Explainable Machine Learning for ETR and Drug Chameleonicity.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Traditional in silico methods for drug design are computationally intensive and do not fully capture the behavior of beyond-rule-of-five (bRo5) drugs.
• Designing orally bioavailable bRo5 drugs requires understanding and optimizing physicochemical properties, particularly polarity.

Purpose of the Study:

• To develop an explainable machine learning model for identifying molecular "hot spots" that influence chameleonicity in bRo5 drugs.
• To guide rapid chemical design for improved oral absorption of complex bRo5 drug candidates.

Main Methods:

• Introduction of the EPSA-to-TPSA ratio (ETR) as a high-throughput measure of polarity reduction.
• Development of an explainable deep learning model using a large dataset of bRo5 molecules (macrocycles, PROTACs).
• Validation of model insights using molecular dynamics simulations.

Main Results:

• The explainable deep learning model accurately predicts EPSA and identifies polarity-reducing "hot spots" influencing drug chameleonicity.
• The EPSA-to-TPSA ratio (ETR) provides a high-throughput measure for assessing polarity reduction in bRo5 compounds.
• Model predictions were validated by molecular dynamics, enabling robust, high-throughput assessment of bRo5 chameleonic behavior.

Conclusions:

• Explainable AI can effectively guide chemical modifications for optimizing physicochemical properties of bRo5 drugs before synthesis.
• This approach establishes new frameworks for designing complex bRo5 drugs with improved oral bioavailability, building upon existing descriptors like Lipinski's rules.