Property prediction · Cheminformatics

CatPred

Predict kcat, Km or Ki for an enzyme and its substrate, with an uncertainty estimate.

CatPred combines an ESM-2 sequence embedding of the enzyme with a message-passing representation of the substrate, and predicts a distribution rather than a point value: each prediction carries total, aleatoric and epistemic standard deviations, and the epistemic part grows as the query moves away from the training data. One run scores one CSV of reactions -- enzyme sequence, substrate or inhibitor SMILES, and a sequence ID -- against one of the three parameters, using the ten-model production ensemble the authors predict with themselves.

Documentation ↗ Input parameters ↗ Examples ↗ Home page ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Input mode

Input

Job name Names the output directory and the predictions CSV inside it.

Kinetic parameter One ensemble per parameter, each a separate model: a run predicts one of them. kcat is predicted for the reaction's whole substrate set, Km for a single substrate, and Ki for a single inhibitor.

Enzyme and substrates

One enzyme box and one box per substrate, which is one reaction. The enzyme's ID names its sequence in the results and is what CatPred caches its embedding under; a substrate's ID is an optional label. For kcat the substrates are joined with ".", as CatPred's own CSV writes a reaction; Km and Ki take a single ligand. Enter a CSV to score several reactions in one run.

Reactions (CatPred CSV) CatPred's own input format: a header with SMILES, sequence and pdbpath, then one row per reaction (up to 200). pdbpath is an identifier, not a file, and must name exactly one sequence across the file; left blank, rows are numbered seq_001 onward. Any extra column, such as the demo files' Substrate, is carried through to the predictions.

Reactions file Upload a CatPred CSV with SMILES, sequence and pdbpath columns. The file contents are sent to the compute node.

Runtime

Device Auto uses CUDA when available, otherwise CPU. An explicitly requested unavailable CUDA device fails clearly. Each row costs one ESM-2 embedding and ten model evaluations; CPU runs are slow but supported.

Ready