Machine Learning Based Quantitative Structure-Dissolution Profile Relationship.

Area of Science:

• Computational chemistry and drug discovery
• Machine learning applications in pharmaceutical sciences

Background:

• Accurate drug dissolution profiling is crucial for estimating oral drug bioavailability.
• Existing methods can predict drug solubility from chemical structures, but not dissolution rate constants.
• A reliable tool for predicting dissolution rate constants is currently lacking.

Purpose of the Study:

• To develop a novel two-stage machine learning approach for predicting drug dissolution profiles.
• To integrate physics-informed neural networks (PINNs) and deep neural networks (DNNs) for this prediction.
• To predict drug dissolution in water with varying Sodium Lauryl Sulfate concentrations.

Main Methods:

• A two-stage machine learning approach: Machine Learning based Quantitative Structure-Dissolution Profile Relationship.
• Stage 1: PINNs extract dissolution parameters (dissolution rate constant 'k' and dissolved mass fraction at saturation 'ϕs') from data, leveraging physical laws and the Noyes-Whitney equation.
• Stage 2: A DNN uses extracted parameters, drug chemical structure, and dissolution medium to predict dissolution profiles.

Main Results:

• The DNN achieved an average testing accuracy of 61.7% using FDA-recommended f1 and f2 metrics.
• This accuracy was achieved with an 80:20 train-to-test split, below the typical 70-80% acceptance range.
• PINNs demonstrated potential in enhancing prediction performance and reducing data requirements by filtering noise.

Conclusions:

• The developed two-stage machine learning approach shows significant potential as a low-cost, time-efficient tool for early drug formulation.
• Further improvements are anticipated with increased data quality and diversity.
• This method offers a promising avenue for predicting drug dissolution profiles, aiding in drug discovery and development.