Chemprop Files

Cuik-molmaker

Chemprop can now use cuik-molmaker (https://github.com/NVIDIA-Digital-Bio/cuik-molmaker), a C++/python package that accelerates atom and bond featurization. Usage of cuik-molmaker accelerates Chemprop training by 1.6X and inference by 2.4X. In addition, memory usage is reduced by ~80% enables larger-scale training and inference workloads. It can be used as a drop in replacement for the featurization classes implemented in Chemprop. Use chemprop/scripts/check_and_install_cuik_molmaker.py to install the correct version for your environment. Then specify --use-cuikmolmaker-featurization if using the command line. Or if you are using chemprop in a python script, import data.LazyMoleculeDatapoint, data.CuikmolmakerDataset, and featurizers.CuikmolmakerMolGraphFeaturizer.

Other notable changes

We continue to make the command line interface easier to use. The train command, chemprop train, no longer requires a test set. Additionally, datapoint descriptors (e.g., temperature, pressure) can now be included in the main input file using a command similar to chemprop train --data-path input.csv --descriptors-columns temperature pressure.

There are also several bug fixes in this release. See the detailed PR list below.

What's Changed

• Zero-indexing refactoring of splits.json used in both tutorial & test by @jxl26 in https://github.com/chemprop/chemprop/pull/1244
• Remove unneeded folders from python distrubution by @JacksonBurns in https://github.com/chemprop/chemprop/pull/1240
• Standardizing the name Chemprop by @akshatzalte in https://github.com/chemprop/chemprop/pull/1259
• Update suggested CheMeleon citation by @JacksonBurns in https://github.com/chemprop/chemprop/pull/1267
• Enable cuik-molmaker for accelerated molecule featurization by @sveccham in https://github.com/chemprop/chemprop/pull/1253
• Fix randomly failing overfit test by @jxl26 in https://github.com/chemprop/chemprop/pull/1271
• Bug Fix: Add handling for --from-foundation argument in hpopt by @JacksonBurns in https://github.com/chemprop/chemprop/pull/1273
• Add support for empty test set by @jxl26 in https://github.com/chemprop/chemprop/pull/1243
• Ensure models loaded for transfer learning are on CPU by @JacksonBurns in https://github.com/chemprop/chemprop/pull/1252
• Fixes multiple quantile regression bugs by @craabreu in https://github.com/chemprop/chemprop/pull/1229
• Support extra feature columns in input datasets by @jxl26 in https://github.com/chemprop/chemprop/pull/1250
• Fix bug for molecule + reaction dataset with --molecule-featurizers (introduced in [#1253]) by @KnathanM in https://github.com/chemprop/chemprop/pull/1274
• Make hpopt use --tracking-metric by @JacksonBurns in https://github.com/chemprop/chemprop/pull/1281
• Clean up code and fix bugs for cuik-molmaker (in [#1253]) by @KnathanM in https://github.com/chemprop/chemprop/pull/1275

New Contributors

• @jxl26 made their first contribution in https://github.com/chemprop/chemprop/pull/1244
• @sveccham made their first contribution in https://github.com/chemprop/chemprop/pull/1253

Full Changelog: https://github.com/chemprop/chemprop/compare/v2.2.0...v2.2.1