DLKcat

What is DLKcat?

DLKcat predicts enzyme turnover numbers (kcat values) from protein sequences and substrate structures. The turnover number represents how many substrate molecules an enzyme can convert to product per second under saturating conditions—a fundamental kinetic parameter for understanding enzyme efficiency.

Developed at Chalmers University of Technology and published in Nature Catalysis (2022), DLKcat combines a convolutional neural network (CNN) for processing protein sequences with a graph neural network (GNN) for analyzing substrate molecular structures. This dual-network architecture allows the model to learn patterns in enzyme-substrate interactions that correlate with catalytic rates.

The model was trained on over 16,000 experimentally measured kcat values from the BRENDA and SABIO-RK enzyme databases, covering both wild-type and engineered enzymes across diverse species.

How does DLKcat work?

DLKcat processes enzyme-substrate pairs through two parallel neural networks:

• Protein encoding: The enzyme sequence is split into overlapping 3-gram amino acid fragments. A CNN with three layers extracts features that capture sequence patterns associated with catalytic activity.
• Substrate encoding: The substrate's SMILES representation is converted to a molecular graph where atoms are nodes and bonds are edges. A GNN with three time steps propagates information through the graph, learning structural features relevant to enzymatic processing.

The outputs from both networks are concatenated and passed through fully connected layers to predict log2(kcat). Training used radius-2 substrate subgraphs and 20-dimensional vector embeddings.

How to use DLKcat online

ProteinIQ runs the pinned DLKcat 1.0.0 prediction program on managed compute without requiring a local installation.

Inputs

Input	Description
Enzyme sequence(s)	FASTA format protein sequences, up to 2,500 residues each. Up to 10 enzymes are supported per job.
Substrate(s)	SMILES strings or compound names, one per line, with up to 10 substrates per job. Every enzyme is paired with every substrate.

For specific enzyme-substrate pairs, switch to paired TSV input and provide one row per pair:

Column	Description
Substrate Name	Compound name used for labeling and PubChem lookup when the SMILES cell is blank or None
Substrate SMILES	Substrate SMILES, recommended for reliable predictions
Protein Sequence	Enzyme amino acid sequence

As in the published prediction program, any columns after the third are ignored, so a DLKcat output.tsv can be submitted again as input. Any other SMILES value, including an invalid one, is scored as given and returns None if DLKcat cannot read it.

Settings

Setting	Description
Input mode	Choose separate enzyme/substrate inputs for all combinations, or paired TSV rows for specific enzyme-substrate pairs.

Results

The output table contains predicted kcat values for each enzyme-substrate pair:

Column	Description
Substrate Name	The substrate label
Substrate SMILES	The substrate structure used for prediction
Protein Sequence	The full input protein sequence
Kcat value (1/s)	Turnover number in reactions per second
Error	Why a row has no prediction, when applicable

Each job is limited to 100 enzyme-substrate pairs and 250,000 aggregate residue-pairs. A job costs 6 credits for up to 50 prediction pairs and 7 credits for 51 to 100 pairs. The downloadable files include output.tsv, the source run log, and a provenance record with the exact source commit and model checksums.

Interpreting predictions

DLKcat predicts log2(kcat), then converts to linear scale. Predictions span a wide range:

kcat (1/s)	Interpretation
> 1000	Fast enzyme, typical of metabolic enzymes
100–1000	Moderate catalytic rate
10–100	Slow enzyme
< 10	Very slow, may indicate poor substrate match

The publication reports Pearson's r ≈ 0.71 and RMSE 1.06 on held-out test data. Performance is best for enzymes similar to the training set—predictions become less reliable as sequence similarity to training enzymes decreases.

Limitations

DLKcat has important constraints to consider:

Sequence similarity dependence: The model performs well when test enzymes share >70% sequence identity with training examples. For dissimilar enzymes, predictions may not outperform simple kcat averages.

No environmental factors: DLKcat was not trained with temperature or pH condition data, so predictions are based on enzyme sequence and substrate structure.

Mutant enzymes: While trained on some mutant data, predictions for novel mutations—especially those affecting catalytic residues—should be validated experimentally.

Substrate coverage: Predictions are most reliable for substrates similar to those in the BRENDA/SABIO-RK training data.

Single-fragment substrates: The published prediction program does not score disconnected multi-fragment SMILES such as salts. Those rows are retained in the result table with a None kcat value and an explanatory error.

Compound-name lookup: Substrate names are resolved through PubChem before prediction. If no SMILES can be resolved, the row is retained with a None kcat value.

Applications

Enzyme turnover predictions support several research applications:

• Metabolic modeling: Parameterizing enzyme-constrained genome-scale models (ecGEMs) with predicted kcat values
• Enzyme engineering: Screening mutant libraries computationally before experimental characterization
• Pathway design: Identifying rate-limiting enzymes in synthetic biology pathways
• Comparative enzymology: Analyzing kcat distributions across species or enzyme families