Implicit Chemical Property Optimization with Graph Energy Based Models

In the generative realm of Machine Learning, optimizing chemical properties is a challenging task due to the inherent complexity and high dimensionality of chemical space. Drug discovery often requires the simultaneous optimization of multiple properties, a task where traditional methods frequently fall short. In this work, we propose a novel unsupervised energy-based model for chemical property optimization, capable of optimizing both single and multiple properties by leveraging the compositionality of energy functions. The framework utilizes Graph Neural Networks and is trained in an unsupervised manner, without the need for explicit property labels. We validated our approach on the optimization of LogP and QED properties, demonstrating that our model can generate molecules that optimize the desired properties without explicit guidance, outperforming state-of-the-art methods. Additionally, we show that our model can generate molecules satisfying multiple properties simultaneously, highlighting its effectiveness in zero-shot multi-property optimization. Our results suggest that the proposed framework offers a promising new approach to compositional concept learning and chemical property optimization in the context of de novo drug design.