Background Feature generation drives machine learning-based drug discovery by producing unique, compact, invariant, and computationally efficient descriptors. In structural protein–ligand binding affinity prediction, machine learning has developed a range of feature generation methods, with geometry-based techniques achieving the highest predictive accuracy—particularly when used with the eXtreme Gradient Boosting (XGBoost) algorithm. Unfortunately, these techniques remain largely inaccessible to end-users due to their reliance on programming expertise. To address this, we introduce <Emphasis FontCategory="NonProportional">REINDEER</Emphasis>, a software tool that streamlines geometry-based feature generation through an intuitive Graphical User Interface (GUI), Command Line Interface (CLI), and Python API. Results <Emphasis FontCategory="NonProportional">REINDEER</Emphasis> implements four state-of-the-art feature generation methods: Occurrence of Interatomic Contact (OIC), Distance-Weighted Interatomic Contact (DWIC), Extended Connectivity Interaction Feature (ECIF), and Multi-Shell Occurrence of Interatomic Contact (MS-OIC). The software also supports parallel processing to accelerate computation. To demonstrate the <Emphasis FontCategory="NonProportional">REINDEER</Emphasis>’s capabilities, and validate its implementations, we evaluate it on a novel protein-peptide dataset, assessing its performance in three key areas: (i) computational speed under CPU parallelization, (ii) binding affinity prediction accuracy using XGBoost, and (iii) uncertainty quantification via conformal prediction. Results confirm the computational efficiency of the implemented methods and demonstrate that the generated features produce predictive models in a practical workflow. Among the four methods, ECIF paired with XGBoost yields the highest Pearson’s correlation coefficient (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\mathrm R_{P}~=~0.651\)</EquationSource> </InlineEquation>). Furthermore, the XGBoost-ECIF conformal predictor proves both valid (mean accuracy of 81%) and efficient (mean confidence region size of 1.442). Conclusions By democratizing access to advanced feature generation methods through GUI, CLI, and Python API, <Emphasis FontCategory="NonProportional">REINDEER</Emphasis> enables researchers with limited programming expertise to harness cutting-edge machine learning techniques in designing machine learning-based scoring functions for their specific problems.

Reindeer: a protein-ligand feature generator software for machine learning algorithms

期刊论文
出版模式：

开放获取
发表日期：

2026年4月22日

查看全文

Milad Rayka,
S. Shahab Naghavi

摘要

Background

Feature generation drives machine learning-based drug discovery by producing unique, compact, invariant, and computationally efficient descriptors. In structural protein–ligand binding affinity prediction, machine learning has developed a range of feature generation methods, with geometry-based techniques achieving the highest predictive accuracy—particularly when used with the eXtreme Gradient Boosting (XGBoost) algorithm. Unfortunately, these techniques remain largely inaccessible to end-users due to their reliance on programming expertise. To address this, we introduce REINDEER, a software tool that streamlines geometry-based feature generation through an intuitive Graphical User Interface (GUI), Command Line Interface (CLI), and Python API.

Results

REINDEER implements four state-of-the-art feature generation methods: Occurrence of Interatomic Contact (OIC), Distance-Weighted Interatomic Contact (DWIC), Extended Connectivity Interaction Feature (ECIF), and Multi-Shell Occurrence of Interatomic Contact (MS-OIC). The software also supports parallel processing to accelerate computation. To demonstrate the REINDEER’s capabilities, and validate its implementations, we evaluate it on a novel protein-peptide dataset, assessing its performance in three key areas: (i) computational speed under CPU parallelization, (ii) binding affinity prediction accuracy using XGBoost, and (iii) uncertainty quantification via conformal prediction. Results confirm the computational efficiency of the implemented methods and demonstrate that the generated features produce predictive models in a practical workflow. Among the four methods, ECIF paired with XGBoost yields the highest Pearson’s correlation coefficient ( \(\mathrm R_{P}~=~0.651\) ). Furthermore, the XGBoost-ECIF conformal predictor proves both valid (mean accuracy of 81%) and efficient (mean confidence region size of 1.442).

Conclusions

By democratizing access to advanced feature generation methods through GUI, CLI, and Python API, REINDEER enables researchers with limited programming expertise to harness cutting-edge machine learning techniques in designing machine learning-based scoring functions for their specific problems.