Automatic chemical design using a data-driven continuous representation of molecules

saw the training set had Br (bromine) and Ca (Calcium), and the output molecules sometimes were Ba (barium) even though that's not in the original dataset at all. There are other reasons why the tokenizer produces bad results- only about 1-10% of vectors decode to valid molecules. Invalid SMILES are mostly useless- they don't correspond to actual structures. To respond to this, the SELFIES format makes a few changes so that it is effectively impossible to produce invalid SELFIES stringes when decoding a VAE. Among other things, tokenization matches the actual elements and so the model will only ever output valid elements. I believe this is the SMILES paper that my own experiments were based on: https://arxiv.org/pdf/1610.02415 (see https://github.com/maxhodak/keras-molecules for an open source attempt at implementation) And this is the paper introducing SELFIES: https://arxiv.org/abs/1905.13741 (open source packages for working with SELFIES, and some example training scripts https://github.com/aspuru-guzik-group/selfies see "Validity of Latent Space in VAE SMILES vs. SELFIES for more detail on the robustness). BTW, as a side note: even though we put a bunch of effort into duplicating the original SMILES VAE, it was extremely slow to train and not very useful. Now you can just ask Gemini to write a full SELFIES VAE and train it in less than a day on a conventional GPU (thanks pytorch transformers!) to get a decent basic set of embeddings useful for exploring chemical space.

interest here is the quite effective smooth(ish) movement through structure-property space using VAEs.[2] [0] https://zinc.docking.org/ [1] "Virtual discovery of melatonin receptor ligands to modulate circadian rhythms" https://www.nature.com/articles/s41586-020-2027-0.pdf [2] "Automatic Chemical Design Using a Data-DrivenContinuous Representation of Molecules", https://arxiv.org/pdf/1610.02415.pdf

to false positives (drugs that have similar "fingerprints" to those seen in training set but are part of a large class of chemicals that mostly don't fight cancer and are therefore unlikely to appear in the small training set of possible cancer-fighting compounds). Fundamentally machine learning works better with large data. It is hard to believe that training on 6000 chemicals, given the overall diversity in chemical space (~72 million chemicals in the dataset they're indexing) is likely to lead to a real "understanding" of what constitutes a cancer-fighting drug as opposed to parroting existing drugs they're trained on. The effort from Ryan P. Adams' group (https://arxiv.org/pdf/1610.02415v2.pdf) is I think more reasonable because it trains on larger sets of chemicals, and is also truly generative in the sense of being able to create new chemicals as opposed to signatures "similar to" existing ones. Though I should note that they also had trouble generating plausible chemical structures despite a larger training set.

course, our paper is just a proof of concept. For instance, instead of encoding to and from SMILES, it would be much better to encode to and from graphs directly. We know how to encode graphs into vectors, but I don’t know of a good way to decode a vector into a graph. Another open problem is that it’s hard to know what to optimize. Our initial experiments optimizing for specific chemical properties produced suggested molecules with crazy structures, such as giant rings. Human chemists have a great intuition for what is easily synthesizable or stable, and it’s hard to articulate all the properties we want the molecule to have programmatically. Alternatively, we could enforce the optimizer to only look at molecules similar to ones we’ve already seen, but this is unsatisfying too - after all, the best result of exploration is when you find something unlike what you’ve seen before.