Designing nanopatterned surfaces using multi-modal ligand self-assembly
Tunable, nanoscale chemical patterns can be formed by interactions between neighboring ligands immobilized on a flat surface. Through developing an understanding of how ligand structure can be used to modulate the self-assembly process, we can design new ligands capable of assembling into a wide variety of complex patterns.
Applying graph neural networks to predict peptide properties
While synthetic peptides can solve problems ranging from treating antibiotic-resistant diseases to catalyzing new reactions, resource and time costs make experimentally exploring their massive design space intractable. We use a hierarchical graph approach to representing peptides in neural networks to predict properties including anti-microbial activity and chromatographic retention time.
Developing Deep Neural Networks to Predict Surface Hydrophobicity
Molecular-scale hydrophobicity plays a key role in a wide variety of problems from protein-protein interactions to colloidal properties. We use a combination of molecular environment-based neural networks and high-throughput simulation methods to develop rapid, high-accuracy approaches for predicting surface hydrophobicity.
Sidestepping the Accuracy-Cost Tradeoff using Multi-Accuracy Active Learning
While molecular simulations are versatile tools that can be applied to quantify many properties, there exists a tradeoff between simulation cost and accuracy. Through developing and applying deep learning techniques such as transfer learning, uncertainty quantification, and active learning, we develop methods for combining simulation data from multiple accuracy levels to enable high quality predictions with reduced computational expense.