Addressing Challenges of Multi-Property Optimization and Synthesizability in Inverse Design

Inverse or generative models that can suggest novel molecules with optimized properties have promise to decrease materials discovery timelines for pharmaceutics, energetics, organic semiconductors, and polymers. However, the utility and promised time-savings of these models have yet to be fully realized due to challenges associated with co-optimizing multiple material properties, reducing model uncertainty when extrapolating beyond a known and trained chemical space to identify new molecules with extreme desired properties, and accounting for molecular synthesizability in a robust manner for suggested molecules. In our work, we have evaluated two state-of-the-art generative models – JTVAE and JANUS – for multi-property optimization of small molecules and have developed strategies for addressing the extrapolation and synthesizability challenges. We specifically focus on identifying novel and synthesizable molecules that have high density and high solid heat of formation – two important qualities for energetic molecules. Between the two models considered, we found that JANUS generally yielded more molecules with both higher density and heat of formation. To address the first challenge of accurately extrapolating to a chemical space beyond the known chemical space used for training, we employ two strategies. First, we created a high-throughput density functional theory pipeline capable of computing the density and heat of formation of ~1,000 molecules/day. We utilize the results from this pipeline to iteratively retrain our model, thereby increasing the data we have available in the chemical space we are extrapolating to and the accuracy of our predictions in this space. Second, we found that incorporating uncertainty quantification (where the uncertainty is the variance from an ensemble of models) in our property predictions also aided in filtering for suggested molecules that are more likely to have the predicted targeted properties. To address the second challenge of synthesizability, we found that the inclusion of a synthesizability score in our objective function is necessary for steering the generative model away from a completely unrealistic chemical space. However, the synthesizability score by itself does not directly correlate to true synthesizability. Thus, we utilize SciFinder’s retrosynthesis planning tool to specifically screen for likely synthesis routes of the top candidate molecules. Using these tools in aggregate, we are identifying novel molecules that are most likely to have the desired properties we aim for and also a means of synthesizing them. Now, synthesizing our best candidates and experimentally validating their properties is the final necessary step towards realizing generative models’ full utility.<br/> <br/>This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.