Dual-conditioning campaigns for pretrained 3D molecular diffusion

This dataset contains the canonical generation outputs and trajectory-level results supporting a dual-conditioning study of a pretrained pocket-conditioned three-dimensional molecular diffusion model. The task combines hard inpainting of a seven-heavy-atom substructure from ligand A with inference-time control toward the global three-dimensional shape of ligand B. The original single-shot campaign comprised three SARS-CoV-2 Mpro ligand pairs, five guidance strengths, and five random seeds, giving 75 runs and 691 pair-specific evaluations. Deduplication of shared unguided baselines yielded 649 unique generated structures. Of the 648 structures that could be chemically sanitized, 555 were fragmented. No guided structure satisfied the connected strict-dual criterion under either Shape Protrude convention. Version 2.0.0 preserves the complete original dataset and adds two adaptive staged-growth campaigns designed to improve global shape control while maintaining molecular connectivity. The variable-increment campaign evaluates greedy and beam search over the number of atoms requested at each growth stage. It contains two ligand pairs, ten paired random seeds, two search variants, and 40 complete runs. Thirty-nine runs reached the graduation criterion, while one terminated after all candidates were eliminated. The centered soft-scaffold campaign evaluates greedy and beam search over stage-dependent scaffold-retention values while using a fixed requested increment of four atoms. It contains two ligand pairs, ten paired random seeds, two search variants, and 40 complete runs. Thirty-six runs reached the graduation criterion, while four terminated after all candidates were eliminated. The record includes generated and retained SDF structures, reference ligands, campaign manifests, per-run beam states, complete trajectory tables, run-status records, raw stage-level candidate evaluations, connectivity and shape metrics, retained anchors, final structural exports, restrained-minimization outputs from the original campaign, environment information, documentation, file inventories, and checksums. Lightweight derived tables and the source code required to generate manifests, launch the campaigns, test the adaptive-search logic, and reproduce the reported analyses are distributed separately with software release v2.0.0 of the associated GitHub repository.