SMILES Mutation Search

The molecule-domain counterpart to Single Mutant Search: a search protocol that proposes novel SMILES each round by mutating the top-K best-labelled training molecules, filtered for RDKit validity. Unlike protein sequences, most single-character SMILES edits are structurally invalid, so this filtering step is required. RDKit’s error logger is silenced, since invalid mutations are expected and handled internally.

Mutating only the single current best (top_k=1) makes the search a pure hill-climb: once no neighbour of the incumbent beats it, the same neighbourhood is regenerated every round and the loop stalls in that local optimum. Setting top_k higher keeps several regions of the search space open at once, so a stall in one neighbourhood doesn’t stall the whole search.