wD-MPNN for Polymer Property Prediction

This repository contains the weighted, directed message passing neural network (wD-MPNN) used for polymer property prediction in the paper "A graph representation of molecular ensembles for polymer property prediction" (Chem. Sci. 2022,13, 10486-10498). For the reference Chemprop code, please go to https://github.com/chemprop/chemprop.

Usage

For general Chemprop usage, please refer to its documentation: https://chemprop.readthedocs.io/en/latest/.

When providing polymer structures as input rather than molecules, add the --polymer flag to the chemprop_train call, e.g.:

chemprop_train --data_path input.csv --dataset_type regression --save_dir chemprop_checkpoints --polymer

In the example above, the file input.csv would contain an extended string representation that includes information on the average repeating unit of the polymer. This is an example of such input:

[*:1]c1ccc2c(c1)S(=O)(=O)c1cc([*:2])ccc1-2.[*:3]c1ccc([*:4])c(N)c1|0.25|0.75|<1-3:0.25:0.25<1-4:0.25:0.25<2-3:0.25:0.25<2-4:0.25:0.25<1-2:0.25:0.25<3-4:0.25:0.25<1-1:0.25:0.25<2-2:0.25:0.25<3-3:0.25:0.25<4-4:0.25:0.25

• The first part ([*:1]c1ccc2c(c1)S(=O)(=O)c1cc([*:2])ccc1-2.[*:3]c1ccc([*:4])c(N)c1) contains the SMILES strings for all monomers in polymer. In this example there are 2 monomers separated by a ..
• The relative abundance of each monomer, reflecting their stoichiometry, is provided in the substring |0.25|0.75|, which indicates a ratio of 1:3 for the first and second monomers provided in the SMILES string, respectively.
• In the SMILES string, numbered wildcards (e.g., [*:3]) are used to indicate the location of possible connections within or between monomers. All these possible extra bonds, and their weights (i.e., frequencies) are specified in <1-3:0.25:0.25<1-4:0.25:0.25< .... The information for each possible bond (in the example above we have 10 possible bonds, both within and between monomers) is specified after a < symbols, e.g. <1-3:0.25:0.75. This would mean that there can be a bond between the atom connecting to [*:1] and the atom connecting to [*:3], that the directed edge 1->3 has a weight of 0.25, and that the directed edge 3->1 has a weight of 0.75. If both the forward and backward edge have the same probability of occurring, the same weight is simply repeated while still separating the weights with :, e.g., <1-3:0.5:0.5.
• The degree of polymerization can also be provided as part of the input, by appending it at the end of the input string after the symbol ~, e.g. ...<3-4:0.5:0.5~502.5. This will be internally transformed into the weighting factor 1 + log(Xn) discussed in the paper.

Citation

If you use this code, please cite:

@article{wdmpnn,
         title={A graph representation of molecular ensembles for polymer property prediction}, 
         author={Matteo Aldeghi and Connor W. Coley},
         journal="Chem. Sci.",
         year="2022",
         volume="13",
         issue="35",
         pages="10486-10498",
         publisher="The Royal Society of Chemistry",
         doi="10.1039/D2SC02839E",
         url="http://dx.doi.org/10.1039/D2SC02839E"
}