Numerical Experiments Bayessian Illumination

Numerical Experiments for Bayesian Illumination This file contains the numerical experiments conducted for the manuscript titled “Bayesian Illumination: Inference and Quality-Diversity Accelerate Generative Molecular Models.” The experiments provide comprehensive benchmarking and validation of the Bayesian Illumination algorithm, which integrates Bayesian optimization with quality-diversity methods to improve molecular discovery. The data includes: Descriptor-Based Rediscovery: Results from a novel benchmark where molecules are rediscovered based on conformer samples and descriptor-based representations (USRCAT and Zernike descriptors). Efficient Organic Photovoltaics: Performance metrics from multiple tasks, including maximising HOMO-LUMO gap values, minimising LUMO energyvalues, maximising the molecular dipole moment and maximising a combined efficiency score). Docking-Based Tasks: Outputs from docking-based optimizations, including stringent structural and physicochemical filtering to ensure realistic molecular designs. This also includes synthetic accessibility-adjusted docking scores. This file serves as a supplement to the manuscript, providing the raw data and detailed performance metrics to support the reproducibility of the findings.贝叶斯光照（Bayesian Illumination）数值实验 本文件包含为题为《贝叶斯光照：推理与质量多样性方法加速生成式分子模型》的手稿所开展的数值实验。本实验对贝叶斯光照（Bayesian Illumination）算法进行了全面的基准测试与验证，该算法将贝叶斯优化与质量多样性方法相结合，以优化分子发现流程。 本数据集涵盖以下内容： 1. 基于描述符的分子重构：来自新型基准测试的结果，该基准通过构象样本与基于描述符的表征（USRCAT描述符与Zernike描述符）实现分子重构。 2. 高效有机光伏：涵盖多项任务的性能指标，包括最大化HOMO-LUMO能隙值、最小化LUMO能级、最大化分子偶极矩，以及最大化综合效率得分。 3. 基于分子对接的任务：分子对接优化的输出结果，包含严格的结构与理化过滤步骤，以确保生成的分子设计具备现实可行性，同时还纳入了经合成可及性调整后的对接得分。 本文件作为该手稿的补充材料，提供原始数据与详细性能指标，以支撑研究结果的可重复性。