Helix-MO: Sample-Efficient Molecular Optimization on Scene-Sensitive Latent Space

Efficient exploration of the chemical space to search the candidate drugs that satisfy various constraints is a fundamental task of drug discovery. Although many excellent deep molecular generative methods have been proposed to produce promising molecules, applying these methods in practice is still challenging since a great number of assessed molecules (samples) are required to provide the optimization direction, which is a considerable expense for drug discovery. To this end, we design a sample-efficient molecular generative method, namely Helix-MO, which can fast adapt to particular optimization scenes with only a small number of assessed samples. Helix-MO explores the chemical space in a scene-sensitive latent space, dynamically fine-tuned by multiple kinds of learning tasks from multiple perspectives. The learning tasks encourage the model to focus on modeling the more promising molecules during the optimization process to promote sample efficiency. Extensive experiments demonstrate that Helix-MO can achieve competitive performance with only a few assessed samples on four molecular optimization scenes. Ablation studies verify the impact of the learning tasks in the scene-specific latent space, efficiently identifying the critical characters of the satisfactory molecules. We also deployed Helix-MO on the website PaddleHelix (https://paddlehelix.baidu.com/app/drug/drugdesign/forecast) to provide drug design service and apply it to produce inhibitors of a kinase to demonstrate its practicability.