--- license: mit language: - en tags: - chemistry - molecule-representation - rmmol - molecular-property-prediction - qm9 - tox21 - sider pretty_name: RMMol --- # RMMol This bundle contains standardized CSV files and a Jupyter sample notebook for RMMol frozen-embedding examples across Biophysics, Physiological, Physical Chemistry, and Quantum Mechanics. ## Contents - `dataset/biophysics/`: BBBP full CSV and HIV train/val/test splits. - `dataset/physiological/`: BACE, ClinTox FDA-approved, Tox21 single-label splits, and SIDER single-label splits. - `dataset/physical_chemistry/`: ESOL, FreeSolv, and Lipophilicity splits. - `dataset/quantum_mechanics/`: QM9 splits. - `examples/RMMol_sample.ipynb`: RMMol-only sample notebook with the selected downstream-head settings. - `examples/RMMol_lipid_conformation_distance.ipynb`: RMMol-only LNPDB heldout IL_SMILES workflow for embedding distance versus RDKit USR conformation distance. - `MANIFEST.csv`: file-level provenance, row counts, columns, and checksums. ## Naming convention Each split follows `_.csv` for split datasets and `__.csv` for Tox21/SIDER single-task files. BBBP is stored as `bbbp.csv` because its protocol uses five-fold cross-validation over the full BBBP table. The validation split is standardized to `val`. ## Evaluation protocol All tasks use frozen embeddings. Classification tasks report ROC-AUC; Physical Chemistry reports RMSE; QM9 reports MAE. The notebook records the selected RMMol downstream-head settings without storing local output tables in the package. ## Hardware note The current inference and downstream-probe reproduction was run on the Spark workstation. RMMol pretraining used 8 x A800 40G GPUs. ## Weights The RMMol model weights are intended to be distributed from the Hugging Face model repository: `https://huggingface.co/Creeperbbs/RMMol`. ## Provenance The CSV files were standardized from the local GraphWorld/MoleculeNet data tree, the HIV split, and the split Tox21/SIDER archives. This bundle was generated on 2026-09-15.