Machine-learning Assisted Discovery of Absorbing Molecules for Tissue Clearing

Optical clearing enables deep imaging in biological tissues, but performance depends critically on dye absorption spectra. In practice, dyes with strong, narrow absorption bands improve transparency and contrast by selectively shaping residual optical attenuation in cleared tissues, as seen in our prior work. Acquiring full molar extinction coefficient (MEC) spectra for each candidate dye is slow and limits screening throughput. We present a data-driven framework that predicts a molecule's complete absorption spectrum from structure to accelerate dye selection for tissue clearing. Starting from each molecule's SMILES identifier, we generate cheminformatic fingerprints and train a supervised regressor to map fingerprints to MEC values across wavelengths; the trained model then predicts full spectra for unseen molecules. Our dataset consists of spectra for a few dyes measured in-house via spectrophotometry, enabling held-out evaluation at the molecule level. We assess recovery of peak positions and band shapes relevant to clearing performance. This approach reduces dependence on serial dilution and instrument time, integrates naturally with tissue-clearing workflows, and has broader utility in chemical sensing, environmental monitoring, and optoelectronic devices such as organic photovoltaics and photodetectors.