YinkaiW/pubchem-faiss-library

PubChem-Scale Molecule Retrieval Library (FAISS)

This dataset provides a prebuilt, global FAISS retrieval library for molecule search at PubChem scale. It includes molecular metadata (SMILES, formula, exact mass), precomputed embeddings, and FAISS indices.

Contents

• index_smi.faiss: FAISS HNSW index built on SMI-TED embeddings (vectors_smi.npy).
• index_chem.faiss: FAISS HNSW index built on ChemBERTa embeddings (vectors_chem.npy).
• vectors_smi.npy: SMI-TED embeddings for each molecule.
• vectors_chem.npy: ChemBERTa embeddings for each molecule.
• meta.parquet: metadata table with columns id, smiles, formula, mass.

Data source

The SMILES list is derived from a PubChem SMILES file used in our local pipeline (De-SpecBridge/data/pubchem.smi). We compute formula and exact mass using RDKit.

Build pipeline

The library and FAISS indices were produced by the Spec-RAG scripts:

• Spec-RAG/scripts/build_library.py (build metadata and embeddings)
• Spec-RAG/scripts/build_faiss.py (build FAISS indices)

FAISS indices use HNSW with inner product on unit-normalized vectors to approximate cosine similarity.

Intended use

This library is designed for global molecule retrieval, e.g. given a query embedding, retrieve nearest neighbors from a large candidate pool. For MS applications, it can be combined with mass or formula filtering using meta.parquet.

Notes

• This is a large dataset (hundreds of GB).
• If you need mass-bucketed indices (one index per mass or ppm window), you can build them by filtering meta.parquet then indexing the corresponding embedding subset.

Contact

If you have questions, please reach out to the dataset owner on Hugging Face.