The primary goal of this work is to compare different deep learning models for molecular property prediction using the ZINC dataset.
Dataset: The ZINC dataset is a collection of commercially available chemical compounds. It provides molecular structures and associated properties, which are used for training and evaluating the models.
Model Descriptions:
- Model 1: GIN (Graph Isomorphism Network):
- A GIN is designed to predict molecular properties directly from the graph representation of molecules.
- Model 2: Pretrained and Fine-tuned Model:
- This model is first pretrained on a node classification task to learn general features of molecular graphs.
- It is then fine-tuned specifically for the property prediction task.
- Model 3: Multitask Learning Model:
- This model simultaneously learns two tasks: node classification and property prediction.
- It shares some components between the tasks to leverage multitask learning benefits.
Code Attribution: This code is adapted from deepfindr/gnn-project.
Training and Validation:
- The training procedure includes splitting the data into training, validation, and test sets.
- Hyperparameters, loss functions, and optimization techniques are kept consistent across models for fair comparison.
Evaluation Metrics:
- The primary evaluation metric is Mean Squared Error (MSE) for property prediction.
Performance Comparison:
- The following table summarizes the MSE results for all three models:
| Model | Mean Squared Error (MSE) |
|---|---|
| GIN | 0.133 |
| Pretrained and Fine-tuned | 0.120 |
| Multitask Learning | 0.143 |
- The pretrained and fine-tuned model achieved the lowest MSE.