SUSadettin Yavuz Ugurlu

Sadettin Yavuz Ugurlu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Sadettin Yavuz Ugurlu has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    A Chemically Interpretable and Leakage‐Safe Machine Learning Framework for Predicting Molecular Refractive Index ( n D )

    by Sadettin Y. Ugurlu, Emre Bal

    The molecular refractive index () governs light–matter interactions and underpins optical biosensors, bioimaging agents, lab‐on‐a‐chip devices, and functional biomaterials. Existing machine‐learning models for emphasize predictive accuracy but provide limited chemical interpretability and safeguards against information leakage. Here, we present an interpretable, leakage‐safe framework for predicting from structure‐derived descriptors and fingerprints. A curated dataset is processed via multi‐toolkit feature generation, removal of low‐information channels, diversity‐aware filtering, and ANOVA F‐test feature selection within a pipeline that performs median imputation and standardization. Sparse and generalized linear models enable attribution of variation to chemically meaningful descriptors. The final Lasso model shows strong generalization, with a mean fold‐specific test performance of (MAE ), the highest fold‐specific test performance of , and a definitive held‐out unseen test‐set performance of (MAE ) after retraining on the complete training set. Ablation studies identify feature selection, pipeline‐based preprocessing, and targeted tuning as the primary contributors to predictive performance in the selected framework.

  • Open Access

    PDB-Quantifier: Reference-Based Protein Structure Quality Assessment and Prioritization via One-Class Learning

    by Sadettin Y. Uğurlu

    A curated ADS benchmark set containing 85 protein structures was used as the reference domain. For each structure, feature vectors were constructed from experimentally reported and validation-related descriptors, including resolution, refinement statistics, atom and residue counts, ligand-related variables, and structural validation measures such as clashscore and Ramachandran outliers. Missing values were handled by mean imputation, and an Isolation Forest model was trained to learn the intrinsic distribution of the reference set. The trained model was then applied to two independent external datasets: an MTi benchmark set and a randomly sampled Protein Data Bank set.

  • Open Access

    Experimental dipole moment prediction with a progressive 2D–3D hybrid framework and twin-pair-based diagnostic evaluation

    by Sadettin Y. Ugurlu, Shan He

    Accurate prediction of molecular dipole moments ( $$\mu$$ μ ) is essential for modeling electrostatic interactions in solvation, molecular recognition, and materials. While modern 3D learning models can approach saturation on low-noise quantum-chemical benchmarks, performance often degrades on experimentally compiled datasets where measurement noise, stereochemical ambiguity, and conformer variability introduce a substantial gap between idealized theory and practical data. Here, we present a progressive 2D–3D hybrid framework for experimental dipole-moment prediction under heterogeneous labels. The framework is constructed in two stages. First, a strong 2D tabular CatBoost model based on SMILES-derived multi-view fingerprints is enriched with physicochemical, dipole-related, and conformer-dependent 3D shape descriptors, introducing a first level of 2D+3D integration and improving performance over the 2D representation alone. Second, this enriched tabular predictor is linearly fused with a geometry-aware 3D graph neural network operating on conformer-derived molecular graphs, yielding a further gain through complementary structural learning. On the experimental test set, the resulting framework achieves $$R^{2}=0.844$$ R 2 = 0.844 with $$\textrm{MAE}=0.399~\textrm{D}$$ MAE = 0.399 D , outperforming a nonparametric fingerprint k -NN baseline ( $$R^{2}=0.65$$ R 2 = 0.65 ) and remaining close to the similarity-conditioned diagnostic reference range estimated from twin-pair dispersion ( $$R^{2}\approx 0.86$$ R 2 ≈ 0.86 –0.91) under Tanimoto similarity thresholds of 0.95, 0.97, 0.98, and 0.99. Scientific contribution: The main scientific contribution of this study is the development of an experimentally grounded progressive 2D–3D hybrid learning framework that integrates enriched tabular chemical representations, geometry-aware graph learning, and twin-pair-based diagnostic evaluation to predict dipole moments under heterogeneous experimental labels. To contextualize performance in this noise-affected experimental setting, we further use twin-pair analysis within a test-to-train similarity-stratified evaluation protocol, while baseline comparisons and component-wise interpretability analyses help clarify the roles of representation diversity and learner complementarity. Relative to three recently published dipole-prediction studies, the proposed framework remains competitively positioned, although these comparisons are intended only for cautious literature contextualization because the underlying datasets, label sources, and noise regimes are not directly matched. In contrast to prior studies focused mainly on lower-noise theoretical benchmarks or narrower molecular domains, the present work targets the comparatively underexplored Stenutz experimental compilation.

  • Open Access

    Interpretable and Robust Hammett Substituent Constant Prediction via Stacking Ensemble with Systematic Feature Engineering and Explainability Analysis

    by Sevval Yontem, Ecenur Soybelli, Baris Ayhan, Mustafa M. Can, Sadettin Y. Ugurlu

