Mapping the Applicability Domain of Lipophilicity Models Through Chemical Space Topology

Read full text Summarise with VIRAT
Authors: Malikussaid, Septian Caesar Floresko, Ade Romadhony, Warih Maharani
Cited by: 0
References: 26

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The applicability domain (AD) of a QSAR model defines the boundary of chemical space within which its predictions are statistically valid. For large-scale lipophilicity (logP) modeling, conventional AD formulations produce a binary in-domain/out-of-domain classification that conceals the internal topology of prediction uncertainty. This paper presents a PCAbased spatial error analysis over a 426,850-molecule bioactive dataset assembled from three independent chemical databases. The seven-dimensional descriptor space reduces to a 2D manifold capturing 76.83% of total variance (PC1: 52.85%; PC2: 23.98%), revealing a bimodal topology: Lipinski-compliant drug-like molecules form a dense central core, while pharmacologically extreme molecules scatter at the periphery. Overlaying ElasticNet prediction residuals (computed exclusively on held-out test molecules) demonstrates that mean absolute error (MAE) is 2.73× higher at the periphery than at the central manifold (1.567 vs. 0.574 logP units), with non-overlapping bootstrap 95% confidence intervals confirming significance at α=0.01. A Lipinski-stratified dual-model strategy motivated by this topology achieves RMSE=0.838 for the Drug-Like population and R²=0.767 for the Extreme population, the latter surpassing the global XGBoost baseline (R²=0.765). The contribution is primarily methodological: PCA-based error topology provides an interpretable, quantitative diagnostic for AD characterization that directly informs production deployment decisions. The approach is general and not specific to logP.

Fields of Study & Subjects

Computational Drug Discovery Methods Machine Learning in Materials Science Advanced Chemical Physics Studies

Publication details

Pages
1–6
Publication date
1 July 2026
Language
EN