DeltaKcat

Introduction

DeltaKcat is a pairwise relative learning framework for robust enzyme kinetics prediction. Instead of directly regressing noisy absolute kinetic values, DeltaKcat learns relative differences between enzyme–substrate complexes within consistent experimental contexts. This approach improves generalization to unseen enzymes and substrates and supports the discovery of highly active enzyme candidates.

1. Environment setup

We recommend creating an isolated Conda environment to ensure consistent package versions.

conda create -n DeltaKcat python==3.11
conda activate DeltaKcat

1.1 Dependencies

We used the following Python packages for core development. We tested on Python 3.11.

name	version
pandas	2.0.3
networkx	3.5
tqdm	4.66.5
torch	2.2.0
torch-geometric	2.7.0
scipy	1.16.2
transformers	4.40.0
scikit-learn	1.2.0
bio	1.8.1
numpy	1.24.4
rdkit
SentencePiece

Make sure that the version of numpy is 1.24.4. Installing other packages may cause errors due to an upgrade in the numpy version.

In addition, DeltaKcat also relies on additional pre-trained models, including ProtT5 and MolT5. These two models are used for extracting features from enzymes and substrates, respectively.

2. Prediction Workflow

Download and unzip the model parameters under DeltaKcat

2.1 Sequence-delta prediction

Script: Train/Predict_Sequence.py

python Train/Predict_Sequence.py `
  --input_csv /path/predict_seq.csv `
  --protein_model_dir /path/prot_t5_xl_uniref50 `
  --substrate_model_dir /path/molt5-base-smiles2caption `
  --deltakcat_model_path /path/Delta_seq_model.pth `
  --output_csv /path/predict_seq_out.csv `
  --device cuda:0

Default input columns:

• Sequence_1
• Sequence_2
• Smiles

Optional column override:

• --sequence1_col
• --sequence2_col
• --smiles_col

2.2 Substrate-delta prediction

Script: Train/Predict_Substrate.py

python Train/Predict_Substrate.py `
  --input_csv /path/predict_sub.csv `
  --protein_model_dir /path/prot_t5_xl_uniref50 `
  --substrate_model_dir /path/molt5-base-smiles2caption `
  --deltakcat_model_path /path/Delta_sub_model.pth `
  --output_csv /path/predict_sub_out.csv `
  --device cuda:0

Default input columns:

• Sequence
• Smiles_1
• Smiles_2

Optional column override:

• --sequence_col
• --smiles1_col
• --smiles2_col

3. Training Workflow

Training scripts:

• Sequence-delta task: Train/train_delta_sequence.py
• Substrate-delta task: Train/train_derta_substrate.py

Both scripts expect this embedding layout under one embedding_path:

• sequences_prottrans.fasta
• prottrans_feature_dict.pt
• SMILES/SMILES_embedding_molT5.pkl

3.1 Generate protein embeddings (ProtT5)

Script: Utils/protein_embeddings_protT5.py. ProtT5 model could be download from https://huggingface.co/Rostlab/prot_t5_xl_uniref50.

python Utils/protein_embeddings_protT5.py `
  --input_csv /path/data_file `
  --output_dir /path/embedding_dir `
  --model_dir /path/prot_t5_xl_uniref50 `
  --sequence_col Sequence `
  --device cuda:0

Notes:

• Writes sequences_prottrans.fasta and prottrans_feature_dict.pt into output_dir.
• Default --sequence_col is Sequence.
• If your data has Sequence_1 and Sequence_2, prepare a temporary CSV with one merged Sequence column.

3.2 Generate SMILES embeddings (MolT5)

Script: Utils/smiles_embeddings_molt5.py. molt5 model could be download from https://huggingface.co/laituan245/molt5-base-smiles2caption.

python Utils/smiles_embeddings_molt5.py `
  --input_csv /path/data_file `
  --output_dir /path/embedding_dir `
  --model_dir /path/molt5-base-smiles2caption `
  --smiles_col Smiles

Notes:

• Writes /path/embedding_dir/SMILES/SMILES_embedding_molT5.pkl.
• Use the same output_dir as protein embedding step.

3.3 Train models

3.3.1 Train sequence-delta model

python Train/train_delta_sequence.py `
  --train_path /path/Train_pair_seq.csv `
  --test_path /path/Test_pair_seq.csv `
  --embedding_path /path/embedding_dir `
  --output_path /path/result_dir/ `
  --device 0 `
  --batch_size 64 `
  --epoch 100 `
  --lr 5e-5

Input CSV should include at least:

• Sequence_1
• Sequence_2
• Smiles
• Value_1
• Value_2
• log2_delta_kcat

3.3.2 Train substrate-delta model

python Train/train_derta_substrate.py `
  --train_path /path/Train_substrate.csv `
  --test_path /path/Test_substrate.csv `
  --embedding_path /path/embedding_dir `
  --output_path /path/result_dir/ `
  --device 0 `
  --batch_size 64 `
  --epoch 100 `
  --lr 5e-5

Input CSV should include at least:

• Sequence
• Smiles_1
• Smiles_2
• Value_1
• Value_2
• log2_delta_kcat

4. Common Issues

4.1 CUDA not available

• Use --device cpu.

4.2 Missing prediction values

• Invalid sequences/SMILES are skipped during preprocessing.
• Check script summary logs for skipped row count.

Contact

• Feiran Li (@feiranl), Tsinghua University, Shenzhen, China
• Xiongwen Li (@xiongwenL), Tsinghua University, Shenzhen, China