Advances and challenges in the prediction of atom transition maps

Understanding how complex phenotypes emerge from large-scale biochemical networks necessitates accurate representation of biochemical reactions that extend beyond mass and charge balances. Atom transition maps provide atom-level representation of biochemical reactions that facilitate discovery of design principles in biochemical networks and, in combination with labeling experiments, enable flux estimation. In this systematic review, we categorize, compare, and contrast the existing computational approaches for the prediction of atom transition maps both on the level of individual reactions as well as entire biochemical networks. Emphasis is placed on the growing number of approaches that predict atom transition maps using deep learning alone or in combination with conventional optimization-based approaches. We stress the problems of data leakage typically arising in prediction of atom transition maps using the existing machine and deep learning approaches for reactions that share mechanisms. Importantly, we illustrate the severity of the problem by re-examining representative deep learning models for the prediction of atom transition maps on a unified test set of reactions. The review of these advances alongside our comparative analysis points at open problems in generation of atom transition maps and paves the way for future research in this area.