Predictions of Colloidal Molecular Aggregation Using AI/ML Models

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Colloidal aggregation of small organic molecules poses challenges in drug discovery screening.
• Predictive models are needed to efficiently triage screening hits and optimize chemical libraries.

Purpose of the Study:

• To develop and validate AI/ML models for predicting colloidal aggregation of small organic molecules.
• To identify key molecular descriptors and chemical features associated with aggregation propensity.
• To apply predictive models for prospective chemical library triage in drug discovery.

Main Methods:

• Utilized various AI/ML techniques including Naïve Bayesian, deep neural networks, logistic regression, recursive partitioning trees, support vector machines, and random forests.
• Trained and tested models on experimentally observed data sets of small organic molecules.
• Employed scaffold tree analysis and matched molecular pair analysis (MMPA) to identify aggregation-driving features.

Main Results:

• Naïve Bayesian and deep neural networks demonstrated the lowest balanced error rate (BER), outperforming other methods.
• Models successfully discriminated between aggregating and non-aggregating molecules.
• Identified hydrophobicity, molecular weight, solubility, fraction of sp3 carbon atoms (Fsp3), and electrotopological state of hydroxyl groups (ES_Sum_sOH) as key descriptors.
• Highlighted the role of scaffolds with high Fsp3 values in preventing aggregation.

Conclusions:

• AI/ML models, particularly Naïve Bayesian and deep neural networks, are effective for predicting colloidal aggregation.
• Fsp3 values and specific chemical scaffolds are important for designing non-aggregating molecules.
• Prospective application of these models enhances chemical library selection and diversity for high throughput screening (HTS).