Tweets
2022-12-14T20:03:07+00:00 | 🔗
A little help is needed, trying to implement a local version of PubChem 881-bit fingerprints in Python/RDKit. On a validation set of 400 orange book structures, can get 331 perfectly right and 360 within 1 bit as compared to PubChem servers. One example of error.. ideas? tytyty https://t.co/DSkSL0fl1B
2022-12-12T21:25:02+00:00 | 🔗
@rapodaca I'm not quite sure about that. When I gave it aspirin it could nicely generate alternate SMILES, but once I gave it something a little different... it couldn't perform the same task. So probably matching strings at least in this case. https://t.co/C1pJ1pESH0
2022-12-06T04:38:42+00:00 | 🔗
Wow this was from so long ago… I would argue that all these negative examples of chatGPT are literally hypnosis! Which makes me think that perhaps this is fairly fundamental to language intelligence, and actually makes me believe in these models MORE, and that idk problem hard https://t.co/2ENLbiQJ5q
2022-12-05T20:53:33+00:00 | 🔗
@gklambauer Interesting… maybe it’s the large dimensions + thousands of units, we’d always gone smaller though maybe that’s a product of ensembling + dataset size. Agreed on better than GNN/MPNN, we’ve implemented a lot but never anything *consistently* better, although often marginally so
2022-12-05T20:32:09+00:00 | 🔗
@gklambauer Do you have favorite MLP layer sizes/activation functions/regularizations? We could never get MLPs to the level of RFs and other ML techniques.
2022-11-30T01:30:20+00:00 | 🔗
RT @ashkanstweets: Doug Leone from Sequoia closed a panel yesterday evening I was in the audience for with a fitting quote for founders giv…
2022-11-28T14:34:25+00:00 | 🔗
@rguha @OlorenAI @OlorenChemEng Think you land on something stricter
2022-11-28T14:34:12+00:00 | 🔗
@rguha @OlorenAI @OlorenChemEng I think the hard part though is making sure that the model path served is able to be loaded in. If there is a single point of access and a single versioning for the database it’s likely not a problem. But then once you get multiple points of access/versions changing over time I
2022-11-28T14:31:32+00:00 | 🔗
@rguha @OlorenAI @OlorenChemEng Not off-hand, but your numbers look extremely doable for the system. I would love to hear more about it, but roughly It seems like what you want might even be simpler than that, e.g. maybe you really just want a database with model path + metadata, with a nicely exposed API