The use of machine learning methods in the development of nasal dosage forms with cerebroprotective action

Abstract

In order to save resource of active pharmaceutical ingredients and excipients, in the early stages of research, when planning an experiment, it is advisable to use data of the predicted and experimental physicochemical properties stored in different aggregation databases. The information found will reduce the time for composition development and for technology processing. However, the variety of active compounds characteristics and excipients is not always reflected in these services.

Recently, machine learning models have been widely used in various scientific fields; they allow to obtain predictions with high reliability. Given the above, it is relevant and promising to develop models of machine learning to predict the presence of pharmaceutical incompatibilities in the formulation of nasal dosage forms.

The aim of the study is to develop models of machine learning for in silico forecast of the rational composition of nasal dosage forms with cerebroprotective action.

Materials and methods. A dataset, containing data on compounds (active and auxiliary) and characteristics on the presence or absence of interaction (pharmaceutical incompatibility), was used as material. Training datasets were filled by content analysis of PubMed library data (pubmed.ncbi.nlm.nih.gov) manually, by keywords “pharmaceutical incompatibilities”, “physico-chemical compatibility”, “incompatible excipients”) for the last 10 years. The resulting dataset comprises 1185 lines. The methods employed were a set of methods for binary classification of machine learning (pycaret.org) using the programming language Python 3.8 (python.org) in the package management environment Miniconda (conda.io). Pipeline programming was performed using Jupyter notebook package (jupyter.org). The generation of MACCS (Molecular ACCess System keys) in the training dataset was performed using RDKit package (rdkit.org). Specifications of the simplified representation of molecules in the input line (SMILES), in automatic mode, were searched using PubChem service (pubchem.ncbi.nlm.nih.gov).

Results. The obtained data allowed to choose two perspective models of machine learning of binary classification, whose quality was checked on a dataset for verification. Statistical evaluations of the selected models indicate a high probability of in silico prognosis for the presence or absence of pharmaceutical incompatibilities in the development of nasal formulations of cerebroprotective dosage forms. They are posted on the web server of the expert system ExpSys Nasalia (nasalia.zsmu.zp.ua) in the calculations section.

Conclusions. As a result of our research, we have developed machine learning models for in silico prediction of the rational composition of nasal dosage forms with cerebroprotective action. Confirmation of the quality of the pharmaceutical incompatibilities prediction, using the developed models, is checked on a dataset for check. The statistical indicators of the tree_blender (AUC 0.9521, F1 0.9747, MCC 0.9094) and boost_blender (AUC 0.9593, F1 0.9821, MCC 0.9352) models were obtained. The use of machine learning models in pharmaceutical development will contribute to resource conservation and optimization of the composition of the formulation.