Datasets:

Modalities:
Image
Text
Formats:
parquet
ArXiv:

Question: duplicate OpenPhenom embedding vector across distinct RxRx3-core metadata identities

#4
by thehenryco - opened

Hi Recursion/OpenPhenom team,

I opened a reproducible note here:
https://github.com/thehenryco/themis-no-narrow-runtime/issues/1

Using the public RxRx3-core metadata and OpenPhenom embedding parquet, I found one duplicate 384-feature embedding digest across two distinct metadata identities:

  • PLK1 / PLK1_guide_1, CRISPR Gene Positive Controls
  • AKR1C3 / AKR1C3_guide_5, Query guides

This is not a claim of dataset error. I’m asking whether duplicate embedding vectors are expected behavior, and whether downstream consumers should treat embedding equality as a representation collision rather than a biological identity key.

The repo includes an independent verifier and a passing public harness.

Question: should consumers preserve well_id and metadata identity separately from embedding-vector equality when using RxRx3-core/OpenPhenom embeddings?

Sign up or log in to comment