MPNNs for Modeling Retention Times from Mixed Mode Chromatography / ESI Negative MS Detection

This repository contains code used to predict retention times given chemical structures (in the form of SMILES).

Environment

The script environment_setup.sh contains instructions for setting up a conda environment with the appropriate versions of Tensorflow and DeepChem needed to run the models.

Data

The raw data used to train and evaluate the model is tracked with Git-LFS in data/combined_time.csv. Each row describes the structure (SMILES in column smiles) and experimental results for a given compound. The column neg RT (min) in this file records the retention time for each compound.

The file data/classes.py contains a list of compound classes used to bucket the compounds in data/combined_time.csv.

Jupyter Notebooks

The notebooks in the notebooks folder illustrate the process of training and evaluating the MPNN models as well as the steps to create the figures in our publication.

1. DeepChem01.ipynb: Train and test a message passing neural network on the dataset (random split)
2. Manuscript_Figures.ipynb: Evaluate the performance of the final MPNN (after hyperparameter tuning) on the dataset
3. contributions_00.ipynb: Generate per-atom contributions by running inference on masked molecules
4. contributions_01.ipynb: Analyzing and generating graphics of per-atom contributions

Code

The src directory contains python files describing the deepchem model and for carrying out hyperparameter searches.