Spirographic AI LLC
The same molecule.
Different biological environments.
Different outcomes.
A drug does not encounter the same conditions everywhere in the body. Spirographic AI models molecular structure and evaluates biological interactions in the context of the organ, membrane, and transporter environment—not simply whether a molecule binds.
Deterministic molecular modeling. CPU-based computation. No machine-learning training data.
Built to be tested. Measured against evidence.
These results come from defined benchmark sets—not marketing estimates. Each metric represents a specific test and should be read with its methods and limitations.
Molecular structures
Successful builds across 13 structural classes. All 112 completed in approximately 119 seconds on one CPU; eight additional complex-drug builds were also successful.
Molecular modeler report ↗Cardiac risk classification
96.3% three-category CiPA torsade-risk classification, including identification of all eight high-risk reference drugs. This model uses measured ion-channel data.
CiPA validation report ↗Blood–brain barrier
324 of 348 transporter classification predictions correct across eight transporters in the reported internal benchmark.
Transporter validation ↗Hepatic transporters
306 of 325 classification predictions correct across nine transporters in the reported internal benchmark.
Transporter validation ↗What these numbers mean: Reported internal retrospective validation results, not independent prospective clinical validation. Structure-build success, cardiac-risk classification, and transporter substrate classification are different endpoints and are not directly comparable.
A transporter is not just a yes-or-no interaction.
A molecule may interact with a transporter, but that alone does not tell us where the molecule goes. The same transporter can contribute to uptake or export depending on where it sits and which direction it moves material across a membrane.
Molecular Modeler VB12 constructs a deterministic three-dimensional structure from SMILES.
Evaluate interactions in the context of the organ barrier, membrane setting, and modeled transporter behavior.
Distinguish influx from efflux at the relevant membrane, rather than treating transporter recognition as the entire answer.
Use those interactions to inform hypotheses about movement, retention, and clearance across physiological compartments.
SMILES → deterministic 3D model
Conceptual illustration — not a measured transport experiment
Validation boundary: The published transporter benchmark figures assess classification against reference substrate data. Predictions of net organ accumulation, retention, or excretion require their own dedicated validation; they should not be interpreted as established by classification accuracy alone.
Connected modeling across biological systems.
Build the structure. Investigate transformation.
Deterministic 3D construction provides a common molecular foundation for downstream modeling, including CYP3A4 metabolite prediction. A reported 20-substrate development benchmark achieved 20/20 top-1 matches; broader validation sets must be considered separately.
Consider where the molecule encounters biology.
Organ-specific transporter panels investigate barrier interactions. Separate receptor and cardiac models address other safety questions, with distinct inputs, endpoints, and validation requirements.
The platform is a set of connected computational models; not every downstream module uses SMILES as its only input. For example, the CiPA cardiac-risk model uses measured ion-channel inhibition and exposure data.
Real results deserve independent testing.
Spirographic AI is seeking external research and pharmaceutical collaborators with existing experimental data to evaluate predictions on molecules not used in model development.
Collaborators can provide SMILES for compounds with withheld experimental outcomes. Predictions can then be scored independently, with test definitions and limitations agreed in advance.
Measured against gas-phase experiment.
Bond lengths from deterministic 3D molecular builds compared directly with experimental gas-phase geometries reported in the NIST Computational Chemistry Comparison and Benchmark Database (CCCBDB). This is not a comparison against computed PubChem conformers.
Absolute bond-length difference by molecule (Å)
Each bar is the mean absolute difference between VB12 and NIST experimental bond lengths for that molecule. Lower is closer to experiment. Multiple measured bonds within a molecule are averaged.
Scale: each track represents 0.030 Å. Source: Molecular Modeler VB12 validation report, “Agreement with experiment,” 35 bond comparisons across 22 molecules. Values are calculated from the displayed VB12 and NIST bond lengths, which are rounded to four decimal places.
These measurements assess bond geometry, not docking pose accuracy or biological predictive performance. The experimental benchmark covers 22 molecules; the 112/112 figure refers to a separate structure-build test.
Read the full molecular modeling validation report ↗