MarkushGlyph

A Markush structure is a chemical drawing that stands for a whole family of compounds rather than a single molecule: a scaffold decorated with variable R-groups, positional-variation attachments, and repeat units, accompanied by text that enumerates what each placeholder may be. Named after Eugene Markush, who first proposed the device and had it upheld in court, these drawings exist essentially only in patents — and they are how chemical intellectual property is actually claimed. Reading one correctly is therefore a prerequisite for knowing whether a molecule is already covered by someone else's patent.

MarkushGlyph, released by Edison Scientific in July 2026, is a vision-language model that reads an entire Markush structure as a single image and writes it out as a machine-readable line notation. That single-stage framing is the model's central architectural claim: prior systems typically decompose the problem into multiple stages that process the drawing and its accompanying text separately, then recombine the pieces. MarkushGlyph treats the whole thing as one image-to-text translation. It was introduced alongside OCSRGlyph, a sibling model for single-molecule optical chemical structure recognition (OCSR), in a preprint that also proposes a stricter way of scoring Markush translations.

The two tasks are at very different stages of maturity. Single-molecule OCSR — the problem addressed by tools such as MolSight — is, in the authors' assessment, quite mature, while Markush parsing remains substantially unsolved. The practical stakes are high: patent-landscape checks for a lead molecule are currently performed by specially trained experts, typically PhD chemists with additional legal training, who hand-review candidate compounds against the Markush structures and textual definitions in the patent literature.

#Key Features

• Single-stage image-to-text translation: The full Markush drawing, including its R-group labels and annotations, is consumed as one image, avoiding the error propagation of pipelines that segment visual and textual content into separate processing stages.
• CXSMILES-opt output: The model emits a tokenization-friendly variant of CXSMILES in which each R-group label sits inline beside the atom it modifies instead of in a trailing table, so labels need not be re-indexed when atom ordering changes.
• Strict parsed-graph equality metric: The paper introduces a stricter scoring rule that requires the entire prediction to be correct, penalizing transposed labels, spurious extra annotations, and misplaced positional variations that earlier metrics let pass.
• Majority-vote decoding: Beyond greedy decoding, sampling eight candidates (temperature 0.7, nucleus sampling at 0.95), canonicalizing each, and taking the most frequent result adds several points of accuracy.
• Open artifacts with a caveat: Code and benchmark datasets are released under Apache-2.0, but the weights are distributed as a LoRA adapter that carries no declared license and requires the separate Qwen/Qwen3.5-2B-Base checkpoint to run.

#Technical Details

MarkushGlyph is a LoRA fine-tune (rank 128, scale 128) of Qwen3.5-2B-Base, a vision-language model that pairs a ViT vision encoder with a transformer decoder. Training used 263,158 records: 152,620 synthetic Markush structures, 52,524 structure-only records derived from real patents, 44,856 molecule-to-Markush conversions, and 13,158 ordinary single molecules. Scored with the official MarkushGrapher-2 scorer, MarkushGlyph reaches 58.2% (greedy) and 60.6% (majority vote) on IP5-M (878 test rows) against 53.7% for MarkushGrapher-2; 61.2% and 62.1% on M2S (103 rows) against 56.0%; and 59.5% and 63.5% on USPTO-Markush (74 rows) against 55.0%. The margins widen under the authors' strict parsed-graph equality metric, where the model scores 61.2%/64.1% on M2S versus 51.5%, and 54.1%/56.8% on USPTO-M versus 41.9%, with a narrower 52.1%/54.7% versus 51.1% on IP5-M. The authors note that the 41.9% USPTO-M baseline is a conservative estimate, reflecting discrepancies they encountered reproducing the comparison system.

#Applications

The target workflow is patent analysis in drug discovery: checking whether a lead molecule falls inside the claims of existing patents, a task presently gated on scarce, expensively trained specialists. Automated Markush parsing also feeds chemical database construction and the assembly of machine-learning training sets from patent corpora, where structures locked in images are otherwise invisible to search. Medicinal chemists, IP and freedom-to-operate teams, and cheminformatics groups digitizing patent literature are the primary beneficiaries.

#Impact

MarkushGlyph is a July 2026 preprint and has not been peer reviewed. Its most useful contribution may be diagnostic rather than absolute: by introducing a metric that refuses partial credit, it shows that headline Markush scores under prior metrics overstated real performance, and the authors state plainly that Markush recognition still lags substantially behind single-molecule OCSR. At roughly 55–64% exact-structure accuracy, the model is a strong triage aid rather than a replacement for expert review. The work is also a case study in Edison Scientific's "autoresearch" program — automating the creation and training of predictive models, with its Kosmos AI scientist contributing to problem framing and proposing the stricter equivalence metric — which makes the release notable as much for how it was produced as for what it scores.