Mol-JEPA Dataset

Multimodal molecular dataset used to train Mol-JEPA (a multimodal Joint Embedding Predictive Architecture for molecules). Each row of metadata.csv describes one molecule (SMILES + InChIKey + source dataset + labels) and points to precomputed per-modality embedding/target files stored as NumPy arrays.

Modalities

These are the modalities included (note that not every modality is available for every row - there is quite some sparsity). For detailed preprocessing steps please refer to the paper.

• ChEMBL: Bioactivity measurements for various endpoints
• MOE: Molecular descriptors of size 227
• Boltz: Embedding vectors from the binding affinity module of the Boltz-2 model against off-targets
• TDC: ADMET measurements for various endpoints
• PCBA: Bioactvity measurements
• NABLA: DFT properties
• CLOOME: Cell painting model embeddings
• ChemGPT: SMILES Transformer embeddings
• BioXMol: Multi-assay phenotypic model
• UMA: Atomistic foundation model embeddings
• Boltz preds: Predictions and uncertainties from the binding affinity module of Boltz-2 against off-targets

Layout

• metadata.csv - one row per molecule. The <modality>_path columns hold the relative path of that molecule's array, e.g. boltz/boltz_0.npy.
• <modality>/<modality>-XXXXX.tar - sharded tar archives containing the .npy files (unpacked they restore the <modality>/<modality>_<idx>.npy layout).

Usage

from datasets import load_dataset

ds = load_dataset("Flogrammer/Mol-JEPA-dataset")
print(ds)

To materialise the .npy arrays for a row, download/extract the corresponding shard and read the path given in the <modality>_path column.

Citation

If you use this dataset, please cite the Mol-JEPA paper.