Chemception: Deep Learning from 2D Chemical Structure Images

ConclusionI doubt that the current accuracy of Chemception's predictions would be of practical use today. Rather, Chemception provides a platform from which such systems may eventually emerge. Recent history suggests that such an emergence may be closer than it seems.Chemception offers a glimpse into a future in which lightly processed chemical datasets can be fed directly into off-the-shelf data learning pipelines to yield highly accurate predictive models. In this future, an iteratively hand-crafted molecular representation is no longer necessary. Instead, the system adapts itself to a much more raw form of structural data, identifying and classifying molecular features on its own and in a matter that may significantly diverge from human intuition.Chemception also hints at the potential for deepening insights into numerous kinds of structure-property relationships. Already it's clear that some predictions will be more amenable to this streamlined approach than others. Could, for example, well-suited problems have some deeper, as yet unidentified, physical connection? Investigations along these lines may be aided by probing the convolutional layers generated during training.Finally, it's worth mentioning the uncanny way in which Chemception seems to operate. For many chemists, and medicinal chemists in particular, 2D structure images are the bread-and-butter for the work they do. Structure images are information-rich, readily parsed, and instantly recognizable. They are the ...
Source: Depth-First - Category: Chemistry Authors: Source Type: blogs