CatalystCompass

From a reaction to an EC number.

A CPU ensemble that combines reaction fingerprints, reaction BERT encoders and chemical-neighbour retrieval. CatalystCompass adds a conservative consensus decision layer and a faster top-one decoder.

Part Purpose
Structural views Recognize substrates and transformation context
Reaction encoders Add learned reaction representations
Consensus lookup Recognize alternate encodings of the same reaction
Output checks Preserve input IDs and return one valid EC

Start here

Download catalyst_compass_setup.zip and run:

mkdir -p ~/Downloads/DL_project
unzip -o ~/Downloads/catalyst_compass_setup.zip -d ~/Downloads/DL_project
bash ~/Downloads/DL_project/catalyst_compass/run_mac.sh

Your input is read from DL_project/data/val_inputs.parquet. To use another file, pass its path after run_mac.sh.

The first run downloads approximately 727 MB of model assets from this repository at a pinned revision. Existing verified files are reused. Setup and image building require internet access; prediction runs with network access disabled.

Item Location
Model folder catalyst_compass
Docker image catalyst-compass
Predictions output/predictions.jsonl
Run details output/run.log
Technical notes docs/MODEL_CARD.md
Measured checks docs/CHECKS.json

What changed

  • Strict canonical reaction matching preserves stereochemistry, isotope labels, reaction direction, reagent roles and molecule counts. A lookup is accepted only when the recorded references agree on a single full EC.
  • Top-one decoding avoids sorting all classes and preserves the base decoder's ordering when scores tie.
  • Fingerprint compatibility failures stop inference instead of allowing a mismatched environment to continue.

These are inference changes around existing neural checkpoints. The public validation inputs have no eligible canonical training matches, so the lookup does not provide a measured score gain on them. Private accuracy has not been measured. See THIRD_PARTY_NOTICES.md for source attribution and reuse information.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support