Structural descriptor and surrogate modeling for design of biodegradable scaffolds

Biodegradable scaffolds are important to regenerative medicine in that they provide an amicable environment for tissue regrowth. However, establishing structure-property (SP) relationships for scaffold design is challenging due to the complexity of the three-dimensional porous scaffold geometry. The complexity requires high-dimensional geometric descriptors. The training of such a SP surrogate model will need a large amount of experimental or simulation data. In this work, a schema of constructing SP relationship surrogates is developed to predict the degraded mechanical properties from the initial scaffold geometry. A new structure descriptor, the extended surfacelet transform (EST), is proposed to capture important details of pores associated with the degradation of scaffolds. The efficiency is further enhanced with principal component analysis to reduce the high-dimensional EST data into a low-dimensional representation. The schema also includes a kinetic Monte Carlo biodegradation model to simulate the biodegradation of polymer scaffolds and to generate the training data for the formation of SP relationships. The schema is demonstrated with the design of polycaprolactone biodegradable scaffolds by connecting the initial scaffold geometry to the degraded compressive modulus.

Prediction of the hydrophobicity of platinum(IV) complexes based on molecular surface properties

A quantitative structure-property relationship (QSPR) study was performed for predicting the hydrophobicity of Pt(IV) complexes. Two four-parameter equations, one based solely on structural descriptors derived from electrostatic potentials (ESPs) on molecular surface, and the other integrated ESP descriptors with molecular surface area (A_S), were firstly constructed. Mechanistic interpretations of the structural descriptors introduced were elucidated in terms of solute-solvent intermolecular interactions. Subsequently, several up-to-date modeling techniques, including support vector machine (SVM), least-squares support vector machine (LSSVM), random forest (RF) and Gaussian process (GP), were utilized to build the nonlinear models. Systematical validations including leave-one-out cross-validation, the validation for test set, as well as a more rigorous Monte Carlo cross-validation were performed to verify the reliability of the constructed models. The predictive performances of the four different nonlinear modeling methods follow the order of LSSVM≈GP > RF > SVM. The pure-ESP-based models are generally inferior to the A_S-integrated ones. Comparisons with previous results were made.