Prediction of Maximum Absorption Wavelength Using Deep Neural Networks

Area of Science:

• Photochemistry
• Computational Chemistry
• Materials Science

Background:

• Fluorescent molecules are crucial for biological detection, requiring tailored photophysical properties like specific wavelengths for optimal signal-to-noise.
• Predicting these optical properties from first principles is challenging due to complexity and solvent effects.
• Existing design guidelines for fluorescent compounds are limited.

Purpose of the Study:

• To establish a comprehensive database of solvated small-molecule fluorophores.
• To develop accurate machine learning models for predicting photophysical parameters.
• To accelerate the rational design and discovery of novel fluorescent dyes.

Main Methods:

• Created SMFluo1, a database of 1181 solvated small-molecule fluorophores across UV-Vis-NIR spectrum.
• Developed deep neural network models for predicting photophysical properties.
• Validated the optimal model on 120 out-of-sample compounds.

Main Results:

• The SMFluo1 database covers a wide spectral range for diverse fluorophore applications.
• The deep learning models achieved high accuracy in predicting photophysical parameters.
• The optimal model demonstrated a mean relative error of 1.52% on unseen data.

Conclusions:

• Deep learning offers a promising computational approach to complement experimental and theoretical studies of fluorophores.
• The developed system significantly accelerates the discovery of new dyes with desired optical properties.
• The model's accuracy can be further improved by incorporating data from newly developed fluorophores.