    The Hammett substituent constant σ is one of the most widely used empirical scales for quantifying the electron-withdrawing and electron-donating effects of aromatic substituents, yet its experimental determination remains resourceintensive. Although several machine-learning (ML) models—including graph neural networks such as GCN and Weave—have been proposed as predictive alternatives, existing approaches either rely on black-box architectures that offer no chemical interpretability or on constrained feature sets that limit generalisability. Here we present a systematic, multi-stage ML framework—designated σ-Stack—that addresses both shortcomings simultaneously. Starting from a cleaned dataset of 984 aromatic compounds described by 1,836 molecular descriptors computed with RDKit, Mordred, and PyBioMed, we applied correlation-based feature selection (F.S-2; 131 descriptors), Yeo-Johnson power transformation of the target, and Optuna-guided hyperparameter optimisation to build six base learners (Extra Trees, XGBoost, LightGBM, GradientBoosting, CatBoost, and a tuned MLP/ANN). These were integrated into a Ridge-meta stacking ensemble that achieved a cross-validated R2 of 0.866 ± 0.025 and a test-set R2 of 0.882 (RMSE = 0.112, MAE = 0.078, on the dimensionless σ 1 scale) on an independent 20% hold-out—values that substantially exceed those of all previously published single-model benchmarks on comparable data. Repeated 5 × 10-fold cross-validation confirmed the absence of split-dependent bias (mean R2 = 0.871 ± 0.025). Explainability was pursued at three complementary levels: global feature importance via SHAP beeswarm plots, local explanation of individual predictions via LIME, and multicollinearity auditing via variance inflation factors (VIF). The intersection of SHAP top-25% importance and VIF < 10 identified three chemically trustworthy descriptors that encode electronic polarisability, hydrogen-bond donor capacity, and molecular topology. An ablation study quantified the additive contribution of each methodological step, and AutoGluon multi-stacking was evaluated as an automated baseline. All training, evaluation, and visualisation scripts are publicly available at https://github.com/sevv007/bioformatic-sigma-Stack.

  • Open Access

    PDBMINER: RANK AGGREGATION-BASED SELECTION OF QUALIFIED PROTEIN DATA BANK STRUCTURES

    by Sadettin Yavuz Uğurlu, Sadettin Y. Ugurlu

    The selection of appropriate Protein Data Bank (PDB) structures is essential for structure-based computational studies such as molecular docking, virtual screening, and molecular dynamics simulations. The quality of the selected structure significantly impacts the reliability and reproducibility of computational outcomes. Researchers typically focus on a limited set of criteria, particularly resolution, for structure selection. However, structural quality encompasses multiple parameters, including crystallographic resolution, R-free, R-observed, R-work, data completeness, mutation status, and structural age. This complexity implies that a PDB entry may excel in one aspect but be inadequate in another, making PDB selection a multi-criteria decision problem. To address this challenge, we developed PDBminer, an automated pipeline designed to systematically identify and prioritize high-quality PDB structures associated with a given UniProt accession. The method retrieves all related PDB entries from the UniProt REST API, extracts structural and experimental quality metrics directly from PDB files, and applies configurable quality filters. Multiple structural criteria—including resolution, R-free, R-observed, R-work, crystallographic completeness, B-factor statistics, mutation presence, and structural age—are integrated using a rank aggregation framework to produce a unified ranking of candidate structures. Additionally, an optional redundancy filtering step removes highly overlapping entries based on UniProt residue coverage, ensuring that only representative structures are retained. Widely studied proteins such as p53 (UniProt: P04637) and the SARS-CoV-2 replicase polyprotein (UniProt: P0DTD1) contain 286 and 2422 associated PDB structures, respectively. Even when only several structural metrics—such as resolution, completeness, R-free, R-observed, R-work, mutation status, and sequence coverage—are considered, manually collecting and comparing these data across thousands of entries can require days of effort. Moreover, identifying optimal structures requires solving a multi-criteria optimization problem involving seven structural quality features, which may take weeks of manual analysis and still fail to guarantee a systematic and unbiased selection. In contrast, PDBminer automatically evaluates all available structures and produces a ranked list of candidate PDB entries within minutes. The final output is a ranked dataset exported as a structured report containing structural metrics, ligand information, and quality indicators for each candidate PDB entry. By combining automated data retrieval, multi-criteria evaluation, and rank aggregation, PDBminer provides a robust and reproducible strategy for selecting optimal PDB structures for structural bioinformatics and computational drug discovery studies. The framework enables systematic prioritization of reliable protein models and serves as a decision-support tool for structure-based computational research.

  • De-TarDis: Decoy-Target Discovery Database Integrating Off-Targets, Promiscuity, Adverse/Side-Effects, and Screening Panels via Reciprocal Rank Fusion for Safety Assessment

    by Sadettin Yavuz Uğurlu

    Computational screening in drug discovery typically concentrates on a single “intended” target; yet as projects approach the clinic, unexpected liabilities—off-target binding, promiscuous (multi-target) proteins, targets implicated in adverse and side effects, and those monitored in safety panels (e.g., Eurofins SafetyScreen™ tiers 1–3)—often drive failure. In order to address such a challenge, this study addresses the lack of a unified resource that enables early, collective checks against such risky targets. We surveyed the literature and public databases, compiling 49 lists organized into five groups: (1) off-target, (2) promiscuous target, (3) adverse-effect target, (4) side-effect target, and (5) safety-check target. Target identifiers were standardized to UniProtKB, and each list was internally ranked using volume and study-specific scores. Reciprocal Rank Fusion (RRF) has been applied to merge these heterogeneous rankings into a single, robust ordering—RRF rewards targets that rank highly across multiple sources, elevating consistently implicated proteins to the top. The resulting resource, “De-TarDis” (Decoy-Target Discovery Database), yields a consolidated “avoid-these-targets” list for computational campaigns. It can be used directly during hit-to-lead and lead-optimization to flag compounds likely to bind safety-relevant proteins, thereby reducing late-stage, ADMET-driven surprises.