Directional Message Passing on Molecular Graphs via Synthetic Coordinates
Authors: Johannes Gasteiger, Chandan Yeshwanth, Stephan Günnemann
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that with this transformation we can reduce the error of a normal graph neural network by 55 % on the ZINC benchmark. We furthermore set the state of the art on ZINC and coordinate-free QM9 by incorporating synthetic coordinates in the SMP and Dime Net++ models. Our implementation is available online. 1 |
| Researcher Affiliation | Academia | Johannes Gasteiger, Chandan Yeshwanth, Stephan Günnemann Technical University of Munich, Germany {j.gasteiger,yeshwant,guennemann}@in.tum.de |
| Pseudocode | No | The paper describes the model architecture and equations but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our implementation is available online. 1 https://www.daml.in.tum.de/synthetic-coordinates |
| Open Datasets | Yes | We use three common benchmarks to evaluate the proposed synthetic coordinates: Coordinate-free QM9 (Ramakrishnan et al., 2014, CC0 license), ZINC (Irwin et al., 2012), and ogbg-molhiv (Hu et al., 2020, MIT license). |
| Dataset Splits | Yes | We use the same data split as Brockschmidt (2020) for QM9, i.e. 10 000 molecules for the validation and test sets, and the remaining 110 000 molecules for training. ... We use 10 000 training, 1000 validation, and 1000 test molecules, as established by Dwivedi et al. (2020)... For ogbg-molhiv it contains 41 127 graphs, out of which 80 % are training samples, and 10 % each are validation and test samples, as provided by the official data splits. |
| Hardware Specification | Yes | The experiments were run on GPUs using an internal cluster equipped mainly with NVIDIA Ge Force GTX 1080Ti. |
| Software Dependencies | No | The paper mentions "Py Torch Geometric" but does not specify version numbers for Python, PyTorch, or other libraries/solvers. |
| Experiment Setup | Yes | More specifically, we use the Deeper GCN (Li et al., 2020) with 12 Res GCN+ blocks, mean aggregation in the graph convolution, and average pooling to obtain the graph embedding. For SMP (Vignac et al., 2020) we use 12 layers, 8 towers, an internal representation of size 32 and no residual connections. For both Deeper GCN and SMP we use an embedding size of 256, and distance and angle bases of size 16 and 18, respectively, with a bottleneck dimension of 4 between the global basis embedding and the local embedding in each layer. We train all models on ZINC with the same training hyperparameters as SMP, particularly the same learning rate schedule with a patience of 100 and minimum learning rate of 1 10 5. For Dime Net++ we use a cutoff of 2.5 Å, radial and spherical bases of size 12, embedding size 128, output embedding size 256, basis embedding size 8 and 4 blocks. We use the same optimization parameters learning rate 0.001, 3000 warmup steps and a decay rate of 0.01. |