Research

The Howard Lab combines supramolecular synthesis, physical organic chemistry, and data science to probe molecular recognition. Selectivity in host-guest chemistry arises from a collection of noncovalent interactions that are poorly described by current models. We build quantitative models that relate molecular structure to the behavior of assemblies using computed molecular properties and machine learning. Our goal is to design chemosensors and stereoselective catalysts using machine learning techniques and to establish general principles for how confinement and noncovalent interactions control supramolecular function. Students in the Howard Lab will develop skills in organic synthesis, NMR spectroscopy, UV-Vis spectroscopy, density functional theory, molecular dynamics, and machine learning.

Machine Learning for Supramolecular Chemosensing

Fluorescent chemosensors built on indicator displacement rely on a dye that is quenched while bound to a host and recovers its fluorescence when a guest displaces it. Which host-dye pairs behave this way, and which guests bind tightly enough to displace the dye, are often determined by trial and error. We are building a library of quantum-chemical descriptors for commonly evaluated guests and dyes and pairing it with a set of pillararene hosts. These combined datasets allow us to construct models that predict binding affinity and the extent of fluorescence quenching from computed molecular properties. By using interpretable molecular features, the models reveal physical properties such as shape, charge distribution, and flexibility that actually control sensor response.

Data-driven Optimization of Supramolecular Catalysts

Chiral metal-organic cages (MOCs) catalyze reactions inside a pocket, and this environment can enforce selectivity that is difficult to achieve with conventional catalysts. Very few MOCs give high enantioselectivity in intermolecular reactions, and there is no general understanding of how cage structure controls enantioinduction. We use computed molecular features of the catalyst derived from molecular dynamics and semiempirical calculations to relate cage structure to stereoselectivity. We are applying this approach to both anionic and cationic MOCs for bimolecular reactions.

Machine Learning for Macrocyclic Peptide Receptors

Macrocyclic peptides bind ions, small molecules, and biological targets, and their binding behavior depends on which backbone conformations they adopt in solution. No current model connects the conformational preferences of a macrocycle to its binding affinity. We are training a neural network on computed conformer libraries to predict how macrocyclic peptide conformations are distributed. The pretrained model will then be fine-tuned to predict association constants measured via NMR spectroscopy. A working model would let us screen sequences computationally and synthesize only those predicted to bind well, targeting receptors for environmental pollutants and peptide motifs of biological interest.