(577a) Iterative Peptide Biomaterials Discovery with Maximum Entropy Methods and Deep Learning

The use of artificial intelligence for molecular design has been rapidly advancing. New generative models propose novel molecular structures to test and deep learning can predict their molecular properties with high accuracy. The goal for these algorithms is optimizing a property by training against a large database of previously measured molecules. A major limitation is that these methods are not designed to be used concurrently with experiments and improving their accuracy as experimental data is gathered is non-trivial. Our research is focused on iterative materials design where experiments are done concurrently instead of post-hoc in molecular design. This is accomplished by using maximum entropy biasing methods that can update predictive physics-based models with new experimental results. This improves predictive accuracy and ensures interpretable models. Another challenge is the scarcity of data in materials design which we’re addressing with meta-learning. Meta-learning translates experience from past related systems to new ones, minimizing the number of molecules necessary to train models. I will present results of these methods applied to peptides where we have studied a variety of tasks including antimicrobial, antifouling, and solubility predictions. Peptides are a model system because there is no ambiguity in encoding and generating them because they are linear biopolymers made of 20 amino acid monomers.