Predicting human liver microsomal stability with machine learning techniques.

Area of Science:

• Pharmaceutical Research
• Computational Chemistry
• Drug Discovery

Background:

• Ensuring a continuous pipeline in pharmaceutical research requires lead candidates with adequate metabolic stability.
• In vitro ADMET (absorption, distribution, metabolism, elimination, and toxicity) screening offers insights into compound metabolic stability.
• Efficient data processing is crucial for managing large compound libraries and high-throughput screening data before synthesis.

Purpose of the Study:

• To establish a predictive relationship between chemical structure and metabolic stability for in-house compounds.
• To evaluate the efficacy of various in silico machine learning models for metabolic stability prediction.

Main Methods:

• Utilized a dataset of 1952 proprietary compounds classified as stable or unstable.
• Calculated 193 molecular descriptors using Molecular Operating Environment.
• Applied machine learning techniques including random forest, support vector machine (SVM), logistic regression, and recursive partitioning for classification modeling.

Main Results:

• All tested classifiers demonstrated satisfactory performance with accuracy > 0.8, sensitivity > 0.9, specificity > 0.6, and precision > 0.8.
• Random forest and SVM models achieved kappa values of approximately 0.7 on an independent validation set, outperforming other classification tools.
• The study identified nonlinear/ensemble-based classification methods as particularly effective for in silico ADME modeling.

Conclusions:

• In silico machine learning models can accurately predict drug metabolic stability based on chemical structure.
• Random forest and SVM are highly effective for classifying compound metabolic stability, aiding early-stage drug discovery.
• These computational approaches streamline the analysis of large datasets, accelerating the identification of viable drug candidates.