Use LSTM coupled with hill climbing algorithm to optimize a certain property in generated molecules.

This code is part of a benchmarking project for molecular discovery using generative models.

Train the model by running

python train.py

In the train.py file, modify the data_path, and string_type (selfies or smiles) used.

The data_path can be a csv or plain text file. The data will be read using utils.get_lists() function. Specify the

• data_path: path to the data file
• sep (default= ' '): delimiter for columns of data in file
• header (default=None): number of rows used as column names
• smiles_name (default=0): name of column for the molecule smiles (if header=None, use smiles_name=0)

Climb the LSTM using

python climb.py

In the climb.py file, modify the fitness_function, which takes in a smiles and returns a target value. Specify the data_path (see above), the model_path, where the trained LSTM is stored, the string_type (selfies or smiles), and the out_path, where the climbing results are stored.

You can change the sampling strategies by changing the keywords.

• num_generation: how many generations of climbing
• num_best: top k values used after sorting
• num_randomize: number of randomized SELFIES generated from top k smiles (SMILES LSTM cannot handle randomized smiles, so the smiles are just duplicated num_randomize times)
• samps_per_seed: number of times each of the randomized seeds are sampled
• num_seed_chars: number of characters in starting sequence used for sampling. If None, a random number of characters from 25% to 75% of the string length is used.
• temperature: controls randomness of sampling (larger is more random)
• retrain: boolean that specifies whether the model is retrained on the new sampled molecules per generation

Note that the total number of sampled molecules per generation is num_best * num_randomize * samps_per_seed.