Chapter4: Deep Learning and Chemical Data

reading notes of《Artificial Intelligence in Drug Discovery》

---

1.Introduction

• Indeed one of the earliest applications of deep learning in chemistry was in an entry to a Kaggle competition in 2012 where the data was supplied by Merck.
• Neural nets are difficult to interpret and are invariably thought of by their developers as “black boxes”, something which is also the case for some other machine- learning methods such as tree ensembles and generalised linear models. This has legal consequences in the European Union with the advent of GDPR and the “right to explanation”.
• A neural network trained using high-precision arithmetic can also be replaced by one that uses eight-bit or even single-bit precision with only a modest decrease in accuracy.

2.Background

2.1.Deep Learning

• Another neural network architecture that has been used in the past in cheminformatics but is no longer popular is the counterpropagation neural network. They have been used in the past in cheminformatics but are more complicated to train as more than one algorithm is involved. Aires-de-Sousa et al. give a good description of how they work in an early application to NMR spectrum elucidation.