Recently, we worked on a patent invalidity analysis involving a chemical patent. We engaged an outside search firm to conduct a prior art search. At the same time, we tested a couple AI-based patent-search tools to see how they would conduct a prior art search. The tools did reasonably well identifying prior art for limitations expressed in text. But one of the claims included a chemical Markush structure, and that is where their performance noticeably declined. The human search identified relevant chemistry that the AI tools had not surfaced.
That experience highlights a current problem for AI tools for chemical patent searching : chemical similarity and textual similarity are not the same thing.
Consider acetaminophen. One reference may call it "acetaminophen," another "paracetamol," another may use its systematic chemical name, and another may show only its structure. A fifth reference may disclose a broader Markush genus encompassing the compound without naming it at all. To a chemist, these disclosures can occupy the same chemical space. To an AI tool searching primarily for similar language, however, they look very different.
A Markush claim makes the problem even harder because it does not identify one molecule. It uses a generic structure, together with definitions such as "R¹ is alkyl," to claim a large family of compounds. Chemical Abstract Services (CAS) describes a Markush structure as one structure defining a set of implied structures.
It is helpful to consider an example. The Markush structure above is reproduced from claim 1 of U.S. Patent No. 9,212,545 for illustrative purposes and is not a claim we have worked with. The mechanical drawing is Figure 1 of U.S. Patent No. 9,347,549, which is directed to a shifting assembly for a vehicular transmission. In both images, significant information is conveyed visually rather than through text.
A patent practitioner can look at a Markush structure, read the accompanying R-group definitions, and identify the compounds and subgenera that fall within its scope. A general-purpose AI workflow must first recognize the drawing, connect each variable to definitions elsewhere in the patent, translate that information into a machine-searchable representation, and then search for overlapping chemistry.
Although large language models may be trained through next-token prediction, the difficulty with Markush structure representations is broader than that training method alone. It is also an image-recognition, chemical-representation, and information-retrieval problem.
Chemistry does have text-based representations for molecules, which one might think would mitigate the problem AI tools have with Markush structures. SMILES converts a molecular graph into a line of characters; for example, ethylene is written as C=C. InChI likewise provides a standardized textual identifier for a chemical substance; ethylene's standard InChI is InChI=1S/C2H4/c1-2/h1-2H2.
But these text-based representations solve the problem of representing a defined molecule in a way that software can understand it. A Markush structure defines a family through variable atoms, substituents, attachment points, optional groups, and combinations of alternatives. Converting a specific molecule into text therefore does not answer the original Markush structure problem: Which individual compounds are encompassed by the claimed Markush structure?
Success for the next generation of patent-search tools will depend on whether they can understand chemistry the way patent practitioners do. That means recognizing that a Markush structure is more than an image on a page, interpreting the accompanying R-group definitions to understand the scope of the claimed genus, searching against chemical information, and understanding the different ways compounds can be named and represented.
General-purpose language models are becoming remarkably capable, but chemical patent searching requires more than language understanding. It requires understanding of chemistry.
There is also a practical implication for patent practitioners. USPTO examiners use AI search tools. If a company wants its patent portfolio to function as strong blocking art, it may want to represent important molecular concepts in ways other than Markush structures. Doing so may improve the likelihood that AI search tools will find their art.
By: Brad Hough