๐Ÿงซ JCVI-syn3A ยท Minimal cell
๐Ÿฆ  E. coli ยท Min oscillation
๐Ÿง Connected Human 3D
โš–๏ธ Fork Test
๐Ÿงฌ Platform
KO|EN
Language
โš–๏ธ Fork-speed consistency test
Two cells, both doubling in the measured 105 min, differing only in replication fork speed. The constant in use does not reproduce the measured chromosome ratio โ€” but a faster fork is not the only way to close that gap.
Cell-cycle clock
0.0 min of 105
A โ€” the constant the field uses
fork speed100 bp/s
replication period C45.2 min
predicted ori:ter1.348
B โ€” a fork fast enough to match
fork speed157 bp/s
replication period C28.9 min
predicted ori:ter1.210
Measured target โ€” sequencing of exponential culture
1.00 stationary1.21 measured1.40
A overshoots the measurement by +0.138; B lands on it. The gap is real and sits inside the field's own parameter set. Its cause is not settled. A faster fork is one reading. The other: cells that are not replicating carry one origin and one terminus, contributing 1.0 โ€” which is exactly what the stationary-phase control measures โ€” so a culture that is roughly 40% idle gives the same 1.21 with the fork left at 100 bp/s. Bulk sequencing cannot separate the two; a single-cell measurement can.
Second inconsistency โ€” ribosome census
ribosomal-protein copies measured10 โ€“ 1001
coefficient of variation0.83
A ribosome carries one copy of each subunit, so these should be equal. They are not, and the repeated floor value marks a detection limit. A count taken from this distribution is not a measurement. With the literature 500 ribosomes and 12 aa/s, an active fraction of 0.54 reproduces 105 min exactly โ€” ordinary for a slow grower, and the one quantity nobody has measured here.

The Differentiable
Virtual Cell

JCVI-syn3A rebuilt in a GPU. Forward = simulate life ยท Backward = design it.
Private ยท R&Dโ“˜ What is this?
๐Ÿงช Live Experiment Console
Inject a perturbation โ†’ the differentiable cell responds โ†’ measure. Every number is our model's real output.
โ‘ข Measure โ€” live readout
ATP productiondifferentiable whole-cell model
3.99 mM/s
ATP yield
1.94
B
Doubling
117m
โ€”
ori : ter
1.31
โ€”
Against public syn3A: ori:ter 1.31 vs 1.21, doubling 117 vs 105 min. No fitting โ€” but both grades were withdrawn on review, so these are reported unscored.
โ‘  Targetโˆ‚ backward ยท inverse design
ENO1.44
PGK1.20
NOX0.37
ACKr0.16
The model computes โˆ‚ATP / โˆ‚enzyme by autodiff through the whole cell (FD-verified to 10.6 digits). It ranks the highest-leverage target โ€” a forward-only simulator cannot.
โ‘ก Inject & runโ–ถ forward ยท simulate
๐Ÿ’Š Inhibit ENO (+1 kcal/mol)
โ†บ Reset
Baseline steady state ยท ATP 3.99 mM/s. Pick a target and inject.
Enzyme barrier from quantum (VQE/DFT) energetics โ†’ Eyring kcat โ†’ whole-cell steady state. No real quantum hardware run (ฮธ is a placeholder); the coupling & sensitivity are FD-verified.
๐Ÿฆ  E. coli โ€” Min system
A self-organizing protein oscillation that tells the cell where to divide โ€” the textbook case where space matters.
Oscillation period
41.7 s
MinD sweeps pole-to-pole; its time-average is minimal at midcell โ†’ marks the division plane. Inside the measured 40โ€“120 s range โ€” predicted, not fit (Huang-Meir-Wingreen 2003).
Why it's oursโˆ‚ differentiable
We differentiate through the pattern (autodiff vs FD, 7โ€“8 digits) โ†’ inverse-design a spatial pattern. Stochastic 4D simulators cannot. Physics gates (conservation, positivity, convergence) all pass โ€” internal checks, not a comparison with measurement.
๐Ÿ”ด Scope: deterministic mean-field (real Min is stochastic, ~4000 molecules) ยท 1D long-axis ยท published params ยท the Min subsystem, not a whole E. coli cell.
๐Ÿง Connected Human ยท live
Real anatomical organ meshes. Choose a sex, inject a public molecule โ€” the mechanism propagates through the shared bloodstream.
Biological sex ยท organ set
โ™‚ Male
โ™€ Female
Inject molecule ยท public
๐Ÿ’Š Sildenafil PDE5 inhibitor
๐Ÿ’Š Chondroitin sulfate cartilage GAG
Scan
Activate
Inhibit
Benefitโ€“risk adjudication ยท 86% on 5 frozen cohorts
Adjudicate
โ†บ Reset
Healthy baseline
Male organ set ยท all organs at rest.
Mechanism-level propagation of our whole-body model โ€” not clinical efficacy. ๐ŸŸข benefit ยท ๐Ÿ”ด harm ยท ๐ŸŸ  stress ยท ๐Ÿ”ต neutral.
Data credits ยท
VIDRAFT ยท Differentiable Cell Platform

