Accelerated Chemical Space Generation of High Molar Extinction Organic Sensitizers via Machine Learning.

Abstract

The development of organic sensitizers with high molar extinction (ε) coefficients is important for various light absorption applications. To accelerate the discovery of such compounds, a machine learning (ML) analysis has been applied to explore their vast chemical space. A dataset of 676 organic chromophores is analyzed by designing their electronic, topological, and molecular descriptors to predict their ε. Among the 10 tested ML models, Gradient Boosting, Random Forest, Extra Trees, and Historical Gradient Boosting regressors show good correlation with their experimental and predicted values (R² ≈ 0.70). Their Shapley Feature importance reveals that Subgraph Density of Secondary Carbon-Hydrogen (SdsCH) and logarithm of the partition coefficient- an Der Waals Surface Area Descriptor 8 (SlogP_VSA8) Descriptors have a significant impact on model performance. Additionally, by leveraging breaking retrosynthetic analysis, 3288 novel structures with potential high ε have been synthesized to validate their feasibility through dimensionality reduction analysis. Their synthetic accessibility (SA) calculations identify the top structures for their experimental synthesis in the future. Interestingly, the findings indicate that new structures with SMILES lengths of 35-80 units can exhibit the highest SA.