Hierarchical Grammar-Induced Geometry for Data-Efficient Molecular Property Prediction
Authors: Minghao Guo, Veronika Thost, Samuel W Song, Adithya Balachandran, Payel Das, Jie Chen, Wojciech Matusik
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On both small and large datasets, our evaluation shows that this approach outperforms a wide spectrum of baselines, including supervised and pre-trained graph neural networks. We include a detailed ablation study and further analysis of our solution, showing its effectiveness in cases with extremely limited data. |
| Researcher Affiliation | Collaboration | 1MIT CSAIL 2MIT-IBM Watson AI Lab, IBM Research. |
| Pseudocode | Yes | Algorithm 1 Molecular Rules Construction. Input :molecular hypergraph H = (V, EH), probability function ϕ( ; θ) Output :molecular rule set Pmol = {pi}M i=1, junction tree T = (VT , ET ) where each v T VT is a subset of V; Algorithm 2 Grammar-induced Geometry Construction. Input :meta production rules PG = {pi}N i=1, maximum BFS depth D, a set of molecular hypergraphs {Hi}M i=1 and their corresponding junction trees J = {Ti}M i=1 Output :geometry in the form of a graph G = (V, E) where each v = Hv = (Vv, Ev) V represents a meta tree or a molecular hypergraph and E is the edge set of the geometry |
| Open Source Code | Yes | Code is available at https: //github.com/gmh14/Geo-DEG. |
| Open Datasets | Yes | We evaluate our approach on eight datasets: CROW (a curated dataset from literature), Permeability (Yuan et al., 2021), Free Solv (Mobley & Guthrie, 2014), Lipophilicity (Wang et al., 2015), HOPV (Lopez et al., 2016), DILI (Ma et al., 2020), PTC (Xu et al., 2018), and Clin Tox (Gayvert et al., 2016). |
| Dataset Splits | No | The paper states: 'For each dataset except Clin Tox, we randomly split the data into 4 : 1 training and testing sets and create five such splits using 5 random seeds.' This describes train/test splits but does not explicitly mention or specify a separate validation dataset or its proportions. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used, such as CPU or GPU models, memory specifications, or cloud computing instance types. It mentions using GIN and MPNN as feature extractors, but this refers to software components rather than hardware. |
| Software Dependencies | No | The paper mentions software components like 'GIN from (Xu et al., 2018)', 'MPNN from (Yang et al., 2019)', 'Adam optimizer', 'Weisfeiler Lehman graph hashing (Shervashidze et al., 2011)', 'networkx for graph isomorphism test', and 'pythonhosted/zss' for tree edit distance. However, it does not provide specific version numbers for any of these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | For our approach, we use 4-degree meta grammar, which contains eight rules. The meta geometry contains all the meta trees whose size is smaller than 11, resulting in 149 nodes and 455 edges. For the molecular rule learning of θ, we follow all the hyperparamters used in (Guo et al., 2022). For the graph diffusion, the input feature of each meta tree node is the Weisfeiler Lehman graph hashing feature (Shervashidze et al., 2011). The encoder for meta tree nodes is an embedding layer that maps hashing features into a 300-dimension continuous vector. For molecular leaves, we use two different encoders: GIN from (Xu et al., 2018) and MPNN from (Yang et al., 2019), both of which output a feature vector of dimension 300. For the decoder, we use a one-layer fully connected network with size 300. For the graph diffusion process, we follow (Chamberlain et al., 2021) and use Dormand Prince adaptive step size scheme (DIORI5) with adjoint method. The diffusivity function a( , ; α) is an attention function. We use Adam optimizer for the training of θ and (φ, ψ, α), with learning rate 0.01 and 0.001, respectively. We train θ for ten epochs. For each training epoch of θ, we train (φ, ψ, α) for 50 epochs. |