BasicAutoencoder encodes molecular SMILES into a latent space and decodes them back into molecular strings. Its usefulness depends on both reconstruction fidelity and chemical validity. section 10.2.4
Evaluate reconstruction quality at several levels:
- Token accuracy reaches 0.994, indicating strong learning of SMILES syntax. section 10.2.4
- Valid reconstruction rate reaches 88.2% when RDKit successfully parses the reconstructed SMILES, leaving over 10% chemically invalid. section 10.2.4
- Exact match rate reaches 0.847 on canonical MOSES SMILES; canonicalization matters because equivalent molecules can have different valid strings. section 10.2.4
- Tanimoto similarity compares non-exact reconstructions using Morgan fingerprints with radius 2 and 2048 bits. section 10.2.4
