(364f) Design of Optoelectronic Self-Assembling ?-Conjugated Peptides Via Combined Experimental-Computational Multi-Fidelity Active Learning

In this work we integrate experimental and computational data streams using multi-fidelity Bayesian optimization and deep representational learning within an active learning platform to accelerate the design and discovery of π-conjugated peptides capable of self-assembling into nanowire-like structures with emergent optoelectronic properties. π-conjugated peptides are tri-block molecules composed of a central π-core flanked by oligopeptide wings demonstrated to self-assemble into pseudo-1D nanoaggregates resembling biocompatible nanowires with potential applications in engineered linkages with living cells for repairing neuronal damage and interfacing with photosynthetic machinery for energy harvesting and storage. The amino acid sequence comprising the oligopeptide wing influences the self-assembly and resultant nanostructure geometry, however exhaustive traversal of possible oligopeptide wings is prohibitively expensive via either simulation or experiment alone due to the combinatorial explosion of the molecular design space. Using graph neural networks and deep representational learning we create a continuous, low-dimensional representational of our discrete, high-dimensional molecular design space that we expose to a multi-fidelity Bayesian optimization platform fusing inexpensive, high-volume, and approximate simulation data with expensive, low-volume, and accurate experimental data to rationally navigate and identify molecules with engineered optoelectronic properties. This active learning search identifies new molecules that are experimentally synthesized and tested to possess superior optoelectronic properties compared to previous works, while simultaneously revealing design rules that govern the self-assembly of π-conjugated peptide systems.