PSoSOQY: A Deep Learning-Driven Singlet Oxygen Quantum Yield Prediction Platform for Expediting Photosensitizer...

Area of Science:

• Photodynamic therapy
• Medicinal chemistry
• Computational chemistry

Background:

• Singlet oxygen quantum yield (SOQY) is crucial for photosensitizer (PSs) efficacy in photodynamic therapy (PDT).
• Current methods for PSs design and SOQY measurement are complex and time-consuming, limiting therapeutic advancement.

Purpose of the Study:

• To develop an accurate and efficient computational model for predicting SOQY.
• To create a platform for rational design and high-throughput screening of novel PSs for PDT.

Main Methods:

• Constructed a comprehensive dataset for SOQY prediction.
• Developed a BiLSTM+attention (BA)-SOQY model integrating deep learning techniques.
• Employed a substructure masking interpretation strategy (SMIS-SMILES) for model explainability.

Main Results:

• The BA-SOQY model achieved high accuracy on a held-out test set (R² = 0.9140).
• Model validation on ESOL and FreeSolv datasets demonstrated strong robustness and generalization (R² > 0.9).
• The PSoSOQY platform successfully predicted SOQY and identified key substructural influences.

Conclusions:

• The BA-SOQY model and PSoSOQY platform offer a powerful tool for accelerating PSs discovery in PDT.
• This approach facilitates rational design and efficient screening, significantly advancing PDT research.