It seems to me a fundamental problem with this paper is that they're training an autoencoder on a relatively small set of drugs that have been tested on cancer. They are then trying to approximately index a much larger dataset of chemicals and claiming these are novel possible cancer treatments because some of them have been previously considered as cancer drugs.
It's not clear that this is doing much more than finding drugs similar to the training drugs in the new dataset. Given that a large part of pharmaceutical/chemical development is based on slightly modifying existing compounds, it is not obvious that this adds much to the discovery pipeline, or that it is robust 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.
Be careful with this one, the premise is sound (coming from a DL background), but I skimmed the paper and didn't see them comparing it to existing simple compounds and don't show if their compounds are actually any better.
Expensive or not, they didn't do it, so we don't know if those compounds actually work.
Am I missing something or did they also not use a held-out validation dataset to assess performance? It seems to me they ran their autoencoder, got some suggested compounds, listed anecdotal evidence about those compounds, and called it a day.
It's not clear that this is doing much more than finding drugs similar to the training drugs in the new dataset. Given that a large part of pharmaceutical/chemical development is based on slightly modifying existing compounds, it is not obvious that this adds much to the discovery pipeline, or that it is robust 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.