One differentiable engine โ€”
from a minimal cell to human disease to our own drugs

Every figure below is our model's own output โ€” forward-simulated and gradient-verified, with limits stated. Letter grades appear only where the reference carries a documented uncertainty; where it does not, the figure is marked unscored rather than given a letter. Self-correcting agentic research.
๐Ÿงซ Minimal cellโ†’๐Ÿฆ  Bacteriaโ†’๐Ÿง  Human diseaseโ†’๐Ÿ’Š Our drugsโ†’๐Ÿซ Organ metabolism
Virtual trial = whole-body sim ร— thousands of virtual patients
๐Ÿงซ
JCVI-syn3A minimal cell
Differentiable whole-cell of the smallest genome (493 genes); it replicates & divides.
cell cycle  replicates & dividesnot scored
No fitting. Prototype; deterministic, not 4D-stochastic.
๐Ÿฆ 
E. coli โ€” Min oscillation
Pole-to-pole protein wave that sets the division site; a differentiable spatial pattern.
period 41.7 s  in 40โ€“120 s rangeA
Published model, not re-fit. Mean-field, not stochastic.
๐Ÿง 
Human disease โ€” protein aggregation
Differentiable neurodegeneration aggregation cascades (Alzheimer-type), with mechanism-based screening.
validated vs published kinetics A
Parameter-free scaling law reproduced (no fit); mechanism-plausibility, not efficacy.
๐ŸŽฏ
Drug mechanism screening
Gradients rank which molecular step a therapy must hit โ€” and flag counterproductive ones โ€” across our disease models.
โˆ‚(outcome)/โˆ‚(mechanism)  FD-verified
Capability shown on public targets; effects are uncertainty bands, not efficacy. Specific compounds & targets are private.
๐Ÿซ
Genome-scale human metabolism
Differentiable human liver metabolism on a genome-scale reconstruction (~10,600 reactions); fatty-liver phenotype emerges from mass balance.
genome-scale  ~10,600 reactionsdirectional
Constraint-based metabolism, not a whole cell; pinpoints the highest-leverage disease step.
๐Ÿ•ฐ๏ธ
Universal clock (reproduction)
Reproduced the equant "universal dynamical clock" method; verified on a Kepler orbit.
31,531ร— more uniform null on ours
Null result: no gain on our symmetric oscillators. Core reproduced, not the full paper.
๐Ÿซ
Lung โ€” fibrosis tipping point
A bistable switch explaining why fibrosis becomes progressive & self-sustaining; differentiable drug-lever contrast (reversal vs prevention).
bistable + hysteresis A
Published-mechanism model; drug effects are bands, not efficacy.
๐Ÿซ€
Heart โ€” HFpEF energetics
The failing heart as an "engine out of fuel": energy reserve (PCr/ATP) drops and collapses under stress.
reserve collapse A
Metabolism only (no electrophysiology); mechanism-plausibility.
๐Ÿง 
Brain โ€” metabolism โ†’ tau circuit
Glucose hypometabolism โ†’ protein-modification stress โ†’ tau, integrated into one mechanism circuit linking energy to neurodegeneration.
integrated circuit A
Causality debated; coupling is a labelled assumption (band).
๐Ÿซ˜
Gut โ€” axis hub
Enterocyte metabolism + leaky-gut bistability; quantifies gutโ†’liver & gutโ†’brain signals that connect the organs.
gut-liver-brain axes A
Not a microbiome ecosystem; axis couplings are bands.
๐Ÿซ˜
Kidney โ€” CKD (metabolicโ†’fibrosis)
A tubular fatty-acid-oxidation defect lowers the fibrosis threshold; the metabolic lever can reverse it.
metabolic driver A
Reuses the fibrosis switch; drug effects are bands.
๐Ÿฅž
Pancreas โ€” T2D + direction check
