CatPred

CatPred webserver overview

CatPred predicts in vitro enzyme kinetic parameters from an enzyme sequence and the SMILES of its substrates or inhibitor. The webserver runs CatPred 1.0.1 (commit b314a28) with the published production checkpoints: one 10-model ensemble each for the turnover number kcat, the Michaelis constant Km and the inhibition constant Ki.

Each prediction is returned as a log10 value, its linear value in the native unit, and three standard deviations that separate total, aleatoric (data) and epistemic (model) uncertainty. The native CatPred CSV, including every individual ensemble model prediction, is available for download.

Pricing

A CatPred job starts at 12 credits. The quote is 11 credits plus 1 credit per 500 residues across the unique enzyme sequences in the job, rounded up. The exact quote is calculated before submission.

The number of SMILES per enzyme and the selected kinetic parameter do not change the price. Repeated sequences are counted once.

Unique enzyme length	1 enzyme	5 enzymes	10 enzymes
300 residues	12	14	17
500 residues	12	16	21
1,000 residues	13	21	31
2,500 residues	16	36	61

Inputs

CatPred has two input modes, selected with Input mode.

Mode	Input	Formats	Limits
Separate enzymes and molecules	Enzyme sequence(s)	.fasta, .fa, .txt	Up to 10 sequences, 2,500 residues each, 20 standard amino acids only
Separate enzymes and molecules	Substrate or inhibitor SMILES	.smi, .smiles, .txt	Up to 10 lines, one SMILES per line with an optional name after a space
CatPred CSV table	CatPred input table	.csv	Up to 100 rows

In separate mode, every enzyme is paired with every SMILES line, up to 100 predictions per job. Enzymes can also be fetched from UniProt and molecules from PubChem.

The CSV table uses CatPred's own format: case-sensitive SMILES, sequence and pdbpath columns, one prediction per row. pdbpath is a sequence identifier and must be unique for each unique sequence. Extra columns, such as a substrate name, are kept in the results.

Every job accepts at most 100,000 residues across its unique enzyme sequences. CatPred rejects the whole job if any SMILES cannot be parsed by RDKit, and reports the affected row.

What to enter as SMILES

Parameter	SMILES to provide
kcat	All substrates of the reaction joined with ., for example A.B
Km	The single substrate whose Michaelis constant is predicted
Ki	The inhibitor

CatPred canonicalizes each SMILES with RDKit. For kcat, it also sorts the dot-separated substrates, so their order does not matter.

Settings

Parameter	Type	Default	Description
Kinetic parameter	enum	kcat	kcat (s⁻¹), Km (mM) or Ki (mM). Selects the matching CatPred production ensemble.
Input mode	enum	Separate enzymes and molecules	Separate enzymes and molecules for every combination, or CatPred CSV table for specific rows.

Outputs

Output	Description
Results table	Every native CatPred column in its original order, one row per prediction
catpred_input_output.csv	Native CatPred predictions with uncertainty and individual ensemble model predictions
catpred_input.csv	The input table after CatPred's SMILES canonicalization
catpred.csv	The input table passed to CatPred, built from FASTA and SMILES in separate mode
provenance.json	Source commit, checkpoint and model weight digests, runtime versions and command
run.log	CatPred command output

In separate mode, the table also includes Enzyme (the FASTA header) and Molecule (the optional SMILES name). Identical sequences share one pdbpath such as seq_001.

Understanding results

Column	Meaning
Prediction_(s^(-1)) or Prediction_(mM)	Predicted value in linear units: s⁻¹ for kcat, mM for Km and Ki
Prediction_log10	Ensemble mean prediction on the log10 scale
SD_total	Total predicted standard deviation, in log10 units
SD_aleatoric	Uncertainty from noise in the training data
SD_epistemic	Uncertainty from disagreement between the 10 ensemble models; higher for unfamiliar inputs
log10kcat_max, log10km_mean, log10ki_mean	Native CatPred target column, identical to Prediction_log10
..._mve_uncal_var	Total predicted variance before the square root
..._model_0 to ..._model_9	Log10 prediction of each ensemble model, numbered in the order CatPred finds the checkpoints

provenance.json lists the checkpoint behind each model number. A range of one total standard deviation is 10^(Prediction_log10 ± SD_total). Enzyme sequences longer than 2,047 residues are truncated to CatPred's ESM-2 window when computing sequence features.

Citation

Boorla, V. S., & Maranas, C. D. (2025). CatPred: a comprehensive framework for deep learning in vitro enzyme kinetic parameters. Nature Communications, 16, 2072. https://doi.org/10.1038/s41467-025-57215-9