GMD-MO-LSO

Codes for Multi-Objective Latent Space Optimization of Generative Molecular Design Models will be uploaded here.

Description

This code repository contains scripts to run the multi-objective weighted retraining on molecular generative model, i.e. JT-VAE.

We followed the codebase from Sample-Efficient Optimization in the Latent Space of Deep Generative Models via Weighted Retraining. For the necessary dependencies, please follow their instructions.

Running multi-objective weighted retraining

The bash script to run the multi-objective weighted retraining can be found in scripts/opt/mo_opt_chem.sh. Currently it has the commands for all six possible pairs of logP, SAS, NP_score and DRD2. One pair can be run at a time. For this reason other 5 commands are commented out.

The important arguments for mo_opt_chem.sh:

• --pretrained_model_file is the path to baseline model. For the case of complete dataset, it is same for all pairs. For reduced dataset, the models for each pair can be found in assets/pretrained_model
• --train_path points to the directory where the tensor data for training molecules are stored. Note that, for the case of reduced dataset, this directory must correpond to the appropriate property pairs.
• --all_new is set 1 if we want to put all the new molecules in the next stage of weighted retraining. If it is 0, then all new molecules are added to the current training dataset from which only 10% random samples are used in weighted retraining stage.

• --rank_weight_k controls the degree of weighting in formulation of weight from Pareto front rank

• --retraining_frequency controls the retraining frequencies (number of epochs between retrainings)
• --result_root is the directory to save the results in