DeepChem Minutes 2/18/2021

DeepChem Minutes

India/Asia/Pacific Call

Date: February 18th, 2021
Attendees: Bharath, Alana, Yuanqi, Vignesh
Summary: Bharath has posted a collection of GSoC Ideas on the open chemistry wiki for students to look at.

Alana has been making progress with the RNN, has got the environment locally set up and has the unit test running, and is working on debugging the RNN class. There’s an error with the input shapes which is a little tricky.

Vignesh has been working on the PAGTNModel. He’s working to implement it in DGL-Life Sciences and is going to write a wrapper function in DeepChem. Vignesh is coordinating with Mufei to do this work. Vignesh asked if this would involve tape.

Yuanqi is continuing to work on MoleculeNet benchmarks. He submitted two PRs last week and is working on more datasets. He found some minor errors in the dataset loader for bandgap PR

Americas/Europe/Africa/Middle East

India/Asia/Pacific Call

Date: February 5th, 2021
Attendees: Bharath, Peter, Stanley, Maria, Seyone, Paul
Summary: Paul is working at Nurix where he’s using DeepChem with Bharath

Bharath gave the same update as he did last week.

Peter is working on test cases, and is trying to fix failures. One issue is the flaky tests failing is a much more serious issue with github actions than with travis. The second issue is that it’s been a while since we ran our slow tests and a bunch of them are failing. It also turns out that TF 2.4 breaks things. Invoking a keras layer no longer returns a tensor and instead returns a keras tensor.

Stanley was out last week moving so was a bit too busy to work on DeepChem work. He hopes that he will have the tutorial for hyperparameter tuning out in a couple of weeks. Stanley also mentioned he’s worked a bit with keras tensor and might be able to take a look at some of the test failures.

Maria has continued working on PR for fragment handling in graph models. Maria’s new architecture uses transformers to perform the necessary transformations, which seems to be a better design than the previous methods that were considered.

Seyone was primarily busy with the ChemBERTa presentation with CMU’s SciML series. He was working on benchmarking with the new smiles tokenizer which boosts numbers across the board.

Nathan fixed an issue with pymatgen fixing the build and is now working on the atomic conv models. This probably won’t be ready by 2.5 but 2.6.

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