Decoupled peak property learning for efficient and interpretable electronic circular dichroism spectrum prediction.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Electronic circular dichroism (ECD) spectroscopy is vital for determining molecular chirality, crucial in drug development and asymmetric synthesis.
• Current predictive methods for ECD spectra face limitations due to data scarcity and poor interpretability, hindering trustworthy predictions.

Purpose of the Study:

• To develop a large-scale dataset for chiral molecular ECD spectra.
• To propose ECDFormer, an accurate and interpretable model for ECD spectrum prediction.
• To enhance the understanding of molecular chirality through spectral data.

Main Methods:

• Established a comprehensive dataset of chiral molecular ECD spectra.
• Developed ECDFormer, a novel architecture utilizing QFormer to decompose ECD spectra into peak entities.
• Employed a decoupled peak prediction strategy for enhanced accuracy and efficiency.

Main Results:

• Achieved a significant improvement in peak symbol accuracy, increasing from 37.3% to 72.7%.
• Dramatically reduced prediction time from 4.6 CPU hours to 1.5 seconds.
• Demonstrated ECDFormer's capability to extract molecular orbital information and its proficiency across various spectroscopic techniques (ECD, IR, MS).

Conclusions:

• ECDFormer offers a breakthrough in accurate and interpretable ECD spectrum prediction.
• The decoupled peak prediction approach significantly outperforms traditional sequence prediction methods.
• ECDFormer exhibits strong generalization capabilities, applicable to diverse spectroscopic data and crucial for chiral molecule analysis.