The conventional wisdom is that biomedical knowledge graphs need full-precision embeddings to stay accurate. We tested it. The conventional wisdom is wrong at this scale.
The result, first
Tilelli Med embeds the entities of biomedicine — drugs, diseases, proteins, side effects, pathways — at
three values per weight. The packed model is 24 MB, roughly 5.3× smaller than the
standard-precision baseline. And it still posts 0.847 MRR on OGBL-biokg, ahead of the public baseline,
while also leading on PrimeKG. Smaller and better, not smaller and worse.
Why ternary is a good fit here
A knowledge graph is mostly structure: which things relate to which, and how. A lot of that structure survives aggressive quantization because the signal lives in the geometry of the embedding space, not in the last few bits of any single coordinate. That's the intuition. The exact recipe — how we train it so the geometry holds up at three values — is its own write-up, and we're not spilling it here.
What this is for
Link prediction. Given a disease, which proteins or pathways are most likely implicated? Given a drug, what else might it touch? A 24 MB model puts that kind of query on a laptop instead of a cluster, which matters for anyone doing biomedical research outside a well-funded lab.
What it is not
Tilelli Med is a research tool, not a clinical product. It ranks hypotheses for human experts to evaluate. It does not diagnose, prescribe, or make treatment decisions, and nothing it outputs should be read as medical advice. The honesty discipline we apply to the chat model applies here too: a confident-looking score is still a hypothesis until a human checks it.
The overview lives on the Tilelli Med page. The drug-ranking preview built on top of it is Petitri.