Proteochemometric modeling in a Bayesian framework.

Area of Science:

• Computational chemistry and cheminformatics
• Bioinformatics and systems biology
• Drug discovery and development

Background:

• Proteochemometrics (PCM) models relationships between protein and chemical data for bioactivity prediction.
• Gaussian Processes (GP) offer objective uncertainty estimation and allow incorporation of experimental error into probabilistic models.
• Evaluating model applicability domain (AD) is crucial for reliable predictions.

Purpose of the Study:

• To apply Gaussian Processes (GP) with various kernels to diverse proteochemometrics (PCM) datasets.
• To assess the statistical soundness and predictive performance of GP models in bioactivity prediction.
• To explore the interpretability of GP models for biological insights.

Main Methods:

• Application of Gaussian Processes (GP) with multiple kernels to three PCM datasets.
• Datasets included adenosine receptors, dengue virus NS3 proteases, and aminergic GPCRs.
• Statistical evaluation using R^2 and Root Mean Squared Error of Prediction (RMSEP) compared to experimental error.

Main Results:

• GP models demonstrated statistical soundness comparable to Support Vector Machines (SVM).
• R^2 values ranged from 0.68 to 0.92, with RMSEP close to experimental error.
• GP models provided confidence intervals consistent with cumulative Gaussian distribution and yielded biologically meaningful interpretations.

Conclusions:

• Gaussian Processes (GP) are a robust and interpretable method for proteochemometrics (PCM) bioactivity modeling.
• GP's ability to estimate uncertainty and incorporate experimental error enhances model reliability and applicability domain assessment.
• The study highlights GP's potential for advancing drug discovery through accurate and interpretable predictive modeling.