EXTENDED DATA — CAPTIONS

============================================================

Extended Data Fig. 1
Extended Data Fig. 1 | DGLD pipeline overview. Encode (LIMO VAE) \(\to\) generate (conditional latent DDPM) \(\to\) guide (multi-task score model) \(\to\) filter (SMARTS, Pareto, GFN2-xTB, DFT). Tier-A/B labels drive the conditional gradient; Tier-C/D labels drive the unconditional CFG branch only.

Extended Data Fig. 2
Extended Data Fig. 2 | Pool-size scaling. Best composite score over top-1 and number of candidates passing every filter versus pool size from 1.5k to 100k. Both curves are still moving at pool = 40k; an M7 five-lane 100k run confirms the trend with 4,639 passing candidates (\(5.1\times\) more than the 40k baseline) and scaffold count expanding from 7 to 24.

Extended Data Fig. 3
Extended Data Fig. 3 | Classifier-free guidance scale sweep. CFG scale \(w \in \{5, 7, 9\}\) at pool = 8k per setting, ranked by post-filter yield over the two-denoiser pool. \(w = 7\) is the empirical sweet spot.

Extended Data Fig. 4
Extended Data Fig. 4 | Per-head sample-time scale sweep. Top-1 composite score (left panel) and top-1 detonation velocity \(D\) (right panel) for six head-condition variants at pool = 2,500 per denoiser: ung = unguided baseline; def = production default \((s_\text{viab}, s_\text{sens}, s_\text{hazard}) = (1.0, 0.3, 1.0)\); low = halved scales; v↑ = boosted viability scale; s↑ = boosted sensitivity scale; +SA = drug-domain SA gradient added (worsens the composite by 13%, retained at \(s_\text{SA} = 0\) in production). The default Hz-C2 condition (def) is the empirical optimum on both axes. Per-head training-loss curves over the 40k-step run are released alongside the model checkpoints on Zenodo.

Extended Data Fig. 5
Extended Data Fig. 5 | Self-distillation refinement of the viability head. Three rounds of mine-then-retrain on the score-model viability head only; the LIMO encoder and denoiser are frozen across rounds. Hard-negative count cumulates 0 \(\to\) 137 \(\to\) 918. The held-out probe stopping criterion (anchors \(\ge 0.86\) AND cheats \(\le 0.84\)) is satisfied at round 2.

Extended Data Fig. 6
Extended Data Fig. 6 | Twelve confirmed leads gallery. Per-lead RDKit 2D depiction, chemotype label, molecular formula, and 6-anchor-calibrated DFT/Kamlet–Jacobs values. Border colour: green = xTB HOMO–LUMO gap \(\ge 1.5\) eV; red = gap below threshold.

Extended Data Fig. 7
Extended Data Fig. 7 | Productive-quadrant scatter. Composite score \(S\) versus viability probability for the twelve DFT-confirmed leads, with the four no-diffusion baselines as reference markers. Marker area is proportional to drop-weight impact sensitivity \(h_{50}\). The green-tinted upper-right quadrant is the productive zone (novel and HMX-class).

Extended Data Fig. 8
Extended Data Fig. 8 | Merged top-100 source-pool provenance. Per-condition provenance of the merged top-100 across guided and unguided lanes, showing that 89/100 originate from the smallest unguided pool reranked by the Pareto scaffold composite; guidance acts as a high-precision lever on the top of the funnel rather than as the source of bulk Pareto coverage.

Extended Data Fig. 9
Extended Data Fig. 9 | Distribution-learning small multiples. Validity proxy, top-100 scaffold uniqueness, internal diversity, and Fréchet ChemNet Distance versus the labelled master, comparing SMILES-LSTM (red) against seven DGLD conditions (blue). Condition codes: C0 = unguided baseline; suffix letters denote which property heads steer sampling (v = viability, s = sensitivity, h = hazard, sa = synthetic accessibility), e.g. C2vsh = viability + sensitivity + hazard. Full per-head specifications in Supplementary Note 6.3.

============================================================

Extended Data Table 1
Extended Data Table 1 | Six-anchor DFT calibration.
 Anchor compounds and their chemotype families; experimental crystal density
 \(\rho_\text{exp}\) and condensed-phase heat of formation \(\Delta H_f^\text{exp}\);
 raw \(\omega\)B97X-D3BJ/def2-TZVP outputs \(\rho_\text{DFT}\) (Bondi-vdW + 0.69 packing)
 and \(\Delta H_f^\text{DFT}\) (atomisation-energy method, 0 K); calibrated values
 under the linear correction \(\rho_\text{cal} = 1.392\,\rho_\text{DFT} - 0.415\) and
 \(\Delta H_f^\text{cal} = \Delta H_f^\text{DFT} - 206.7\) kJ/mol. Joint-fit RMS over all
 six anchors is 0.050 g/cm^3 on \(\rho\) and 53.8 kJ/mol on \(\Delta H_f\);
 leave-one-out RMS is 0.078 g/cm^3 and 64.6 kJ/mol respectively. Per-anchor
 LOO residuals and the DNTF 7th-anchor exclusion are in Supplementary Note 3.4 and
 3.11.