Home > Journals > Minerva Biotechnology and Biomolecular Research > Past Issues > Minerva Biotechnology and Biomolecular Research 2021 December;33(4) > Minerva Biotechnology and Biomolecular Research 2021 December;33(4):210-8



Publishing options
To subscribe
Submit an article
Recommend to your librarian


Publication history
Cite this article as



Minerva Biotechnology and Biomolecular Research 2021 December;33(4):210-8

DOI: 10.23736/S2724-542X.20.02669-5


language: English

Estimation of maximum recommended therapeutic dose of anti-retroviral drugs using diversified sampling and varied descriptors

Roopshikha SAHU, Amisha YADAV, Abhigyan NATH

Department of Biochemistry, Pt. Jawahar Lal Nehru Memorial Medical College, Raipur, India

BACKGROUND: The maximum recommended therapeutic dose (MRTD) is defined as the upper limit of a drug’s dose beyond which no improvement in efficacy is observed and detrimental effects exceeds the benign effects. Estimates of MRTD facilitate in the assessment of toxic dose levels of drugs and can reduce the toxicological attrition of putative drugs in drug development pipelines.
METHODS: In the present work, we developed a machine learning based prediction model for MRTD of anti-retroviral (ARVs) drugs, using a varied range of molecular descriptors (JoeLib, ChemmineR, OpenBabel and RDKit). Representative and diversified training and testing data are of utmost importance for proper/complete learning and for robust evaluation respectively of the full input descriptor space by the machine learning algorithms. We implemented a dissimilarity based uniform sampling approach (Kennard-Stone algorithm) for obtaining the diversified and representative training/testing set.
RESULTS: Using ReliefF feature selection algorithm and representative training/testing set, a set of top 30 feature subset is extracted, which resulted in locally weighted predictor to achieve a root mean square error of 5.71 and a correlation coefficient of 0.71. Further, we compared the dissimilarity based uniform sampling with similarity-based K-means sampling approach for train/test split, on the performance accuracy of the learning algorithm.
CONCLUSIONS: Our results show that, K-means based sampling cannot guarantee a representative training set on a single run, while the performance of the uniform based sampling resulted in enhanced performance evaluation metrics than the previously reported results.

KEY WORDS: Anti-retroviral agents; Pharmaceutical preparations; Drug dosage calculations

top of page