OCMR: A comprehensive framework for optical chemical molecular recognition

01 / Research question

How can a chemical drawing become a machine-readable molecular graph without losing local chemical details?

02 / Innovation

Recover molecular topology and integrate local information, including noncanonical drawings and abbreviated atom groups. The public pipeline outputs a structured representation such as SMILES.

03 / Mechanism

OCMR recovers molecular topology and resolves local atom-group details, including abbreviations, from structure drawings. The public Docker implementation exposes image-to-SMILES inference alongside MolVec, OSRA and Imago.

04 / Experimental setup

Public benchmarks are CLEF, JPO, UOB and USPTO. SMI evaluates equality of canonical isomeric SMILES after retaining the longest disconnected fragment. The evaluator compares all four recognition systems; training-set size and held-out counts are unavailable in the README.

Public repository README · benchmark table · Source ↗

05 / Results

Repository SMI scores are 0.650202, 0.604444, 0.859930 and 0.746284 respectively. These are implementation-reported values, separate from the final journal tables.

06 / Ablation & limitations

Abbreviations, unusual drawing styles and stereochemistry remain important error modes. Fragment normalization can discard disconnected components. The README calls its second metric Tanimoto, but the evaluator computes Dice similarity on radius-3 Morgan fingerprints; that column is therefore not plotted here. A journal ablation table remains unavailable; the original Figure 1 is reproduced under the coauthor’s retained reuse rights.

High fingerprint similarity does not guarantee an identical decoded molecule. Chemical validity and structure identity need separate checks.

07 / My contribution

Second listed author on the published article.

08 / Paper & code