XIMP: Cross Graph Inter-Message Passing

Requirements

• Linux with an x86 Processor

Installation

Install all dependencies in a conda environment defined in environment.yml

conda env create -f environment.yml

and activate the environment

conda activate ximp

Experiments

Every individual run will output a unique file in the results folder (ensure this folder exists before running), containing the hyperparameter configuration, the mean validation loss, the standard deviation on the validation loss, and the MAE on the test scaffold.

Single Run

To execute a single run, pass the hyperparameter configuration as flags to main.py. Details on default values and available options can be shown by calling the help function:

python main.py --help

Example 1: Run a sample run with the default hyperparameter configuration:

python main.py

Example 2: Use XIMP to perform molecular property prediction on the admet task, specifically the MLM endpoint. Do not use an ErG tree abstraction, but do use a junction tree abstraction with a coarseness values of 2

python main.py --repr_model="XIMP" --task="admet" --target_task="MLM" --use_erg="FALSE" --use_jt="TRUE" --jt_coarsity=2

Batch Run

To execute multiple hyperparameter configurations in parallel, use main_batch.py and define the hyperparameters to be used in a csv file. Sample hyperparamters to reproduce the results shown in the paper can be found in the hyperparameters folder. Be sure to set the hyperparameter filename at the top of the main_batch.py file. Default: "global_best_params.csv".

python main_batch.py

evaluation.ipynb lets you evaluate the results to reproduce the tables mentioned in the paper.

Datasets

We investigate 2 datasets, each containing multiple regression tasks:

MoleculeNet:

• ESOL
• FreeSolv
• Lipophilicty

Polaris:

• HLM
• KSOL
• LogD
• MDR1-MDCKII
• MLM
• pIC50 (MERS-CoV Mpro)
• pIC50 (SARS-CoV-2 Mpro)