Machine-learning reveals the virtual screening strategies of solid hydrogen-bonded oligomeric assemblies for thermo-responsive applications

The assembly of hydrogen-bonding (H-bonding) motifs enables the phase transition of oligomeric materials, and facilitates thermo-responsive properties such as self-healing and melt-castable capabilities. However, the dis-covery of solid H-bonded oligomers remains in the laboratory because of the absence of practical virtual screening (VS) strategies. Herein, a machine-learning (ML) approach is proposed to realize the vS for solid H-bonded oligomers with high accuracy (>93 %). A synthetic library comprising 770 oligo-dimethylsiloxane (oDMS) with structurally diverse H-bonding motifs was constructed, and the solid/fluid-state labels of oDMS were obtained via rheological measurements. Tailored descriptors for H-bonding motifs, which were derived from quantum chemistry calculation and analysis, were adopted for ML algorithms, including interaction energy, AlogP98, molecular flexibility, van der Waals area and Connolly surface occupied volume values. The eXtreme Gradient Boosting (XGBoost) model presents the best prediction accuracy among the investigated ML algorithms and the interpretation of XGBoost model provides the feasible vS routines for the discovery of solid H-bonded oligomers.

Automated summary

This study proposes a machine-learning approach to achieve virtual screening (VS) for solid hydrogen-bonded oligomers with high accuracy (>93%). A synthetic library consisting of 770 structurally diverse oligo-dimethylsiloxane (oDMS) with different hydrogen-bonding motifs was constructed, and the solid/fluid-state labels of oDMS were obtained through rheological measurements. Tailored descriptors for hydrogen-bonding motifs, derived from quantum chemistry calculation and analysis, were used for machine learning algorithms, including interaction energy, AlogP98, molecular flexibility, van der Waals area, and Connolly surface occupied volume values. The eXtreme Gradient Boosting (XGBoost) model showed the best prediction accuracy among the investigated algorithms, and the interpretation of the XGBoost model provides feasible VS routines for the discovery of solid hydrogen-bonded oligomers.