Abstract
De novo, in-silico design of molecules is a challenging problem with applications in drug discovery and material design.
Here, we introduce a masked graph model which learns a distribution over graphs by capturing all possible conditional distributions over unobserved nodes and edges given observed ones. We train our masked graph model on existing molecular graphs and then sample novel molecular graphs from it by iteratively masking and replacing different parts of initialized graphs. We evaluate our approach on the QM9 and ChEMBL datasets using the distribution-learning benchmark from the GuacaMol framework. The benchmark contains five metrics: the validity, uniqueness, novelty, KL-divergence and Fr{\'e}chet ChemNet Distance scores, the last two of which are measures of the similarity of the generated samples to the training, validation and test distributions. We find that KL-divergence and Fréchet ChemNet Distance scores are anti-correlated with novelty scores. By varying generation initialization and the fraction of the graph masked and replaced at each generation step, we can increase the Fréchet score at the cost of novelty.
In this way, we show that our model offers transparent and tunable control of the trade-off between these metrics, a key point of control in design applications currently lacking in other approaches to molecular graph generation. Our model outperforms previously proposed graph-based approaches and is competitive with SMILES-based approaches. Finally, we observe that minimizing validation loss on the training task is a suitable proxy for improving generation quality, which shows the suitability of optimizing the training objective for improving generation.
Here, we introduce a masked graph model which learns a distribution over graphs by capturing all possible conditional distributions over unobserved nodes and edges given observed ones. We train our masked graph model on existing molecular graphs and then sample novel molecular graphs from it by iteratively masking and replacing different parts of initialized graphs. We evaluate our approach on the QM9 and ChEMBL datasets using the distribution-learning benchmark from the GuacaMol framework. The benchmark contains five metrics: the validity, uniqueness, novelty, KL-divergence and Fr{\'e}chet ChemNet Distance scores, the last two of which are measures of the similarity of the generated samples to the training, validation and test distributions. We find that KL-divergence and Fréchet ChemNet Distance scores are anti-correlated with novelty scores. By varying generation initialization and the fraction of the graph masked and replaced at each generation step, we can increase the Fréchet score at the cost of novelty.
In this way, we show that our model offers transparent and tunable control of the trade-off between these metrics, a key point of control in design applications currently lacking in other approaches to molecular graph generation. Our model outperforms previously proposed graph-based approaches and is competitive with SMILES-based approaches. Finally, we observe that minimizing validation loss on the training task is a suitable proxy for improving generation quality, which shows the suitability of optimizing the training objective for improving generation.