Generate novel, chemically valid analogs of a seed molecule — or sample the molecular prior directly — with zjunlp/MolGen-large, a molecular language model pre-trained on 100M+ SELFIES strings. Because generation happens in SELFIES space, every output decodes to a valid 2D molecule; results are canonicalized, deduplicated, and audited with SELFIES, canonical SMILES, Bemis–Murcko scaffold, and seed linkage so downstream ADMET, docking, filtering, and clustering can triage the library.
0 seeds · 10 sequences/seed
Estimated minimum cost: $—
Examples
Seed molecules
Paste SMILES or InChI one per line, or mol blocks / SDF records. At most 1,000 seeds — leave this empty to sample the model prior unconditionally. Seeds are converted to SELFIES internally, the representation the generator was trained on; standardize salts, charges, and tautomers first for the cleanest analogs.
or drop a file here
Decoding
Beam search decodes deterministic, close analogs of each seed; sampling with temperature and nucleus/top-k filtering explores chemical diversity. Higher beam width finds higher-likelihood analogs; the model never proposes chemically invalid output because generation happens in SELFIES space.