Edison Scientific launches research hub with open models for reading chemical patents

Why it matters

Edison Scientific is using open weights and benchmarks to build technical credibility for Kosmos as AI-scientist vendors move from research demos into enterprise biopharma workflows.

On July 23rd, Edison Scientific, led by co-founder and CEO Sam Rodriques (@SGRodriques), launched Edison Advances, a site for research findings, engineering posts, open model weights and benchmarks. Rodriques promoted the site in a thread on X on July 31st, eight days after the dated launch post. The first release targets a narrow bottleneck in drug development: turning chemical structures embedded in patent images into machine-readable data. (advances.edisonscientific.com)

Rodriques is a physicist and bioengineer who earned his doctorate at MIT under Ed Boyden after studying physics at Haverford College and engineering at Cambridge. Before starting FutureHouse, he worked as an entrepreneur-in-residence at Petri Bio and led an applied biotechnology laboratory at the Francis Crick Institute. Edison Scientific lists Rodriques as co-founder of both Edison Scientific and FutureHouse, the nonprofit AI research organization that began pursuing an AI scientist in 2023. (edisonscientific.com)

The first models

The initial Edison Advances release includes MarkushGlyph and OCSRGlyph. MarkushGlyph reads Markush structures, patent diagrams used to describe families of related chemical compounds. OCSRGlyph handles standard optical chemical structure recognition, converting a molecule image into a SMILES line representation. Edison Scientific released the paper, training and inference code, datasets and model weights for both systems. The arXiv paper was submitted on July 30th by Alex Andonian, Rodriques, Edison co-founder and CTO Andrew White, and Siddharth Narayanan. (advances.edisonscientific.com)

MarkushGlyph is based on Alibaba's Qwen3.5-2B-Base vision-language model and accepts an image containing a chemical diagram and accompanying text. Edison Scientific says the model outperformed earlier systems across its evaluated Markush benchmarks, including under a stricter metric introduced by the researchers. OCSRGlyph, a transformer encoder-decoder model, reached 93.8% accuracy when stereochemistry had to match and 96.2% without that requirement on a benchmark of molecules extracted from U.S. patent data, according to Edison Scientific. Those are research-team results rather than independent evaluations. (advances.edisonscientific.com)

The release also provides a test of Edison's central product claim. Edison Scientific says its Kosmos AI scientist helped researchers frame the chemical-recognition problem, propose hypotheses and run experiments. Kosmos independently proposed oversampling chiral molecules during OCSRGlyph training and helped devise the stricter Markush evaluation metric, according to the research post. The human researchers remained responsible for directing the project and reporting the results. (advances.edisonscientific.com)

A public proof layer for Kosmos

Edison Advances gives Edison Scientific a venue to expose the narrower models, datasets and benchmarks produced around Kosmos while selling the broader agent to research organizations. Edison Scientific describes Kosmos as a persistent code-writing agent that can review scientific papers and patents, generate hypotheses, provision computing resources and work with researchers through interfaces including Slack, Teams and email. Edison Scientific says Kosmos can also be deployed inside regulated organizations with limited external data access. (advances.edisonscientific.com)

Edison Scientific is the commercial spinout of FutureHouse, which Rodriques founded in 2023 to build semi-autonomous systems for scientific research. Edison Scientific incorporated in Delaware in 2025. In December 2025, Edison Scientific announced a $70 million seed round led by Triatomic Capital, Spark Capital and an unnamed U.S. institutional biotech investor, with Pillar VC, Susa Ventures, Striker Venture Partners, Hawktail VC and Olive VC participating. A December 9th Form D listed a $68.89 million offering, with $65.89 million sold to 26 investors at the time of filing. (edisonscientific.com)

Edison Scientific has since pushed Kosmos into commercial drug-development work. A May partnership with Incyte covers target discovery, target validation and translational biology. A June agreement with Population Health Partners applies Kosmos to creating biotechnology ventures focused on cardiovascular, inflammatory lung and gastrointestinal diseases. The companies have not published financial terms for those agreements. (edisonscientific.com)

AI-for-science groups are competing over how much of the research process agents can perform and how results should be measured. Google has positioned its Gemini-based AI co-scientist as a collaborative system for generating hypotheses and research proposals. Edison Scientific is pursuing a wider enterprise role for Kosmos across proprietary data, computing and long-running R&D workflows. Edison Advances gives researchers a way to inspect concrete outputs from that strategy, starting with models built for a specialized patent-analysis task where errors can affect which chemical compounds a drug developer chooses to pursue. (research.google)