Molecular Generative Models

Among the wide range of available models for molecular design, evolutionary algorithms stand out because of their simplicity and their ease of use without requiring a significant amount of initial data. Accordingly, they are standard baseline models for inverse molecular design models, in particular for goal-directed optimization tasks. They can be the stepping stone for more sophisticated approaches such as reinforcement learning, which has provided ground-breaking advances in the field of artificial intelligence in recent years when fed with very large amounts of data. We seek to combine the best of both these worlds by implementing inverse molecular design algorithms that do not require any initial data to be started, however, at the same time, continuously learn from previous results to reduce the number of property evaluations significantly. This is important for real world applications in the laboratory as going beyond hundreds of experimental evaluations for catalytic activity tends to prove highly prohibitive.