorded decisions

0 / 8 queried
State match Pearson r—
Killing readout GDSC1+2 · model prediction—
Observed queryCurrent candidate
TARGET

A target cellular state

Step 1 / 11 0 recommendations

Recorded model predictions, not measured drug activity. This condition-specific result is not a population-average effect or a synthesis clearance.

Experiment record & data
Download replay data

Live inference · not connected

New designs run on the model server. Archived replay is independent of that service.

Open design tool