CrossKP: integrating protein, substrate and reaction information for enzyme kinetic parameter prediction

CrossKP predicts k_cat and K_m for enzyme–substrate pairs. It combines a bidirectional cross-attention neural network with an ExtraTrees branch. Their predictions are combined using fixed weights of 0.40 and 0.60, respectively.

CrossKP is available for download.

• CrossKP model package (CrossKP.zip, 129 MB)

Package contents

The downloaded package includes:

• ten neural-network checkpoints for each prediction task;
• code for rebuilding and applying the ExtraTrees branch;
• the fixed 0.40 neural-network and 0.60 ExtraTrees fusion rule;
• feature assembly, prediction and training scripts;
• fixed data splits, row-identity checks and small runnable examples.

The following large or third-party files are not included:

• serialized ExtraTrees models and full task-level feature matrices;
• ProtT5, ESM-2 and MolT5 model files;
• P2Rank outputs, AlphaFold structures and original datasets.

The supplied neural-network weights can be used directly with correctly assembled features. The complete two-branch estimator requires rebuilding ExtraTrees with the included code.

Instructions for CrossKP installation

Download and unpack the CrossKP package, then install the listed dependencies.

unzip CrossKP.zip
cd CrossKP
python -m pip install -r requirements.txt

Optional dependencies for generating the 167-dimensional structure descriptor are listed in requirements-feature-extraction.txt. A GPU is recommended for neural-network inference and training.

Quick start

Run the packaged examples and integrity checks:

python smoke_test.py

Run neural-network prediction for the example k_cat inputs:

python predict.py --task kcat \
  --features examples/kcat_neural_features.npy \
  --row-ids examples/kcat_row_ids.npy \
  --mode neural --output predictions.csv

The output CSV contains predictions in log10 space and values converted back to the original units. Instructions for rebuilding ExtraTrees and reproducing training are provided in the package README files.

Input feature layouts

Neural network (4,174 columns):

[ProtT5(1024) | ESM-2(1280) | structure(167)] | [MolT5(768) | MACCS(167) | reaction(768)]

ExtraTrees (4,039 columns):

[ProtT5(1024) | ESM-2(1280) | structure(32)] | [MolT5(768) | MACCS(167) | reaction(768)]

The 167-dimensional descriptor summarizes a candidate protein pocket and an independently generated substrate conformer. It does not represent a docked enzyme–substrate pose or atom-level intermolecular contacts.

Intended use and license

CrossKP is intended for research, benchmarking and candidate prioritization. Its predictions are not substitutes for experimental kinetic measurements.

No public license has yet been selected for CrossKP. License information and third-party notices are included in the downloaded package.