Abstract<\/strong><\/p>\n Organic chemistry has largely been conducted in an ad hoc manner by academic laboratories that are funded by grants directed towards the investigation of specific goals or hypotheses. Although modern synthetic methods can provide access to molecules of considerable complexity, predicting the outcome of a single chemical reaction remains a major challenge. Improvements in the prediction of \u2018above-the-arrow\u2019 reaction conditions are needed to enable intelligent decision making to select an optimal synthetic sequence that is guided by metrics including efficiency, quality and yield. Methods for the communication and the sharing of data will need to evolve from traditional tools to machine-readable formats and open collaborative frameworks. This will accelerate innovation and require the creation of a chemistry commons with standardized data handling, curation and metrics.<\/p>\n<\/div>\n<\/div>\n<\/section>\n \n<\/div>\n Main<\/strong><\/p>\n Now, the scepticism of synthetic chemists seems to be on the verge of changing. Using computer-aided synthesis planning (CASP), it is now possible to take the molecular structure of a desired product and output a detailed list of reaction schemes that connect the target molecule to known and often purchasable starting materials through a sequence of intermediates that are likely to be unknown5<\/a>,6<\/a><\/sup>\u00a0(Box\u00a01<\/a>). For example, the decision-tree-like search engine\u00a0Chematica<\/i>\u2014which has a user-friendly graphical user interface and has been coded with human-curated rules over the past decade\u2014has received laboratory validation of the predicted synthesis of medicinally relevant targets7<\/a><\/sup>. Approaches towards such programs usually reflect the priorities and prejudices of the programmers, and others have used different approaches\u2014for example, using machine-learning algorithms or Monte Carlo Tree Search (as in AlphaGo8<\/a><\/sup>) to guide the search, and a filter network to pre-select the most promising retrosynthetic steps that is trained on essentially all reactions ever published in organic chemistry9<\/a>,10<\/a>,11<\/a><\/sup>. In the future, it will be substantially faster for such programs to learn automatically from the primary data rather than rely on extracted rules and hand-designed heuristics, in analogy to the differences in strategy between Stockfish and AlphaZero in learning chess12<\/a><\/sup>.<\/p>\n The digitization of multistep organic synthesis is fast approaching, and the automation of the synthesis planning is just the first component that must be considered before automated reaction prediction can become a reality. The selection of reaction conditions is a key element of automated reaction prediction and is potentially a far more challenging task13<\/a><\/sup>\u00a0(Fig.\u00a01<\/a>). This Perspective surveys the current prospects for the prediction of above-the-arrow conditions and addresses the challenges that are involved in integrating them into optimal methods of synthesis. For one, it has been stated that \u201csyntheses are reported in prose\u201d14<\/a><\/sup>. Not only are the reactions conditions often poorly communicated, but details are also omitted when explaining exactly how operations were carried out, meaning that many assumptions are made about the skills of the researcher repeating the synthesis. The prediction problem must then consider an even broader range of variables in order to master or fully execute a synthesis or optimization, depending on the context of academic research and medicinal or process chemistry.<\/p>\n <\/p>\n
![\"figure1\"](\"\/\/media.springernature.com\/m685\/springer-static\/image\/art%3A10.1038%2Fs41586-019-1288-y\/MediaObjects\/41586_2019_1288_Fig1_HTML.png?as=webp\")
<\/picture><\/a><\/div>\n