Search for a command to run...
The new wave of successful generative models in machine learning has\nincreased the interest in deep learning driven de novo drug design. However,\nassessing the performance of such generative models is notoriously difficult.\nMetrics that are typically used to assess the performance of such generative\nmodels are the percentage of chemically valid molecules or the similarity to\nreal molecules in terms of particular descriptors, such as the partition\ncoefficient (logP) or druglikeness. However, method comparison is difficult\nbecause of the inconsistent use of evaluation metrics, the necessity for\nmultiple metrics, and the fact that some of these measures can easily be\ntricked by simple rule-based systems. We propose a novel distance measure\nbetween two sets of molecules, called Fr\\'echet ChemNet distance (FCD), that\ncan be used as an evaluation metric for generative models. The FCD is similar\nto a recently established performance metric for comparing image generation\nmethods, the Fr\\'echet Inception Distance (FID). Whereas the FID uses one of\nthe hidden layers of InceptionNet, the FCD utilizes the penultimate layer of a\ndeep neural network called ChemNet, which was trained to predict drug\nactivities. Thus, the FCD metric takes into account chemically and biologically\nrelevant information about molecules, and also measures the diversity of the\nset via the distribution of generated molecules. The FCD's advantage over\nprevious metrics is that it can detect if generated molecules are a) diverse\nand have similar b) chemical and c) biological properties as real molecules. We\nfurther provide an easy-to-use implementation that only requires the SMILES\nrepresentation of the generated molecules as input to calculate the FCD.\nImplementations are available at: https://www.github.com/bioinf-jku/FCD\n