A Generative Model For Electron Paths
Authors: John Bradshaw, Matt J. Kusner, Brooks Paige, Marwin H. S. Segler, José Miguel Hernández-Lobato
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We propose an electron path prediction model (ELECTRO) to learn these sequences directly from raw reaction data. Our model achieves excellent performance on an important subset of the USPTO reaction dataset, comparing favorably to the strongest baselines. Furthermore, we show that our model recovers a basic knowledge of chemistry without being explicitly trained to do so. |
| Researcher Affiliation | Collaboration | John Bradshaw University of Cambridge Max Planck Institute, Tübingen jab255@cam.ac.uk; Matt J. Kusner University of Oxford Alan Turing Institute mkusner@turing.ac.uk; Brooks Paige Alan Turing Institute University of Cambridge bpaige@turing.ac.uk; Marwin H. S. Segler Benevolent AI marwin.segler@benevolent.ai; José Miguel Hernández-Lobato University of Cambridge Microsoft Research Cambridge Alan Turing Institute jmh233@cam.ac.uk |
| Pseudocode | Yes | Algorithm 1 The generative steps of ELECTRO (given that the model chooses to react, ie c0 = 1). Input: Reactant molecules M0 (consisting of atoms A), reagents Me, atom embedding function h A( ), graph embedding functions greagent( ) and gcont( ), additional logit functions f start( ), f remove( ), f add( ), time steps T max; Algorithm 2 Predicting electron paths at test time. Input: Molecule M0 (consisting of atoms A), reagents Me , beam width K, time steps T max |
| Open Source Code | No | The paper does not explicitly provide a link or statement confirming that the source code for the methodology described in this paper is publicly available. Footnote 5 links to the code for baseline models, not the authors' own work. |
| Open Datasets | Yes | To evaluate our model, we use a collection of chemical reactions extracted from the US patent database (Lowe, 2017). We take as our starting point the 479,035 reactions, along with the training, validation, and testing splits, which were used by Jin et al. (2017), referred to as the USPTO dataset. Daniel Lowe. Chemical reactions from US patents (1976-Sep2016). 6 2017. doi: 10.6084/m9.figshare.5104873.v1. URL https://figshare.com/articles/Chemical_ reactions_from_US_patents_1976-Sep2016_/5104873. |
| Dataset Splits | Yes | We take as our starting point the 479,035 reactions, along with the training, validation, and testing splits, which were used by Jin et al. (2017), referred to as the USPTO dataset. Applying these steps, we discover that 73% of the USPTO dataset consists of LEF reactions (349,898 total reactions, of which 29,360 form the held-out test set). |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. It only mentions 'We train our models using Adam' without hardware info. |
| Software Dependencies | Yes | Each of these steps can be easily implemented with the open-source chemo-informatics software RDKit (RDKit, online). We train up the required models from scratch using the Open NMT library (Klein et al., 2017). |
| Experiment Setup | Yes | We train everything using Adam (Kingma and Ba, 2015) and an initial learning rate of 0.0001, which we decay after 5 and 9 epochs by a factor of 0.1. We train for a total of 10 epochs. For training we use reaction minibatch sizes of one, although these can consist of multiple intermediate graphs. |