Beta-cell glucoseโ†’insulin, plus a check that caught a wrong-direction drug mechanism โ€” which binding/docking alone cannot see.
wrong-direction caught A
Mechanism-plausibility; context-dependent (band), not clinical.
๐Ÿฉธ
Blood & immune
RBC oxygen metabolism + macrophage inflammation โ€” grounds the systemic inflammation shared across every organ model.
grounds inflammation A
Two cell types, not the whole immune system; bands.
๐Ÿ”—
Connected human โ€” whole-body integration
All 7 organs coupled through a shared bloodstream + physiological axes: a perturbation in one organ propagates body-wide, and a whole-body differentiable inverse-design ranks the best systemic intervention.
7 organs ยท 1 body A
Coarse mechanism coupling โ€” NOT a complete human, not PK/PD; self-flags where it doesn't match clinical biology.
๐Ÿ”ฌ
Whole-body drug trial (QSP)
Dose a drug into the whole body โ†’ see efficacy AND off-target effects across organs at once; differentiable therapeutic-window & tissue-selectivity optimization.
efficacy-safety window A
Calibration/PK-PD prototype; mechanism bands, not a clinical/safety prediction.
๐Ÿ‘ฅ
In-silico virtual clinical trial
Sample the whole-body model into a virtual population (~5,000 virtual patients with individual variation) โ†’ predict responder fraction, the responding subpopulation (biomarker), and the safety-event distribution.
responder % ยท go/no-go
Turns a candidate ranking into a development decision. Virtual-population prototype.
๐Ÿงฌ
Digital twin (personalized)
Personalize the whole body to an individual, calibrate to their data โ†’ their own optimal therapy. The same drug helps one person but not another.
per-person optima A
Concept prototype; synthetic individuals; identifiability limits apply.
๐Ÿ”’ Private R&D preview ยท research prototype ยท mechanism-plausibility models, not efficacy or clinical predictions ยท method patent-pending.
Molecular Legend
Plasma membrane
Circular chromosome ยท ori/ter
Ribosomes ยท ATP energy core
Proteins ยท enzyme sites
Membrane
Chromosome
Ribosomes
Crowding
Auto-rotate
โ–ถ Cell cycle
Cell cycle  Replication
VIDRAFT Virtual Cell Platform ยท Quantum โ†’ Kinetics โ†’ Whole-cell
oriC replication origin
ter terminus ยท ยฝ genome
ENO enzyme target
VIDRAFT ยท Virtual Cell Platform

Not a simulation you watch.
A cell you can run backward.

JCVI-syn3A โ€” the smallest genome that builds a living cell โ€” rebuilt inside a GPU as a differentiable model. Run it forward to simulate life; run it backward to compute exactly what to change.
493-gene genome ยท differentiable end-to-end
cell cycle from real genome coordinates โ€” unfitted, unscored
quantumโ†’cell gradients ยท 10.6-digit verified
๐Ÿ—บ๏ธ
Mechanistic whole-cell (SOTA)
A precise map. But static โ€” to ask "what if?" you re-run blindly. Not differentiable, no inverse design.
๐Ÿงญ
AI cell models
A compass with no map โ€” data-driven, no real physics. Struggles to beat a linear baseline on perturbations.
๐Ÿ›ฐ๏ธ
VIDRAFT โ€” map + GPS
Real mechanism you can differentiate: it tells you which knob to turn and re-routes instantly. Forward AND backward.
Enter the cell โ–ถ
Research prototype ยท differentiable minimal-cell method patent-pending ยท every figure is our own validated model output.
Assembling minimal cell
493 genes ยท 543 kbp ยท placing ribosomesโ€ฆ