Scholarlink

<p>Degree-based topological indices derived from molecular graphs provide compact, interpretable descriptors of structural connectivity with established molecular properties. This study investigates whether analytically computed indices can serve as chemically interpretable representations for predicting the properties of oncology-relevant compounds and whether estimated surrogates of these indices retain sufficient structural information for competitive downstream prediction. Using a curated dataset of 82 oncology-relevant small molecules spanning diverse mechanistic classes, we computed ten degree-based indices and modeled them from SMILES-derived RDKit descriptors using a stacking ensemble of XGBoost and HistGradientBoosting base learners with a Bayesian Ridge meta-learner. The resulting estimated indices were then used as surrogate features to predict the eight physicochemical properties under the same grouped validation protocol applied to the analytically computed indices. Pearson correlation analysis confirmed that the analytical indices produced stronger and more direct property–index relationships, establishing them as interpretable references. Both performed well for size- and connectivity-related properties, although the estimated indices occasionally returned slightly higher the coefficient of determination values, likely due to the smoothing effect of the learning process rather than any structural advantage over the exact analytical values. Predictive accuracy was lower for thermophysical properties under both settings, as 2D topological indices do not capture three-dimensional molecular behavior. Paired error-propagation analysis using Wilcoxon signed-rank tests and bootstrap confidence intervals confirmed that replacing analytical indices with estimated did not cause a statistically significant degradation in downstream prediction. Grouped Y-scrambling and applicability-domain analyses based on Williams plots further supported the validity of the structure–property relationships. Benchmarking against classical baselines, direct descriptor- and fingerprint-based models, and neural architectures demonstrated that the proposed index-mediated framework achieved competitive predictive performance while uniquely preserving explicit topological interpretability. The present experiments were conducted on a modest-sized dataset, the modular design of the framework makes it readily extensible to larger-scale benchmarking and virtual screening applications, where parallel computation and high-performance computing resources could enable rapid screening of oncology-relevant compounds at pharmaceutical scales.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph-based stacking ensemble approach for physicochemical properties prediction of oncology-relevant compounds

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Irfan Haider,
  • Mingchu Li,
  • Muhammad Kamran Jamil

摘要

Degree-based topological indices derived from molecular graphs provide compact, interpretable descriptors of structural connectivity with established molecular properties. This study investigates whether analytically computed indices can serve as chemically interpretable representations for predicting the properties of oncology-relevant compounds and whether estimated surrogates of these indices retain sufficient structural information for competitive downstream prediction. Using a curated dataset of 82 oncology-relevant small molecules spanning diverse mechanistic classes, we computed ten degree-based indices and modeled them from SMILES-derived RDKit descriptors using a stacking ensemble of XGBoost and HistGradientBoosting base learners with a Bayesian Ridge meta-learner. The resulting estimated indices were then used as surrogate features to predict the eight physicochemical properties under the same grouped validation protocol applied to the analytically computed indices. Pearson correlation analysis confirmed that the analytical indices produced stronger and more direct property–index relationships, establishing them as interpretable references. Both performed well for size- and connectivity-related properties, although the estimated indices occasionally returned slightly higher the coefficient of determination values, likely due to the smoothing effect of the learning process rather than any structural advantage over the exact analytical values. Predictive accuracy was lower for thermophysical properties under both settings, as 2D topological indices do not capture three-dimensional molecular behavior. Paired error-propagation analysis using Wilcoxon signed-rank tests and bootstrap confidence intervals confirmed that replacing analytical indices with estimated did not cause a statistically significant degradation in downstream prediction. Grouped Y-scrambling and applicability-domain analyses based on Williams plots further supported the validity of the structure–property relationships. Benchmarking against classical baselines, direct descriptor- and fingerprint-based models, and neural architectures demonstrated that the proposed index-mediated framework achieved competitive predictive performance while uniquely preserving explicit topological interpretability. The present experiments were conducted on a modest-sized dataset, the modular design of the framework makes it readily extensible to larger-scale benchmarking and virtual screening applications, where parallel computation and high-performance computing resources could enable rapid screening of oncology-relevant compounds at pharmaceutical scales.