Generating readily synthesizable dye scaffolds with SyntheFluor

Ruhi Sayana, Jennifer Xu, Kate Callon, Jonathan Deutsch, James Zou, John Janetzko, Rabindra V. Shivnaraine, Kyle Swanson

Keywords: generative AI, reinforcement learning, fluorescence, molecular design, chemical space exploration, virtual screening, structure-property prediction

TL;DR: In this paper, we introduce SyntheFluor, a generative model that can design novel, synthesizable fluorescent compounds.

Abstract: Developing new fluorophores needed for advanced bioimaging techniques requires the exploration of previously unexplored chemical space. Generative AI approaches for the creation of novel dye scaffolds are promising in that they explore diverse regions of chemical space, but previous attempts have yielded synthetically intractable dye candidates due to the absence of reaction constraints, thus impeding experimental validation. Here, we present SyntheFluor, a generative AI model that employs known reaction libraries and molecular building blocks to create readily synthesizable fluorescent molecule scaffolds. SyntheFluor designed 11,590 molecules, which were filtered to a set of 19 diverse candidate molecules predicted to have dye-like properties. These 19 candidates were further examined by time-dependent density functional theory calculations, and 14 were successfully synthesized and 13 were experimentally validated. The photophysical properties of the three most fluorescent molecules were characterized in depth, and the top scaffold in particular showed robust fluorescence properties comparable to a known dye, demonstrating the utility of SyntheFluor.

Submission Number: 128