Originality: The proposed model (Molecule-Chef) is a novel combination of existing deep learning models - like Graph Neural Networks, RNNs, etc. to solve the task of molecule searching and to provide a synthesis recipe for the same. The authors also ensure the validity of products by restricting the latent space to chemical reactants that are readily available to chemists. Quality: The authors compare Molecule-Chef with state-of-the-art baselines and report improved/comparable results. By restricting the latent space, the model produces more valid molecular products as compared to other models. The paper also presents results of retrosynthesis, in which given some molecular products, the decoder of Molecule-Chef can be used to generate the possible combinations of reactants that were used to create the product. Clarity: The paper is well written and adequately covers the background for the task. More details could be added to the description of the model and tuning in the appendix. The authors could also include a figure showing the full architecture of Molecule-Chef since it has multiple parts. Significance: The paper has tried to tackle a useful problem of jointly solving molecule search and deciphering recipe tasks. It may be helpful for researchers developing methods to solve similar tasks.

The authors present an innovative model addressing two crucial problems in cheminformatics: the molecular search problem and the molecular recipe problem. The article is written a clear and understandable way, which leaves no loose ends by good argumentation and making nice use references. Furthermore, by providing the code makes the reviewing process easier. Based on this, the paper should be accepted as it is.

There are other previous methods that can generate valid molecules. For example, the relationship with the following paper is not clear. Hiroshi Kajino: "Molecular Hypergraph Grammar with Its Application to Molecular Optimization", ICML-19, Long Beach, CA, 2019. The authors in the above paper insist that most generated molecules are chemically valid. Did the authors compare with the method in the above paper? The detailed of the proposed method should be provided. I understand the page limitation. I would appreciate if the authors could provide the detailed explanations on the proposed method in the supplement. How did the authors use the method in the following paper? Marwin HS Segler, Mike Preuss, and Mark P Waller. Planning chemical syntheses with deep neural networks and symbolic AI. Nature, 555(7698):604, 2018. Data used in this paper were not provided, so the code did not work. There is no README file. Please provide data and the README file that describes how to run the program. Otherwise, it is impossible to evaluate the reproducibility of the proposed method. There are several typos. Line 98: we hope gives Line 128: has has Line 248: a target molecules