license: mit
language:
- en
tags:
- molecular-generation
- conditional-generation
- graph-tokenization
- chemistry
pretty_name: MOSAIC Conditional Motif Generation
MOSAIC: Conditional Motif Generation
Conditional generation samples + metrics for the three MOSAIC tokenizers (SENT, HDT-MC, HDTC) prompted with each of 6 shared ring/aromatic motifs, across MOSES, GuacaMol, and COCONUT.
The conditioning prompt for each model is the standalone-motif tokenization, with closing tokens stripped so the model continues from "the molecule starts with this motif." For HDT-MC, the outer ENTER block is left open so the model can add additional communities; for HDTC, the trailing super-graph block is stripped so the model can add more typed communities.
Layout
{dataset}__AVG.png— per-dataset metric table averaged over the 6 motifs{dataset}__AVG__grid.png— per-dataset 6×5 visual grid (one row per motif, one column per model, plus reference + motif primer){dataset}/_reference/{motif}.txt— k motif-containing reference SMILES drawn from training (used as the comparison set for SNN / Frag / motif-distribution MMDs){dataset}/{motif}/{dataset}__{motif}.png— per-cell metric table{dataset}/{motif}/{dataset}__{motif}__grid.png— per-cell 5-column visual grid (Reference | Motif Primer | SENT | HDT-MC | HDTC){dataset}/{motif}/{model}/generated_smiles.txt— k=100 conditional generations{dataset}/{motif}/{model}/generated_metadata.json— gen-time stats (prompt token sequence, decode rate, motif retention, sampling params, elapsed){dataset}/{motif}/{model}/metrics.json— full computed metrics
Motifs
| Name | SMARTS / SMILES |
|---|---|
benzene |
c1ccccc1 |
pyridine |
c1ccncc1 |
naphthalene |
c1ccc2ccccc2c1 |
indole |
c1ccc2[nH]ccc2c1 |
cyclohexane |
C1CCCCC1 |
cyclopentane |
C1CCCC1 |
Datasets
- moses — MOSES drug-like (training set ~1.6M)
- guacamol — GuacaMol drug-like (test split fallback, ~941 reference SMILES)
- coconut — COCONUT natural products (training set ~10K)
Models
Listed in the column order used by the rendered tables (flat → unsupervised hierarchies → supervised hierarchy → typed hierarchy):
- sent — flat random-walk tokenizer (SENT)
- hdt_lou — HDT with Louvain community coarsening (HDT-Lou)
- hdt_hac — HDT with HAC-avg community coarsening (HDT-HAC)
- hdt — HDT with motif-community coarsening (HDT-MC)
- hdtc — HDT-Compositional, typed two-level (RING / FUNC / SINGLETON)
All models are gpt2xs (~11M params) trained with next-token prediction. Per-cell sample budget: k=100 decodable molecules per (dataset, model, motif), obtained via rejection sampling capped at 800 attempts.
Pipeline
Generation/metric/render code lives in the MOSAIC repo at
scripts/conditional_motif/.