Скачать или смотреть Message-Passing Neural Networks for Molecular Property Prediction Using Chemprop

2022.04.13, Kevin Greenman, Massachusetts Institute of Technology (MIT)
Chemprop demo tool can be found at: https://nanohub.org/tools/chempropdemo
Part of Hands-on Data Science and Machine Learning Training Series at: https://nanohub.org/groups/ml/handson...

Table of contents below.

Chemprop is an open-source implementation of a directed message passing neural network (D-MPNN) that has been demonstrated to be successful in predicting a variety of molecular properties, including solvation properties, optical properties, infrared spectra, and toxicity. The framework has also been used to predict reaction properties and to identify new antibiotics. Chemprop can be used via a command-line interface or within a Python script or Jupyter notebook. In this tutorial, we will discuss the D-MPNN algorithm and several successful use cases, followed by a demonstration of some of Chemprop’s core functionalities (e.g. training and testing models for single- and multiple-molecule properties and reactions, transfer learning, and estimating uncertainty).

Table of Contents:
00:00 Message-Passing Neural Networks for Molecular Property Preduction Using Chemprop
01:41 What is Chemprop?
02:29 Acknowledgements
03:13 What we'll cover today
03:49 How can we represent a molecule as a vector?
05:10 D-MPNN method
05:49 In Contrast – Fixed Encodings
06:31 MPNN – Trainable Encodings
06:37 Chemprop Structure
07:52 Dataset Types and Targets
08:16 Data Input Formatting
09:04 Training
09:38 Major Hyperparameters
10:07 Data Splitting
10:52 Data Splitting
11:27 Loss Functions
11:45 Multiple Molecules
12:45 Predicting
13:08 Fingerprinting
13:48 Fingerprint Type
14:03 Heuristics
15:52 Features Not Covered Today
16:47 Papers Using Chemprop
17:13 Papers Using Chemprop (cont)
17:32 Resources
18:20 Chemprop

This presentation and related downloads can be found at: https://nanohub.org/resources/36082