Cytotoxicity of phytosynthesized silver nanoparticles: A meta-analysis by machine learning algorithms

Digital Innovation Enabled Nanomaterial Manufacturing; Machine Learning Strategies and Green Perspectives

“…The major battlement remains related to the access to clinical applications data, and actual paradigms such as immunotherapy [ 40 ]. For example, it is known that it is easier for the smaller nanoparticles to penetrate the cellular membranes and interact with organelles [ 27 ], but other individual features also affect the nanomaterial behavior and interactions. Digital infrastructures and databases are expected to bridge this gap of knowledge and support the discovery of nanomaterials that can embrace the activation and inhibition of therapeutical mechanisms based on their individual features; though it required a massive effort to create general rules [ 37 ].…”

“…Cell uptake, damages to DNA, mitochondria and cell membrane, and ROS production have been identified as effects that are mainly size-dependent, which rises the importance of a standardized and widely accepted monitoring technique, which can provide high-throughput information about the statistical representation of size distribution. Current techniques suffer from the agglomeration state of nanoparticles in the case of DLS, while in TEM observation the nanoparticles might be affected by the test media [ 27 ].…”

“…Understanding nanomaterials in depth has been the motivation to collect key materials descriptors to support the universality of the computational models [ 25 ]. Tensile strength, elongation at break, thermal and electrical conductivity, Young’s modulus, porosity, (eco-)toxicity indexes [ 10 , 26 ], topology, crystallography, aspect ratio, zeta potential, surface energy, chemical structure, and protein corona fingerprint have been common descriptors used to establish structure–property–response relationships [ 27 , 28 , 29 ]. Coming with a single and robust model challenging [ 28 ], and requires time investment to collect, curate, and analyze the data.…”

“…Regarding model development, the best-case scenario is to have a verified model concerning the correctness and reliability, and to focus on the validation of the model predictions based on the training dataset, in order to avoid high levels of complexity and possible overfitting of the derived model [ 12 , 17 , 23 , 31 , 32 ]. Cross-validation is a common procedure to optimize the prediction metrics and achieve similar fitting accuracy to unseen datasets, but it should be noted that sampling is important, as both the training and validation datasets should be representative of the data [ 27 , 32 , 33 , 34 ]. The repeated learning test and k-fold cross-validation have been shown to be the most promising approach to reduce the computational costs in this matter [ 35 ].…”

“…Another factor is that existing data usually have not been annotated and costly man-made annotation processes have to be followed in order to ensure that data are well-structured for the efficient establishment of predictive models [ 38 ]. Among datasets, different descriptors/features (rows or columns) are evidenced, and often one consideration is the “dimensionality curse”, which can lead to overtraining and induce bias in the developed models based on their observation frequency, and their ability to reveal hidden relationships in a family of materials is affected [ 12 , 27 ]. Some impressive demonstrations of machine learning include models to detect potential Heusler compounds and properties.…”