Molecular Modeler VB12 Validation Report

Section 01

Executive summary

Molecular Modeler VB12 V12 built a full 3D structure for each of 112 molecules spanning 13 structural classes, from water to strychnine, starting from nothing but each molecule's SMILES string. Every structure is a single deterministic build with no training data. The builds were scored against two references: experimental gas-phase geometries from the NIST CCCBDB for 22 molecules, and PubChem 3D computed structures for all 112. Against experiment, bond lengths came out within 0.56% and bond angles within 0.52° on average. All 112 molecules built successfully, in 119 seconds total on a single CPU with no GPU.

A separate pressure test then built eight large, complex drugs, from doxorubicin (38 heavy atoms) to vancomycin (101 heavy atoms), including the macrocycles cyclosporine A, rapamycin and rifampin. All eight built, with median bond-length error under 0.24% for every drug; vancomycin, the largest, finished in 56.6 seconds.

The headline 0.56% averages each molecule's error first, then averages across the 22 molecules. Averaging all 35 individual bond comparisons directly gives 0.58% (median 0.61%). Angles are handled the same way: 0.52° per molecule, 0.55° across all 14 angle comparisons. RMSD compares one deterministic Molecular Modeler VB12 structure to one PubChem structure, with no ensemble and no best-of selection.

Section 02

Agreement with experiment

NIST CCCBDB gas-phase geometries are measured, not computed, which makes them the strictest reference available. Every bond comparison but one fell within 1.3% of the measured value.

35 bond comparisons across 22 molecules: mean 0.58%, median 0.61%, largest 1.72% (toluene ring-to-methyl C–C).

Bond angles

14 angle comparisons across 13 molecules: mean 0.55°, median 0.27°. Methane, benzene and cyclopropane match exactly.

Section 03

Results across 112 molecules

Mean bond error stays between 0.5% and 1.4% in every class, including steroids and the ergoline family. RMSD rises with size and flexibility, as expected for a single structure compared against a single computed conformer.

By molecular weight

Bond-length error stays close to 1% as molecules grow; the larger RMSD in heavier molecules mainly reflects more ways for a flexible chain to sit in 3D, not bond errors.

Section 04

Large-molecule pressure test 8 / 8 built

Eight large drugs with fused, bridged and macrocyclic ring systems were built in a separate run on the same day. Vancomycin, with 101 heavy atoms and 12 rings, is the largest molecule Molecular Modeler VB12 has built to date.

Rifampin had the lowest median error (0.014%) and Taxol the lowest mean (0.033%). Where a maximum is high, it comes from a single bond (for vancomycin and rapamycin, a macrocyclic ring closure); the median for every drug stays at or below 0.24%. Taxol and digoxin were also built in a September 28 run, and their results were stable between the two runs.

Section 05

Limitations

Reference structures

Only 22 molecules have experimental gas-phase geometries in NIST CCCBDB. The broader 112-molecule comparison uses PubChem 3D structures, which are computed with the MMFF94 force field rather than measured, so part of the difference against them may come from the reference itself.

Single structure, not an ensemble

Molecular Modeler VB12 produces one deterministic structure per molecule. RMSD here is not comparable to benchmarks that report the best match from an ensemble of many generated conformers. Long, flexible molecules such as retinol (3.68 Å) and capsaicin (3.14 Å) show the highest RMSD for this reason.

Single-bond outliers in large drugs

In the pressure test, a few individual bonds carried much larger errors than the rest of the molecule: vancomycin 24.2%, digoxin 7.3%, rapamycin 5.8% and doxorubicin 3.2%. In vancomycin and rapamycin these sit at macrocyclic ring closures. Medians for these drugs remain low, but those bonds are the clearest target for further work.

Small sample sizes

Only 3 of the 112 benchmark molecules are above 400 Da, and the pressure test covers 8 drugs, so results for large molecules should be read as early evidence. RMSD was unavailable for 11 molecules, and vitamin B12 was not included in the October 1 pressure test.