Adapting Differential Molecular Representation with Hierarchical Prompts for Multi-label Property Prediction
Official Pytorch implementation of the paper: "Adapting Differential Molecular Representation with Hierarchical Prompts for Multi-label Property Prediction".
HiPM is an innovative hierarchical prompted multi-label molecular representation learning framework designed to effectively manage the complexities of multi-label molecular property prediction. This framework comprises two main components: the Molecular Representation Encoder (MRE) and the Task-Aware Prompter (TAP). The novel use of hierarchical prompts within TAP allows our model to learn task-specific prompts while also capturing shared information across multiple tasks. By constructing a hierarchical prompt tree based on task affinity—measured through the angle of gradients—our method uncovers latent associations between molecular labels. This approach mitigates the negative transfer effects often seen in multi-label tasks and enhances model interpretability. Moreover, HiPM adapts to complex multi-label scenarios without requiring extensive task-specific tuning.
- python == 3.10.11
- dgl == 2.0.0.cu116
- dgllife == 0.3.2
- rdkit == 2023.12.25
- torch == 2.0.1
You can run the following commands line by line:
conda create -n HiPM python=3.10.11
conda activate HiPM
pip install torch==2.0.1
conda install -c dglteam/label/cu116 dgl
pip install dgllife
pip install rdkit
pip install optuna
For the dataset we used, we have uploaded them to data.
To train a model from scratch, run:
python main.py --dataset <datasetname> --device cuda:0 --epoch 200 --mode tune --n_trials 20
| Model | Clintox | SIDER | Tox21 | ToxCast |
|---|---|---|---|---|
| HiPM |
| Model | QM8 | QM9 |
|---|---|---|
| HiPM |
We randomly selected three seeds for each experiment with random scaffold splitting, then reported the mean and standard deviation of AUC or MAE values.