DimeNet

Overview

DimeNet is a research repository containing reference implementations of the DimeNet and DimeNet++ neural-network models for molecular graphs. It accompanies the papers Directional Message Passing for Molecular Graphs and Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules. Its workflow role is model training and molecular-property prediction, rather than molecular database access or a hosted prediction service. The README recommends DimeNet++ over the original model.

The models use molecular graph information with distances and angles in their directional message-passing calculations. The documented QM9 targets include alpha, R2, U0, U, H, G, Cv, Mu, HOMO, LUMO and ZPVE. The source excerpts do not specify a complete input schema or prediction-file format. Training is organized through train.ipynb, while predict.ipynb generates test-set predictions using a trained model. Two sets of pretrained models are provided in pretrained, and train_seml.py supports cluster training with Sacred and SEML.

This repository uses TensorFlow 2 and documents training and initialization differences from the original TensorFlow 1 implementation. Interpretation of results requires attention to the listed issues: an off-by-one molecule-filtering error in qm9_eV.npz, distance and basis-layer discrepancies, an embedding-layer discrepancy, and a checkpointing problem affecting older TensorFlow AddOns versions. The README also identifies an MD17 benchmark-comparison mismatch. For energy and force prediction, the authors instead recommend GemNet; this entry should therefore be treated as a documented research implementation, not an independently validated scientific workflow.

Key Features

• Includes reference implementations of both DimeNet and DimeNet++ for directional message passing on molecular graphs.
• Provides `train.ipynb` for model training and `predict.ipynb` for test-set prediction with a trained model.
• Supplies two sets of pretrained models in the `pretrained` folder for experimentation.
• Includes `train_seml.py` for cluster training using Sacred and SEML.
• Documents target-specific output-layer initialization in the TensorFlow 2 implementation.
• Lists known dataset, geometry, basis-layer, embedding-layer and checkpointing issues relevant to interpreting or reproducing results.

Use Cases

• Intended evaluation: use the training and prediction notebooks to assess a QM9 molecular-property prediction workflow, accounting for the documented molecule-filtering error.
• Intended evaluation: compare DimeNet and DimeNet++ within a controlled experimental setup without assuming that published results transfer to a new dataset.
• Intended evaluation: examine the supplied pretrained models before committing resources to retraining.
• Intended evaluation: assess `train_seml.py` as a starting point for cluster-based molecular-model experiments using Sacred and SEML.

How to Use

1. Read the official README to choose between DimeNet and DimeNet++. It recommends DimeNet++ and points energy-and-force users toward GemNet.
2. Review the dependencies in setup.py, including TensorFlow, TensorFlow AddOns, NumPy, SciPy and SymPy. Treat these as declared requirements, not evidence of compatibility with your environment.
3. Inspect train.ipynb, predict.ipynb and the pretrained folder in the repository. Check the expected data representation and target selection before attempting training or prediction.
4. Review the README's known issues and the QM9 filtering discussion. For an intended evaluation, record dataset provenance, initialization choices and any corrections separately from upstream results.
5. If cluster training is needed, examine train_seml.py and the linked SEML project. Review the code licence before use or redistribution.