RamonY
Building a mundaneum of the Neuroscience field
We wrote this document to show our progress, open for feedback, share with you the potential we discovered while doing our project, and ask for further resources to complete the first mundaneum of Mammalian neuroscience.
We started building the knowledge graph of neuroscience we planned for.
“a production-grade extraction system … an agentic interface where Claude navigates and reasons over the graph … and rigorous benchmarks demonstrating that structured graph context measurably improves Claude's scientific reasoning”Proposal for Claude for Life Sciences Collaboration, April 2026
When we wrote the proposal, our prototype held eleven thousand nodes and fifty thousand edges, from ten papers and one atlas.
We first strived to cut extraction costs.
We optimized to shrink per-paper cost to a minimum.
We made it down to just $0.16 a paper.
Therefore we were able to start tackling the entire neuroscience.
As we scaled the extraction pipeline up, we realised we could do more than a knowledge graph.
What we do differently is qualify our edges at scale, with an expressive but structured vocabulary. A knowledge graph stores a subject, a relation and an object: it can tell you that GABA regulates a neuron. Ours carries, on every edge, the conditions under which that was observed — species, cell type, developmental stage, brain region, direction of effect, assay — drawn from 1,121 condition keys.
This led to the idea of creating a mundaneum of neuroscience. A mundaneum is a universal archive: every fact in a field written down once, in one shared classification, so that any of it can be found and joined to any other.
“Mankind is at a turning point in its history. The mass of data acquired is astounding. We need new instruments to simplify it, to condense it, or intelligence will never be able to overcome the difficulties imposed upon it or achieve the progress that it foresees and to which it aspires.”Paul Otlet, Traité de documentation, 1934
The word is not ours. It belongs to Paul Otlet and Henri La Fontaine, who began theirs in Brussels in 1895 and had filled more than twelve million index cards before the war scattered it — and our extractor is named after Otlet.
The scale of the mundaneum is well beyond the largest knowledge graph, OptimusKG.
We took the field and found it is 2.7 million papers.
Including the fundamental biochemistry, cytology, physiology and neurology needed to understand neuroscience.
We go back to the origins of the field.
We sorted it, to get a good representation with the budget we had.
We started formalising legal agreements to be able to do text and data mining on it.
What leaves this project is triplets — a subject, a relation, an object, and the conditions the fact holds under — never the text itself.
Then we needed to transform this pile of papers into a set of facts observed in our field.
We designed a canonical schema that could cover the vast majority of a paper's content: which kinds of thing exist, which relations can hold between them, and which conditions each relation may carry. Everything downstream, the databases included, is written into that same vocabulary.
We defined our mundaneum's facts and tuned Otlet, our Claude-based named-entity-recognition and relationship-extraction pipeline, to harvest them comprehensively across subfields.
So we hired biologists to annotate the pipeline results with a dedicated platform, and incorporated their feedback on the go.
Thanks to them, we realised the unsolved challenge of having one biological object represented at once, so that different papers can talk to one another. This is called entity linking.
So we invested much time in creating universal linkers throughout our canonical schema — and then in the same problem between databases, at a scale no one can handle by hand.
This enabled us to make all papers talk to one another, and we soon realised that the value of a paper was, counterintuitively, increasing as more papers were added.
Because it created paths that would not otherwise exist.
This made it clear that we cannot stop at our planned million papers, or several questions would be unanswerable simply for the lack of connecting papers. But that comes with a cost.
To increase coverage within our budget, we expanded beyond papers.
We prepared to parse textbooks, which we collected through the EPFL library.
And mostly through databases, which are free and contain highly detailed, curated information about subfields. A paper reports that something was observed; a database row reports how much, in how many samples, at what age, and in which region.
We realised we could use Claude subagents to bring heterogeneous databases into our canonical world, and went big, including more than a thousand databases. The list itself was manually curated by biologists with the help of Claude, and every adapter into our taxonomy was written by a Claude subagent and manually curated by a biologist.
Among those we also included quantitative datasets, such as scRNA-seq average gene expression, with rich qualifiers as in the literature, to achieve a field-scale view.
We therefore created the first field map of Neuroscience.
One point per fact. Each fact is verbalised into a sentence and embedded, so that a point sits where it does because of everything in it, conditions included.
This field-scale mundaneum can be asked questions no single curated set of papers or databases could tackle.
But you need a Cypher expert to ask questions of the mundaneum, so we started building the RamonY Claude agent.
As promised, we also started curating a benchmark.
It compares RamonY against the state of the art on a sensible set of neuroscience questions.
We then built a portal to use RamonY.
And we made sure Claude subscribers will be able to talk to RamonY, by shipping an MCP server they can call from the client they already use.
We therefore felt ready to ask the field for feedback, and the response at FENS in Barcelona was amazing.
But to let the field ask anything, RamonY has to stay up to date. We made sure it will read all new papers every week using our newsletter Paperboat, so it grows as the field grows.