Orbit-Equivariant Graph Neural Networks
Authors: Matthew Morris, Bernardo Cuenca Grau, Ian Horrocks
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 EXPERIMENTS |
| Researcher Affiliation | Academia | Matthew Morris, Bernardo Cuenca Grau, & Ian Horrocks Department of Computer Science University of Oxford {matthew.morris,bernardo.grau,ian.horrocks}@cs.ox.ac.uk |
| Pseudocode | Yes | The 1-WL algorithm (Weisfeiler & Leman, 1968) can be formalized algorithmically as follows: 1. Initialize C0 : V (G) C by C0(v) = c1 for each node v V (G), where C is a codomain of colors and c1 C. C0 is a node coloring function at iteration 0 of the algorithm. 2. At each iteration n of the algorithm (starting from n = 0), we define a new coloring function Cn+1. For each node v V (G), set Cn+1(v) := h(Cn(v), {{Cn(w) | w N(v)}}), where h is a hash function and N(v) is the set of all neighbors of v in G. 3. At each iteration n of the algorithm, also compute a partition Pn+1 of V (G), where for any nodes v, w V (G), v, w Q for any Q P if and only if Cn+1(v) = Cn+1(w). In other words, the nodes are partitioned by their colors. 4. If at any point, Pn+1 = Pn, the algorithm halts and returns h({{Cn(w) | w V (G)}}). |
| Open Source Code | Yes | Code, data, and instructions for how to reproduce our experimental results are given in the supplemental material. |
| Open Datasets | Yes | To construct this dataset, we retrieved 961 small drug molecules from Ch EMBL (Mendez et al., 2019) and then used MMPDB (RDKit, 2023a) to compute many different bioisosteres for each molecule that each differ by at most one atom from the original molecule. Using RDKit (RDKit, 2023b), we computed the lipophilicity of each bioisostere and selected the one that yielded the lowest value as the target for each molecule. |
| Dataset Splits | No | Every experiment is run 10 times, with different seeds, and 10% of the data is randomly left out of training to be used as a test dataset. |
| Hardware Specification | No | All experiments were run in a distributed manner on an internal university cluster, on exclusively CPU nodes. |
| Software Dependencies | No | No specific version numbers for key software components (like deep learning frameworks or libraries) were provided. |
| Experiment Setup | Yes | Fixed hyperparameters across all experiments are shown in Table 3. The fixed hyperparameters were chosen by preliminary optimization experiments of GCN on Bioisostere. Hyperparameters for the experiments on Bioisostere are shown in Table 4. Hyperparameters for the experiments on Alchemy Max-Orbit-2 are shown in Table 5. Hyperparameters for the experiments on Alchemy-Max-Orbit-6 are shown in Table 6. |