Skip to content

Performance

All figures are measured at 100 denoising steps on 100,000 samples generated from the 1,000 CCDC Virtual Screening [4] reference compounds, in batches of 100 samples.

Core Model

24M EDM + 22M AdjMatSeer parameters

Metric Value
Avg time to generate 50 valid samples 11.46 s (NVIDIA H100)
Generation speed 4.18 valid molecules / s
GPU memory per generation thread up to 14.0 GB (float16, 39 atoms, 100 samples)
Avg Shape Tanimoto similarity 53.32% (basic) — 69.97% (Inertial Fragment Matching)
Max Shape Tanimoto similarity 99.69%
Avg chemical Tanimoto similarity (Morgan r=2, 2048 bits) 10.87%
Avg fragment integration success (IFM) 63.80%
Chemically novel vs. training set 99.84%
Valid after standardisation 48% (ML bond prediction) — 93% (OpenBabel bond prediction)
Unique in generated set 99.94%
Fréchet fingerprint distance ChEMBL 4.13 · PubChem 2.64 · ZINC-250k 4.95

Distilled Model

8.9M EDM + 1.6M AdjMatSeer parameters

Note: AdjMatSeer distillation was accompanied by fine-tuning to match OpenBabel bond perception to maximise validity. This may affect the general chemical quality of the generated structures.

Metric Value
Avg time to generate 50 valid samples 10.9 s (NVIDIA A40)
Generation speed 4.59 valid molecules / s
GPU memory per generation thread up to 8.6 GB (float16, 39 atoms, 100 samples)
Avg Shape Tanimoto similarity 50.5%
Max Shape Tanimoto similarity 88.65%
Avg chemical Tanimoto similarity (Morgan r=2, 2048 bits) 7.53%
Valid after standardisation 71.14% (ML bond prediction)

Choose the distilled weights when memory or throughput is the constraint, or when validity under ML bond prediction matters more than chemical fidelity to ChEMBL. Choose the core weights for maximum shape similarity and IFM workflows.

Generation Quality

PoseBusters validity [5]

PB-valid molecules: 91.33%

Check Failure rate
position 0.01%
mol_pred_loaded 0.0%
sanitization 0.01%
inchi_convertible 0.01%
all_atoms_connected 0.0%
bond_lengths 0.24%
bond_angles 0.70%
internal_steric_clash 2.31%
aromatic_ring_flatness 3.34%
non-aromatic_ring_non-flatness 0.27%

Synthesizability

Average SA Score [6]: 3.18 (scale 1 = easy to 10 = very difficult)

Practical Scaling Notes

  • Runtime scales linearly with diffusion_steps and roughly linearly with n_samples until the device saturates.
  • Memory is dominated by the EDM's edge tensors and scales with n_samples × max_n_nodes²; reduce the batch for large molecules on small GPUs.
  • resample_steps = k multiplies the denoising cost by approximately k + 1.
  • MMFF94 optimisation (optimize_geometry=True) runs on CPU per molecule and becomes noticeable for batches of hundreds.

Generation Examples

ex1 ex2 ex3 ex4

References

  1. CCDC Virtual Screening set — ccdc.cam.ac.uk
  2. Buttenschoen et al., PoseBusters — 10.1039/D3SC04185A
  3. Ertl & Schuffenhauer, SA Score — 10.1186/1758-2946-1-8