DeepChem Minutes 3/18/2021

DeepChem Minutes

India/Asia/Pacific Call

Date: March 18th, 2021
Attendees: Bharath, Alana, Vignesh V., Vignesh S., Aditya, Rajath
Summary: Aditya is a final year undergrad student from BITS Pilani. Aditya is interested in learning more about deep learning and wanted to learn more.

Vignesh S. is building a startup based on deep learning for drug discovery.

Rajath is a senior at Harvard currently on a leave of absence working at Postera.

Alana is working on getting the RNN PR working. It’s been more challenging than expected, with an unexpected type mismatch slowing it down.

Aditya asked about pretraining as a GSoC project. Bharath said it might be too tricky for a summer project.

Vignesh V. asked about the Megnet model and whether we could wrap a long running test. Bharath said tests should run within a minute.

Rajath asked about adding in support for physics based featurizations. Bharath said this would be a good idea but might be good for us to lean on an underlying library.

Americas/Europe/Africa/Middle East

India/Asia/Pacific Call

Date: March 19th, 2021
Attendees: Bharath, Stanley, Seyone, Nathan, Peter, Vignesh, Daniel
Summary: Bharath gave the same update as at the last meeting.

Peter has been working on updating the book examples to make them work correctly with 2.5.0. There are still a couple of book examples which are behaving poorly.

Vignesh has been working on a wrapper for megnet.

Seyone this week has been working on MoleculeNet benchmarking for ChemBERTa. Elayne from the ChemBERTa project has also got multitask regression working. We should be close to starting to focus on porting all the code to DeepChem now that we’ve finished building the benchmarking + pre-training MLM and regression scripts. We just need to copy over the Roberta Sequence classification pytorch implementation and we should be good to go!

Stanley has been starting to work on using Ray Tune with his company infrastructure. He’s in particular also started experimenting with Hyperband as a next step past Bayesian hyperparameter tuning. This weekend, Stanley will be working on a notebook for hyperparameter tuning.

Nathan has been working on adding a benchmark for atomic convs to MoleculeNet. There’s a bit of snag since it runs fine locally, but the featurization fails on colab for some reason. It could be something to do with running a script within Colab.

Daniel had to focus on his research for a bit but can now take up looking at the low data models again.

Joining the DeepChem Developer Calls

As a quick reminder to anyone reading along, the DeepChem developer calls are open to the public! If you’re interested in attending either or both of the calls, please send an email to, where X=bharath, Y=ramsundar.

1 Like