Search for predictive generic model of aqueous solubility using Bayesian neural nets

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Existing aqueous solubility models often lack applicability in drug hunting due to limited structural diversity in training data.
• There is a need for robust predictive models that generalize well to novel chemical structures relevant to drug research.

Purpose of the Study:

• To develop and evaluate predictive models for aqueous solubility using diverse chemical datasets.
• To assess the performance of different modeling approaches, including Bayesian neural networks and descriptor selection methods.

Main Methods:

• Gathered a diverse dataset of compounds from literature and proprietary sources, divided into literature (I), proprietary (II), and mixed (III) sets.
• Calculated approximately 100 surface-property descriptors for each compound.
• Employed Bayesian learning of neural networks with descriptor selection via Gram-Schmidt (GS) or automatic relevance determination (ARD) on datasets I, II, and III.
• Validated model performance using unrelated datasets and introduced new metrics: Normalized Descriptor Distance (NDD) and Combination of Descriptor Distance (CD).

Main Results:

• A generic predictive model could be derived from the literature dataset (I).
• The proprietary dataset (II) was too limited in diversity to produce a model with good generalization ability.
• The automatic relevance determination (ARD) method applied to the mixed dataset (III) yielded the best predictive model.

Conclusions:

• The study highlights the importance of dataset diversity for developing generalizable predictive models in drug discovery.
• Bayesian neural networks combined with ARD on a mixed dataset offer a promising approach for accurate aqueous solubility prediction.
• The developed model demonstrates improved applicability for drug hunting compared to models trained on less diverse datasets.