UV-adVISor: Attention-Based Recurrent Neural Networks to Predict UV-Vis Spectra.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Ultraviolet-visible (UV-Vis) absorption spectra are crucial for identifying chemical reaction products in High-Performance Liquid Chromatography (HPLC).
• Predicting UV-Vis spectra computationally from molecular structure alone can streamline chemical analysis.

Purpose of the Study:

• To introduce UV-adVISor, a novel computational tool for predicting UV-Vis spectra directly from molecular structures.
• To develop and validate deep learning models for accurate UV-Vis spectrum prediction.

Main Methods:

• UV-Vis spectrum prediction framed as a sequence-to-sequence problem.
• Utilized Long-Short Term Memory and attention-based neural networks with Extended Connectivity Fingerprint Diameter 6 or molecule SMILES.
• Trained and validated models on two distinct spectrum datasets (N=949 and N=2222).

Main Results:

• Achieved strong prediction accuracy, with key statistical measures including RMSE (0.064), R² (0.71), and dynamic time warping (DTW, 0.194) on the test set.
• Model performance was validated by demonstrating degraded accuracy when molecular structures were scrambled during training.
• The tool provides rapid and precise spectral predictions for compound libraries.

Conclusions:

• UV-adVISor successfully predicts UV-Vis spectra from molecular structures using advanced neural network architectures.
• The developed models are accurate and reliable, offering a valuable tool for chemical analysis and compound identification.
• The computational approach significantly enhances the efficiency of spectral prediction for large compound sets.