The Pnictogen Oracle Provider: Multi-Fidelity Graph Landscapes To Benchmark Molecular Optimization
- 1Instituto de Micro y Nanotecnología (IMN-CNM), CSIC, Isaac Newton 8, Tres Cantos, Madrid, Spain, 28760
- 2Department of Chemistry and Chemical Biology, McMaster University, 1280 Main Street West, Hamilton, Canada, L8S 4L8
Abstract
The informed discovery and inverse design of new molecules and materials is among the most active applications of data-driven procedures, such as statistical learning, to chemical problems. These approaches involve two key aspects: the dataset, and the exploration algorithms. Here, we cover both of these dimensions by introducing POP (Pnictogen Oracle Provider), a platform comprised by an extensive graph database and an optimization toolkit. The database contains almost 200,000 structures for phosphines, amines and arsines at different fidelity levels, explicitly including failed calculations and enforcing a topology based on substituent-level mutation operations. On the other hand, the software toolkit streamlines the application of optimization algorithms (search policies and predictive oracles) to such databases, pushing forward the development of benchmarks. We explore the baseline application of a varied set of such methods—evolutionary algorithms, Bayesian Optimization, LLM-based procedures and multifidelity approaches—to the optimization of a key property: the HOMO-LUMO gap across each of the three ligand families. In this manner, we reveal the relevance of graph topology in the guidance of these algorithms, as well as the performance differences these baseline methods depending on the particularities of each of the ligand families.
Related
- Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark
- Rescuable Stagnation: A Diagnostic for Benchmark Suitability in LLM-Guided Materials Optimization
- Accuracy, Reliability, and Efficiency of a MACE Machine-Learning Interatomic Potential for Aspirin: A Reproducible Benchmark for Molecular Simulation
- Bicarbonate/carbonic acid equilibria in aqueous solutions: Conformational Dynamics, Ion-Pair Speciation, and Free Energy Landscapes from Ab Initio Molecular Dynamics and Metadynamics
- Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark
BibTeX
@misc{garayruiz2026pnictogen,
title = {{The Pnictogen Oracle Provider: Multi-Fidelity Graph Landscapes To Benchmark Molecular Optimization}},
author = {Diego Garay-Ruiz and Rodrigo A. Vargas-Hernández and Sergio Pablo-García},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010180/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010180/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - The Pnictogen Oracle Provider: Multi-Fidelity Graph Landscapes To Benchmark Molecular Optimization
AU - Diego Garay-Ruiz
AU - Rodrigo A. Vargas-Hernández
AU - Sergio Pablo-García
PY - 2026
DA - 2026/10/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010180/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010180/v1
AB - The informed discovery and inverse design of new molecules and materials is among the most active applications of data-driven procedures, such as statistical learning, to chemical problems. These approaches involve two key aspects: the dataset, and the exploration algorithms. Here, we cover both of these dimensions by introducing POP (Pnictogen Oracle Provider), a platform comprised by an extensive graph database and an optimization toolkit. The database contains almost 200,000 structures for phosphines, amines and arsines at different fidelity levels, explicitly including failed calculations and enforcing a topology based on substituent-level mutation operations. On the other hand, the software toolkit streamlines the application of optimization algorithms (search policies and predictive oracles) to such databases, pushing forward the development of benchmarks. We explore the baseline application of a varied set of such methods—evolutionary algorithms, Bayesian Optimization, LLM-based procedures and multifidelity approaches—to the optimization of a key property: the HOMO-LUMO gap across each of the three ligand families. In this manner, we reveal the relevance of graph topology in the guidance of these algorithms, as well as the performance differences these baseline methods depending on the particularities of each of the ligand families.
ER -