PROPMAP: A Novel GPU-accelerated Structure-Property Relationship Mapping Algorithm Utilizing Volunteer Computing

化学文摘服务处最近在其登记册中记录了第53百万种独特的化学物质，而第4000万种化学物质仅在9个月前才被编目。随着化学物质的爆炸性增长，这些物质具有哪些物理、化学和生物特性的问题就产生了。现有的基本物理/化学性质的数据，如离解能、logP、生成焓、折射率、沸点和熔点，定义了结构性质关系(SPR)，并可通过SciFinder和Beilstein获取。同样，基本的生物学特性，如与酶的结合常数，可以通过ChEMBL和PubChem数据库获得。本研究创建了一个依托志愿者计算的创新工具，使用一种新的计算算法PROPMAP？用于教育和研究。PROPMAP的新体系结构利用蒙特卡罗随机游走结构生成器和基于图形处理单元(GPU)加速支持向量回归算法的定量结构属性关系(QSPR)模型。PROPMAP以交互方式将用户选择的物理、化学和生物属性映射到感兴趣的任何输入化学结构上。图形表示在基础化学教育中促进对SPR的理解，并使研究中能够有针对性地修改性质。这种新颖的方法通过伯克利网络计算开放基础设施(BOINC)利用志愿者GPU计算来克服训练和交叉验证超过100万种物质的数据集的模型所带来的高昂的计算费用。PROPMAP对科学界的影响进一步扩大，通过万维网接口免费提供该工具，将其用作教育和研究工具，同时还通过讲习班和研讨会直接培训各机构使用该工具。

The Chemical Abstract Services (CAS) recently recorded the 53 millionth unique chemical substance in the CAS registry with the 40 millionth being cataloged only 9 months prior. With this explosive growth in chemical substances, the question of what physical, chemical, and biological properties are possessed by these substances arises. The existing data for fundamental physical/chemical properties, such as dissociation energies, logP, enthalpies of formation, refractive indexes, boiling points, and melting points define Structure Property Relationships (SPR) and is accessible through SciFinder and Beilstein. Similarly, fundamental biological properties such as binding constants to enzymes are available through the ChEMBL and PubChem databases. This research creates an innovative tool relying on volunteer computing for predicting and mapping SPRs three-dimensionally using a novel computational algorithm ?PROPMAP? for education and research. The novel architecture of PROPMAP utilizes a Monte Carlo random walk structure generator and Quantitative Structure Property Relationship (QSPR) models based on a graphics processing unit (GPU) accelerated Support Vector Regression algorithm. PROPMAP interactively maps user-selected physical, chemical, and biological properties onto any input chemical structure of interest. Graphical representation fosters understanding of SPRs in basic chemistry education and enables targeted property modifications in research. This novel approach utilizes volunteer GPU computing through the Berkeley Open Infrastructure for Network Computing (BOINC) to overcome the otherwise prohibitive computational expense of training and cross validating models for data sets in excess of one million substances. The impact of PROPMAP on the scientific community is broadened by making the tool freely available through a WWW interface for use as an educational and research tool while also directly training institutions on its utility through workshops and seminars.