Lots of SMILES: a large deduplicated collection of real-world SMILES

Lots of SMILES (LOS) is a single corpus of 15,260,616,134 unique, canonical SMILES strings, assembled by merging and deduplicating three large public molecular collections: PubChem, ZINC20, and the Enamine REAL™ Database. Every record is a single SMILES string, with no catalogue identifiers, prices, computed properties, building blocks, synthons, or source labels. Format. The corpus is a single Zstandard file holding one canonical SMILES per line in plain ASCII, byte-sorted and deduplicated, split into 5 GB parts named corpus.canonical.smi.zst.part-* for upload. Reassemble and decompress with cat corpus.canonical.smi.zst.part-* > corpus.canonical.smi.zst then zstd -d --long=31 corpus.canonical.smi.zst (the 2 GB long-distance window requires the matching flag). Stereochemistry uses the OpenSMILES extended tetrahedral notation and aromatic rings are written in Kekule form. The sha256 of the reassembled compressed file is 8924d2a171562d8ef23649ef49c5b5790c795fe4c952071761d6209380d091e2. Construction. Each source was read in full: PubChem (about 120 million), the ZINC20 two-dimensional catalogue (about 1.93 billion), and the entire Enamine REAL™ Database across all heavy-atom-count partitions (about 13.6 billion). The only structural filter was a single-connected-component requirement. Records were deduplicated first by the InChIKey shipped with each source, then re-canonicalized with a pure-Rust canonicalizer and deduplicated again on the canonical string. The funnel was 15,685,365,054 ingested, 15,399,489,475 unique by shipped InChIKey, and 15,260,616,134 unique canonical SMILES. Reproducibility. Produced by the open-source lots-of-smiles pipeline (https://github.com/earth-metabolome-initiative/lots-of-smiles), using the smiles-parser crate (https://github.com/earth-metabolome-initiative/smiles-parser). License and attribution. Released under CC BY 4.0. Reuse, including commercial use, is permitted as long as the original sources (PubChem, ZINC20, and the Enamine REAL™ Database) are attributed. The REAL-derived structures are published with the express written permission of Enamine: the Enamine REAL™ Database must be credited as their source, and the notice that REAL™ is a registered trademark of Enamine must be preserved.