Tweets

2023-07-08T03:38:24+00:00 | 🔗

Anyone know of anyone doing keypoint detection (with detectron2 perhaps) for optical structure recognition?

2023-07-08T01:08:27+00:00 | 🔗

@johnjnay @Stanford @UCBerkeley I mean that’s why we have introductions and conclusions right, or maybe that is “prompt engineering” now. Seems like a good heuristic

2023-07-08T00:04:22+00:00 | 🔗

The chat with PDF I want to see is to be able to be able to ask the Blue Book which IUPAC naming rules apply…

2023-07-07T17:49:12+00:00 | 🔗

@marwinsegler @j0hnparkhill @RGBLabMIT why couldn't the paper starting from an initial population of patented similar compounds get there with REINVENT? Though I noticed they aren't using the latest version of REINVENT though, maybe that + hyperparameters is it https://t.co/RyHWXQruwZ

2023-07-07T17:46:19+00:00 | 🔗

@marwinsegler @j0hnparkhill @RGBLabMIT I noodled on it a bit more, and was in fact able to do it! I think the issue was the population (with CrEM) had gotten to a weird place which GraphGA wasn't able to escape from (nor was able to reach), and which CrEM would have escaped from. But that also raises, ...

2023-07-07T17:23:39+00:00 | 🔗

A workflow to take a screenshot of an IUPAC, use optical character recognition to extract the string, GPT to clean it up, then OPSIN to change it to a SMILES (and then if erroring, feed the error back to GPT to clean up the IUPAC more). https://t.co/yYpjK6J1T1

2023-07-06T16:52:31+00:00 | 🔗

RT @doctorveera: A mind-blowing paper has come out today in @Nature In 2016, JC Venter Institute scientists trimmed a bacterial genome t…

2023-07-06T16:04:03+00:00 | 🔗

@j0hnparkhill @RGBLabMIT think the failure mode is interesting.

2023-07-06T16:03:45+00:00 | 🔗

@j0hnparkhill @RGBLabMIT Note they had already tried REINVENT too (have not replicated). You’re quite right in that it seems like being able to exactly rediscover is the “bare minimum”. Now perhaps in a “real” setting the metrics are more continuous and “easy” relative to tanimoto similarity, but I…

2023-07-06T16:01:20+00:00 | 🔗

@j0hnparkhill @RGBLabMIT That’s what I thought too, and that’s why I was surprised that they had used it in the paper. But then I decided to put my time where my mind was and try some fairy reasonable methods slicing, GB-GA, and CrEM and was surprised that they couldn’t “get there”.