OCSRGlyph

Chemistry is published as pictures. Patents and journal figures encode molecules as line drawings, and any programmatic use of that chemistry — indexing a patent corpus, assembling a training set for a property model — first requires converting the drawing back into a machine-readable string. Optical chemical structure recognition (OCSR) is the single-molecule version of that task: image in, SMILES out. It is a mature area with a decade of published systems, and accuracies on the standard USPTO benchmark had converged in the low 90s.

OCSRGlyph, released by Edison Scientific in July 2026, is built around a diagnosis of what the residual errors actually were. Analysis of early training runs traced them not to connectivity but to stereochemistry: the model recovered the correct molecular graph and then mis-assigned the configuration at a stereocenter. Since stereoisomers sharing a connectivity graph can be distinct compounds with distinct pharmacology, those errors matter disproportionately downstream.

The fix was not a new architecture. OCSRGlyph is a conventional image-to-sequence transformer; what changed was the training curriculum, which deliberately oversamples chiral molecules and other stereochemically difficult motifs. Edison credits Kosmos, its AI-scientist system, with proposing that oversampling strategy, and frames the work as an instance of what the company calls autoresearch — automating the creation and training of predictive models. OCSRGlyph was released alongside MarkushGlyph, a companion vision-language model for the harder Markush case, in which a single drawing stands for a whole family of molecules.

#Key Features

• Stereochemistry-enriched training curriculum: Two stereochemistry-rich sources are oversampled on top of the standard OCSR pools, raising full-stereochemistry accuracy without degrading accuracy on non-chiral molecules.
• Compact encoder-decoder: At roughly 94 million parameters it is about 20 times smaller than its 2B-parameter sibling, and runs on a single GPU or on CPU.
• Reported under three stereochemistry conventions: Accuracy is given for canonical (all stereochemistry correct), chirality-kept (tetrahedral chirality only), and graph (connectivity only) exact match, so results are comparable to prior papers that each adopted a different convention.
• Deterministic postprocessing: Isolated explicit-hydrogen fragments that the decoder occasionally hallucinates are stripped and the prediction is re-canonicalized with RDKit, yielding 99.6% molecule validity.
• Released artifacts with mixed licensing: Training code, configs, and frozen self-contained benchmark files are Apache-2.0, while the published checkpoint carries no declared license — its model card lists one as pending.

#Technical Details

OCSRGlyph pairs a Swin-B Transformer encoder at 384-pixel resolution, initialized from ImageNet-pretrained weights, with a six-layer Transformer decoder trained from scratch that emits a bare SMILES string one character at a time over a character-level vocabulary. Decoding is greedy. Training runs 100,000 steps at an effective batch size of 256 with a cosine learning-rate schedule peaking at 4e-4 and label smoothing of 0.1. The base data pool is the PubChem-1M and USPTO-680K OCSR sets distributed together by MolNexTR, with PubChem images rendered on the fly by EPAM Indigo and RDKit canonical SMILES as targets. The enrichment consists of Stereo-200K, preprocessed to a white background and a tight crop, plus a 64,752-molecule subset of PubChem-1M mined for adjacent-ring stereocenters — a motif the authors identified as especially difficult — with each enrichment source sampled twice as often as the base pool.

On the 5,719-image USPTO OCSR benchmark, OCSRGlyph reaches 93.8% exact match under the strictest canonical convention, 93.9% chirality-kept, and 96.2% graph, with 99.6% molecule validity. An ablation varying only the data mixture, with the architecture, objective, optimizer, and schedule held fixed, isolates the source of those gains: the base pools alone plateau near 90% canonical exact match, oversampling a 50,000-molecule stereo-rich subset alone reaches 93.5%, and the full mixture reaches 93.8%.

#Applications

The immediate use is bulk digitization of chemical drawings: converting structure images in patents, journal articles, and internal documents into SMILES for database indexing, prior-art search, and the construction of training sets for downstream property and reaction models. Because stereochemistry is where the model's gains sit, it is particularly relevant to medicinal-chemistry and intellectual-property workflows where a single stereocenter distinguishes a claimed compound from an unclaimed one. The small footprint lets large corpora be processed on modest hardware, and the released training code and configs let groups retrain on their own drawing conventions.

#Impact

OCSRGlyph advances a benchmark that had largely stopped moving, showing that targeted data curation, not a larger model or a new architecture, was what the remaining accuracy gap required. It is also a counterpoint to the scale-up trend in molecular image understanding, where systems such as MolSight attach graph-topology modules to 8B-parameter vision-language backbones. The authors are candid about what remains: with stereochemistry errors largely addressed, wrong connectivity is now the largest residual error class, and stereochemistry-focused curation does not touch it. The work is a July 2026 arXiv preprint awaiting peer review, and while code, datasets, and weights are all public, the checkpoint's licensing status is unresolved.