TripK_a: Accurate and Scalable Molecular Ionization Property Prediction via Triplet Interaction Networks and Physical Knowledge

This work is based on the paper TripK_a: Accurate and Scalable Acid–Base Dissociation Property Prediction via Triplet Interaction Networks and Physical Knowledge.

TripK_a is an advanced macro- and micro-pK_a predictor designed to address the challenges of modeling complex atomic interactions and limited high-quality data in molecular discovery. By leveraging a triplet interaction network and integrating quantum-level physical knowledge, TripK_a effectively captures cooperative and competitive effects in multi-site protonation equilibria. It achieves state-of-the-art performance on the SAMPL6–8 and Novartis datasets. Pretrained on pK_a data, TripK_a serves as a robust foundation for molecular property prediction and interaction modeling, providing deeper insights into molecular ionization for downstream applications.

Installation

TripKa requires a customized Uni-Core 0.0.1: first install official Uni-Core compatible with your Python, PyTorch, and CUDA environment, then overlay the bundled unicore/ and unicore_cli/ sources onto the installed packages as described in the configuration instructions. The official package alone does not include the TripKa-specific modifications.

Datasets

Datasets and checkpoints are available at Google Drive. Download and unzip all files, then place data files in the ./Datasets/pickle directory, as well as checkpoints in the ./checkpoint/ directory.

Data Preprocessing

To preprocess the data, navigate to the TripKa directory and run the process.py script with the appropriate arguments:

cd TripKa
python Datasets/process.py --name <str> --num-conf <int> --mm

Parameters

• --name: Specifies the dataset name for preprocessing (e.g., dwar_small).
• --num-conf: Defines the number of conformers to process (integer, default: 1).
• --mm: Enables conformer generation using MMFF optimization (default: False).

Training

Pre-training

First, download and place chembl_train.pickle and chembl_valid.pickle in ./Datasets/pickle. Then preprocess the ChEMBL dataset and run the pretraining script:

cd TripKa
python Datasets/process.py --name "chembl_train" --num-conf 6 --mm
python Datasets/process.py --name "chembl_valid" --num-conf 6 --mm
bash scripts/pretrain_tripka.sh

Fine-tuning

After pretraining, TripK_a is ready for fine-tuning.

Fine-tuning with pK_a Data Only

Preprocess the dwar-iBond dataset and execute the fine-tuning script:

python Datasets/process.py --name "dwar_small" --num-conf 6
bash scripts/finetune_tripka.sh

Fine-tuning with Quantum-Level Knowledge Guidance

After initial fine-tuning on pK_a data, further fine-tune TripK_a with quantum-level knowledge guidance. First, download the QM-features data and place it in the appropriate directory (update the path in the script), then run:

bash scripts/finetune_tripka_qm.sh

Inference

To perform macro- or micro-pK_a predictions on new or standalone data, ensure the base and QM-guided TripK_a models are trained. Update the weight paths in scripts/infer_tripka_base CONDITIONS.sh and scripts/infer_tripka_qm.sh, then use the following commands.

Macro-pK_a Prediction

python infer_pKa.py --smiles "<SMILES_STRING>" --dataset "molecule_name" --iter-num 4 --sampl-num 10 --cuda-idx 0 --mode "A/B"

Micro-pK_a Prediction

python infer_pKa.py --micro-a "acid_micro_state_smiles" --micro-b "base_micro_state_smiles" --dataset "molecule_name" --sampl-num 10 --cuda-idx 0

Parameters

• --name: Specifies the dataset name for preprocessing (e.g., dwar_small).
• --num-conf: Defines the number of conformers to process (integer, default: 1).
• --mm: Enables conformer generation using MMFF optimization (default: False).

Testing

Base Model Testing

Preprocess the test dataset and run inference on the base model. For example, on the novartis_a test set:

for ((i=0; i<10; i++)); do
    python Datasets/process.py --name "novartis_a" --num-conf 1 --i $i
done
bash scripts/infer_tripka_base.sh novartis_a 10 ETKDG 0 8

QM-Guided Model Testing

Run inference for the QM-guided TripK_a model:

for ((i=0; i<10; i++)); do
    python Datasets/process.py --name "novartis_a" --num-conf 1 --i $i --mm
done
bash scripts/infer_tripka_qm.sh novartis_a 10 MM 0 8

Acknowledgement

This project builds upon the codebases of Uni-pKa and TGT. We express our gratitude to the authors for their valuable contributions.

Citation

If you use TripK_a in your research, please cite:

@article{wei2026tripka,
  title   = {{TripK}$_a$: Accurate and Scalable Acid--Base Dissociation Property Prediction via Triplet Interaction Networks and Physical Knowledge},
  author  = {Wei, Wentao and Rao, Jiahua and Xue, Bai and Xie, Jiancong and Xu, Dahao and Sun, Xichen and Lu, Yutong and Wang, Yu and Yang, Mingjun and Yang, Yuedong},
  journal = {Journal of Chemical Information and Modeling},
  year    = {2026},
  volume  = {66},
  number  = {8},
  pages   = {4525--4537},
  doi     = {10.1021/acs.jcim.6c00330},
  url     = {https://pubs.acs.org/doi/10.1021/acs.jcim.6c00330}
}