Machine Learning For Predicting Cytotoxicity Of Bioactive Glass Nanoparticles

Bioactive glass nanoparticles(BGNP)have been widely used in oral and craniomaxillofacial surgery due to their superior osteoinductive ability and potential to promote soft tissue regeneration/repair and remineralization.As high-throughput experiments and high-throughput theoretical predictions have gradually shifted biomaterials research from the traditional “Edison model”(trial and error)to a data-driven paradigm,understanding the physicochemical properties-activity/toxicity relationships of BGNPs is essential for predicting the fate of material in vivo,as well as aiding material screening and optimizing material design.The complex and diverse chemical composition and highly interrelated physicochemical properties of BGNPs make it especially challenging to study the specific biological effect of BGNPs using the traditional way,which needs to synthesize BGNPs with different compositions and morphological structures one by one.Machine learning is a method to discover patterns in data “by itself” through computational means or learning algorithms to accomplish classification tasks or find optimal solutions to problems.Excelling at processing high-throughput multidimensional data,machine learning can extract valuable patterns from complex and diverse physicochemical properties and highly heterogeneous toxicological studies of BGNPs to predict cellular responses to BGNPs.To understand the structure-composition-cytotoxicity relationship of BGNPs used in dental and bone repair and regeneration,data samples and relevant properties of BGNPs in vitro cytotoxicity studies were extracted from previous studies by literature search.Regression and classification models were developed by the Random forest(RF)algorithm to predict in vitro cytotoxicity based on the physicochemical properties and experimental conditions of BGNPs,respectively.The key attributes affecting in vitro cytotoxicity of BGNPs were screened by the Recursive feature elimination algorithm(RFE).An exhaustive search through the Pub Med and Web of Science databases yielded 697 relevant papers.33 eligible studies were screened according to the inclusion and exclusion criteria,from which 944 samples of BGNPs cytotoxicity data and 37 quantitative or qualitative attributes,including synthesis method,physicochemical properties,morphological structure,and experimental conditions,were extracted.The results of feature importance analysis showed that the top 5 most relevant attributes related to BGNPs in vitro cytotoxicity were BG concentration,Exposure time,SSA,Zeta potential,and Ca O composition(wt%).RFE screened two sets of critical attributes for the regression and classification models.The regression model based on 15 key features explained 75.8% of the sample variance in the test set with a mean absolute error(MAE)of 0.075;the classification models based on 15 features performed well in the test set,with an Accuracy of 96%,a Macro-F1 of 87% and a ROC-AUC of 96%.The results of this study suggest that by combining literature data mining and machine learning,the physicochemical properties-bioactivity/toxicity relationships of BGNPs can be extracted from previous studies.Also,the in vitro cellular response to the BGNPs can be predicted by building appropriate models.Thus,guidance for nano bioactive glass screening and design is provided.