[image: :seedling:] *What if the answer isn’t in a database—but in the connections between databases?*
Join us on *Wednesday, October 7* for a webinar with Ehsan Estaji of the Umeå Plant Science Centre, Sweden:
[image: :link:] Trapped knowledge in the AI-agent era: how PlantGraph turns multi-database questions into one traversal you can check
PlantGraph connects plant data across the transcriptome, epigenome, regulome, and proteome, helping researchers turn complex, multi-database questions into one connected, traceable query.
As part of the monthly AgBioData Webinar Series https://www.agbiodata.org/webinars, Ehsan will show how this approach can help AI agents retrieve evidence more reliably—and how we can check whether AI-generated answers actually stay grounded in that evidence.
Wednesday, October 7 | 1P ET | 12P CT | 11A MT | 10A PT | Find your local time here https://www.timeanddate.com/worldclock/fixedtime.html?msg=AgBioData+Oct+2026+Webinar+-+Ehsan+Estaji+%28PlantGraph%29+&iso=20261007T13&p1=3918&ah=1 .
Join Zoom Meeting https://us06web.zoom.us/j/82038356125?pwd=YVFMRElMdEpHZmtObXFvZlA4QVFXQT09 Meeting ID: 820 3835 6125 Passcode: 160683
Title: Trapped knowledge in the AI-agent era: how PlantGraph turns multi-database questions into one traversal you can check
Abstract: A real biological question is rarely a record — it is a path. “Which salt-responsive genes also carry a promoter methylation change, a transcription-factor binding site, and a known phosphosite?” has no home database: the answer exists only as a traversal across four of them. This is why so much of what we collectively know sits trapped. Not because the facts are missing — they are curated, somewhere — but because the connections between them were never stored anywhere. Researchers rebuild those connections by hand, one question at a time, and lose them the moment they close the tabs. Capable AI agents raise the cost of that fragmentation rather than resolving it. An agent handed a set of endpoints has to plan the join itself, guessing identifiers and directions at every hop, and every hop is an opportunity to improvise. In our own benchmarking, giving a strong model tools and API access over fragmented sources did not reliably fix retrieval; connecting the data first, and retrieving over it deterministically, did. PlantGraph (plantgraph.se) integrates community plant resources into one connected knowledge graph and separates two jobs that agents usually conflate: the graph retrieves the evidence, the language model only writes over it. A multi-database question becomes a single traversal — and because that traversal actually ran, every claim can be followed back to the query that produced it and the source record that curated it. I will walk through questions spanning transcriptome, epigenome, regulome and proteome resolving in one step, how we measure when the generated narration drifts from the retrieved evidence, and where the approach still fails. PlantGraph https://plantgraph.se/ is not a replacement for the databases it draws on — it stands on them, and every resource that connects makes the rest worth more.
See also this news blog https://www.agbiodata.org/node/652.
Marcela