Images of chemical structures as molecular representations for deep learning

Abstract

Implementing Artificial Intelligence for chemical applications provides a wealth of opportunity for materials discovery, healthcare and smart manufacturing. For such applications to be successful, it is necessary to translate the properties of molecules into a digital format so they can be passed to the algorithms used for smart modelling. The literature has shown a wealth of different strategies for this task, yet there remains a host of limitations. To overcome these challenges, we present two-dimensional images of chemical structures as molecular representations. This methodology was evaluated against other techniques in both classification and regression tasks. Images unlocked (1) superior augmentation strategies, (2) application of specialist network architectures and (3) transfer learning, all contributing to superior performance and without prior specialised knowledge on cheminformatics required. This work takes advantage of image feature maps which do not rely on chemical properties and so can represent multi-component systems without further property calculations. Graphical abstract

Predicting pharmaceutical crystal morphology using artificial intelligence

“…Recent work assessed different molecular representations for use in chemical applications focusing specifically on tasks relevant to the field of solid form engineering. 11 This demonstrated that images of chemical structures offer the best accuracy among the methods tested. Despite this, the work is recent, and as such other chemical applications beyond those presented have not been tested using images as inputs.…”

“…ResNet models were chosen as a result of previous evidence of image-based deep learning models used in chemical applications and wider image recognition tasks. 11,13 These models were trained and evaluated using stratified cross-validation, from which the mean accuracy was recorded across the splits. The batch size was 256 for the CSD dataset and 8 for the in-house experimental dataset.…”

“…In this work, the process of representing molecules was approached using two methods: (i) chemical descriptors and (ii) images of chemical structures. Recent work assessed different molecular representations for use in chemical applications focusing specifically on tasks relevant to the field of solid form engineering 11 . This demonstrated that images of chemical structures offer the best accuracy among the methods tested.…”

“…When using images to represent molecules, ResNet neural models were used, and when using chemical descriptors, random forests were applied. ResNet models were chosen as a result of previous evidence of imagebased deep learning models used in chemical applications and wider image recognition tasks 11,13 . These models were trained and evaluated using stratified cross-validation, from which the mean accuracy was recorded across the splits.…”

“…It can be appreciated that images displayed much higher accuracy scores than chemical descriptors across all case studies. This performance difference is a result of images' improved ability to distinguish molecular features as highlighted across wider chemical applications 11 . Indeed, by leveraging the information provided by 99 data points, this work demonstrates the advantages of transfer learning for solid-state applications, an area where big data is (at present) exceedingly scarce.…”