Machine Learning Models for Predicting Molecular UV-Vis Spectra with Quantum Mechanical Properties.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Accurate ultraviolet-visible (UV-vis) spectra are crucial for high-throughput compound synthesis in drug discovery.
• Experimental determination of UV-vis spectra for numerous novel compounds is costly and time-consuming.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting UV-vis spectra.
• To leverage quantum mechanics (QM) and experimental data for improved spectral predictions.
• To assess the performance of different deep learning architectures for molecular property prediction.

Main Methods:

• Devised four machine learning architectures: UVvis-SchNet, UVvis-DTNN, UVvis-Transformer, and UVvis-MPNN.
• Utilized both QM-predicted and experimentally measured UV-vis spectra as input features.
• Optimized 3D molecular coordinates were employed as input for the models.

Main Results:

• The UVvis-MPNN model demonstrated superior performance compared to other architectures.
• The UVvis-MPNN model achieved a training Root Mean Square Error (RMSE) of 0.06 and a validation RMSE of 0.08.
• The developed model successfully predicted spectral differences between regioisomers.

Conclusions:

• Machine learning, particularly the UVvis-MPNN model, offers a cost-effective and accurate alternative for predicting UV-vis spectra.
• This computational approach accelerates the drug discovery process by enabling rapid property assessment of novel compounds.
• The model's ability to differentiate regioisomers highlights its potential for detailed molecular analysis.