NeurIPS 2020

Guiding Deep Molecular Optimization with Genetic Exploration


Meta Review

Thank you for your submission. The paper was borderline. R2 main issue was that molecules are not looking realistic. This ties back to observations in https://arxiv.org/pdf/2002.07007.pdf. I fully agree with R2 that we need better models for generating ‘naturally-looking’ molecules. However, I also believe we need better benchmarks for de novo models, and it is not fair to demand from every paper in the literature to invent its own way of evaluating models against this (very important!) objective. The authors also empirically demonstrated the approach can be steered towards more realistic molecules by a proper regularization. To address R2 remark better, I propose that the authors also evaluate how easy it is to synthesize proposed molecules (there are available public APIs, and you could also use ASKCOS). I think that would be much better than the proposed somewhat ad-hoc filtering procedure. All reviewers found the evaluation thorough, and on this grounds I am happy to accept the paper. It presents a simple idea that I believe has chances to nudge the community towards better exploration methods in de novo generation. One issue that came up is that the solution is a bit poorly motivated because genetic operators are hand-crafted. A more perhaps natural rules could be based on feasible chemical reactions as in Molecule Chef. Please include a discussion of such methods in the related work section. Please also remember to address all remarks by the reviewers.