A series of baseline model implementations for the guacamol benchmark for generative chemistry.
A more in depth explanation of the benchmarks and scores for these baselines is can be found in our paper.

To install all dependencies:

conda install rdkit -c rdkit
pip install -r requirements.txt

Some baselines require the guacamol dataset to run, to get it run:

bash fetch_guacamol_dataset.sh

Dummy baseline, always returning random molecules form the guacamol training set.

To execute the goal-directed generation benchmarks:

python -m random_smiles_sampler.goal_directed_generation

To execute the distribution learning benchmarks:

python -m random_smiles_sampler.distribution_learning

Dummy baseline that simply returns the molecules from the guacamol training set that best satisfy the score of a goal-directed benchmark.
There is no model nor training, its only purpose is to establish a lower bound on the benchmark scores.

To execute the goal-directed generation benchmarks:

python -m best_from_chembl.goal_directed_generation

No distribution learning benchmark available.

Genetic algorithm on SMILES as described in: https://www.journal.csj.jp/doi/10.1246/cl.180665

Implementation adapted from: https://github.com/tsudalab/ChemGE

To execute the goal-directed generation benchmarks:

python -m smiles_ga.goal_directed_generation

No distribution learning benchmark available.

Genetic algoritm on molecule graphs as described in: https://doi.org/10.26434/chemrxiv.7240751

Implementation adapted from: https://github.com/jensengroup/GB-GA

To execute the goal-directed generation benchmarks:

python -m graph_ga.goal_directed_generation

No distribution learning benchmark available.

Monte Carlo Tree Search on molecule graphs as described in: https://doi.org/10.26434/chemrxiv.7240751

Implementation adapted from: https://github.com/jensengroup/GB-GB

To execute the goal-directed generation benchmarks:

python -m graph_mcts.goal_directed_generation

To execute the distribution learning benchmarks:

python -m graph_mcts.distribution_learning

To re-generate the distribution statistics as pickle files:

python -m graph_mcts.analyze_dataset

Long-short term memory on SMILES as described in: https://arxiv.org/abs/1701.01329

This implementation optimizes using hill climbing algorithm.

Implementation by BenevolentAI

A pre-trained model is provided in: [smiles_lstm/pretrained_model](TODO real URL)

To execute the goal-directed generation benchmarks:

python -m smiles_lstm_hc.goal_directed_generation

To execute the distribution learning benchmark:

python -m smiles_lstm_hc.distribution_learning

To train a model from scratch:

python -m smiles_lstm_hc.train_smiles_lstm_model

Long-short term memory on SMILES as described in: https://arxiv.org/abs/1701.01329

This implementation optimizes using proximal policy optimization algorithm.

Implementation by BenevolentAI

A pre-trained model is provided in: [smiles_lstm/pretrained_model](TODO real URL)

To execute the goal-directed generation benchmarks:

python -m smiles_lstm_ppo.goal_directed_generation