A Machine Learning Model trained on a High-Throughput Antibacterial Screen Increases the Hit Rate of Drug Discovery

Introduction

This repository provides trained weights and materials associated with the manuscript "A Machine Learning Model trained on a High-Throughput Antibacterial Screen Increases the Hit Rate of Drug Discovery".

The code of directed message passing neural network (D-MPNN) used in this study was implemented in the open-source package Chemprop, which is an effective graph neural network for molecular property prediction. The introduction, implementation, and application of D-MPNN can be found in below links:

• Yang et al (2019). Analyzing Learned Molecular Representations for Property Prediction. JCIM, 59(8), 3370–3388. https://doi.org/10.1021/acs.jcim.9b00237
• The original code: https://github.com/chemprop/chemprop
• Stokes et al (2020). A Deep Learning Approach to Antibiotic Discovery. Cell, 180(4), 688–702.e13. https://doi.org/10.1016/j.cell.2020.01.021

Result summary

Classification

Random split

Model	Additional Features	ROC-AUC	PRC-AUC
1	No molecule-level features	0.788	0.186
2	RDKit descriptors	0.795	0.232
3	RDKit descriptors + count-based Morgan fingerprints	0.808	0.091
4	RDKit descriptors + binary Morgan fingerprints	0.815	0.238

Scaffold split

Model	Additional Features	ROC-AUC	PRC-AUC
5	No molecule-level features	0.845	0.170
6	RDKit descriptors	0.823	0.241
7	RDKit descriptors + count-based Morgan fingerprints	0.784	0.191
8	RDKit descriptors + binary Morgan fingerprints	0.820	0.180

Regression

Random split

Model	Additional Features	RMSE	MAE
9	No molecule-level features	2.326	1.125
10	RDKit descriptors	2.290	1.104
11	RDKit descriptors + count-based Morgan fingerprints	2.806	1.130
12	RDKit descriptors + binary Morgan fingerprints	2.818	1.094

Scaffold split

Model	Additional Features	RMSE	MAE
13	No molecule-level features	2.823	1.135
14	RDKit descriptors	2.548	1.138
15	RDKit descriptors + count-based Morgan fingerprints	2.508	1.077
16	RDKit descriptors + binary Morgan fingerprints	2.564	1.058