ROGI-XD

Evaluating the roughness of learned representations from chemical foundation models

Table of Contents

• Setup
• Pretrained models
• Results
• Figures

Setup

0. (if necessary) install conda

1. clone this repository and its submodules:

   git clone --recurse-submodules THIS_REPO
   

2. Choose one option below to install all required packages

   a. Conda (recommended)

   conda env create -f environment.yml
   

   Note: environment.yml contains CUDA 11.6 versions of several packages. If you want to use this environment on CPU, replace cu116 with cpu before running the above command

   sed -i -e 's/cu116/cpu/g' environment.yml
   conda env create -f environment.yml
   

   b. Manual

   Note: this is only recommended in the event that you'd like specify different package versions

   conda create -n rogi_xd -y python=3.9 && conda activate rogi_xd
   CUDA=cu116  # NOTE: depending on your machine, this can be any one of: ("cpu" | "cu116" | "cu117") 
   pip install torch==1.13 --extra-index-url https://download.pytorch.org/whl/${CUDA}\
   && pip install pyg-lib torch-scatter torch-sparse \
       torch-cluster torch-spline-conv torch-geometric \
       -f https://data.pyg.org/whl/torch-1.13.1+${CUDA}.html \
   && pip install git+https://github.com/DeepGraphLearning/torchdrug \
   && pip install -r requirements.txt
   

3. install the ae_utils and rogi_xd pacakges:

   pip install autoencoders/
   pip install -e . --no-deps
   

Pretrained models

• the original pretrained models are stored in the following directories:

  • GIN: models/gin/zinc
  • VAE: models/vae/zinc

• both the VAE and GIN were pretrained over 100 epochs on the ZINC 250k dataset using a learning rate of 3e-4 and early stopping on the validation loss

• to train your own models, run one of the following commands:

  rogi_xd train -m (gin | vae) -d zinc -c 8

  The models will be saved to the following directory models/{gin,vae}/zinc, which can be supplied to the rogi command later via the --model-dir argument.

  NOTE: this script trains a simple GIN or VAE and doesn't allow for custom architectures to specified. That's because the goal of this repository was not to provide another VAE implementation. If you wish to reuse the VAE object model, then you'll want to head to the autoencoders submodule

Results

All results can be generated via the following command: make all

ROGI data

Use the rogi_xd rogi command line entry point to run your desired calculations.

usage: rogi_xd rogi [-h] [--logfile [LOGFILE]] [-v] (-i INPUT | -d DATASETS_TASKS [DATASETS_TASKS ...]) [-f {descriptor,morgan,chemberta,chemgpt,gin,vae,random}] [-r REPEATS] [-N N] [-o OUTPUT] [-b BATCH_SIZE]
                    [-m MODEL_DIR] [-c NUM_WORKERS] [--coarse-grain] [-k [NUM_FOLDS]] [--reinit] [--orig] [-l [LENGTH]]

optional arguments:
  -h, --help            show this help message and exit
  --logfile [LOGFILE], --log [LOGFILE]
                        the path to which the log file should be written. Not specifying will this log to stdout. Adding just the flag ('--log/--logfile') will automatically log to a file at 'logs/YYYY-MM-
                        DDTHH:MM:SS.log'
  -v, --verbose         the verbosity level
  -i INPUT, --input INPUT
                        A plaintext file containing a dataset/task entry on each line. Mutually exclusive with the '--datasets-tasks' argument
  -d DATASETS_TASKS [DATASETS_TASKS ...], --datasets-tasks DATASETS_TASKS [DATASETS_TASKS ...], --dt DATASETS_TASKS [DATASETS_TASKS ...], --datasets DATASETS_TASKS [DATASETS_TASKS ...]
  -f {descriptor,morgan,chemberta,chemgpt,gin,vae,random}, --featurizer {descriptor,morgan,chemberta,chemgpt,gin,vae,random}
  -r REPEATS, --repeats REPEATS
  -N N                  the number of data to subsample
  -o OUTPUT, --output OUTPUT
                        the to which results should be written. If unspecified, will write to 'results/raw/rogi/FEATURIZER.{csv,json}', depending on the output data ('.json' if '--cg' is present, '.csv' otherwise)
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        the batch size to use in the featurizer. If unspecified, the featurizer will select its own batch size
  -m MODEL_DIR, --model-dir MODEL_DIR
                        the directory of a saved model for VAE or GIN featurizers
  -c NUM_WORKERS, --num-workers NUM_WORKERS
                        the number of CPUs to parallelize data loading over, if possible
  --coarse-grain, --cg  whether to store the raw coarse-graining results.
  -k [NUM_FOLDS], --num-folds [NUM_FOLDS], --cv [NUM_FOLDS]
                        the number of folds to use in cross-validation. If this flag is present, then this script will run in cross-validation mode, otherwise it will just perform ROGI calculation. Adding only the flag
                        (i.e., just '-k') corresponds to a default of 5 folds, but a specific number may be specified
  --reinit              randomize the weights of a pretrained model before using it
  --orig                whether to use the original ROGI formulation (i.e., distance threshold as the x-axis). By default, uses the ROGI-XD formulation (i.e., 1 - log N_clusters / log N as the x-axis)
  -l [LENGTH], --length [LENGTH]
                        the length of a random representation

Cross-validation and coarse-graining results

Use the same entrypoint as before with the addition of the --cv and --cg flags, like so: rogi_xd rogi --cv --cg

Note: The scripts rely datasets from both TDC [1] and GuacaMol oracle functions evaluated for random molecules sampled from ZINC250k [2]. The script will first search for the corresponding dataset in the $ROGIXD_CACHE directory (where ROGIXD_CACHE is an environment variable) and if it doesn't find them, will then download them to that directory. If this environment variable is not set, the scripts will use $HOME/.cache/rogi_xd instead.

Finetuning

To finetune a VAE model, run the following command:

rogi_xd finetune -m models/vae/zinc/ -b 64 -d DATASET_TASK -f FRACTION

For our experiments, we used -d LIPOPHILICITY_ASTRAZENECA and -f 0.8

Figures

See the corresponding notebook:

• correlation.ipynb: correlation plots and $r distribution plots
• rogi-dist.ipynb: ROGI distribution boxplots and parity plots
• toy_surfaces.ipynb: toy example figures

Note: for these notebooks to work out of the box, data should be generated using the make all command from above`