文献类型：期刊文献

中文题名：基于化学语言的分子物化性质深度学习通用建模框架

英文题名：General deep learning modeling framework for molecular physicochemical properties based on chemical language

作者：曹培林[1];邱钰鑫[1];张翔[1];成洪业[1];漆志文[1];宋震[1]

机构：[1]华东理工大学化工学院,化学工程与低碳技术全国重点实验室,上海200237

年份：2026

卷号：77

期号：1

起止页码：126

中文期刊名：化工学报

外文期刊名：CIESC Journal

收录：;EI(收录号：20262621012692);北大核心:【北大核心2023】；

基金：国家自然科学基金项目(22578115,22208098,22278134);国家重点研发计划项目(2024YFA1510302)。

语种：中文

中文关键词：模型;热力学性质;分子工程;物化性质预测;深度学习;Transformer;SMILES

外文关键词：model;thermodynamic properties;molecular engineering;physicochemical property prediction;deep learning;Transformer;SMILES

摘要：分子性质预测模型可以基于分子结构信息预测其相应的物化、热力学、环境健康安全等关键性质,能够正向或反向指导从庞大的化学空间中筛选和设计满足特定性质需求的功能分子,广泛应用于物质分离、反应工程、产品工程等领域,是化工过程设计与优化中的共性关键基础。随着深度学习自然语言处理方法的突破性发展,基于文本结构信息的化学语言分子性质预测方法极具前景,但尚未应用于常规分子关键物化性质的系统性建模。全面选择常规分子的25个关键物化性质,包括4个临界性质(临界温度、临界压力、临界体积、偏心因子)、4个基础物性(沸点、熔点、自燃温度、闪点)、5个标准热力学性质(生成焓、生成Gibbs自由能、熔化焓、汽化焓、液体摩尔体积)、5个环境相关性质(半数致死浓度、半数致死量、光化学氧化潜能、生物富集因子、允许接触限值)和7个溶解性关键参数(酸解离常数、水溶性、辛醇-水分配系数、Hildebrandt溶解度参数、3个Hansen溶解度参数),以分子SMILES为模型输入,采用自然语言处理Transformer模型方法进行了严格的分子性质预测建模研究。相较于基于同一数据库开发的现有模型,采用分子SMILES以及Atom-in-SMILES分词方法的深度学习模型在14个预测任务中表现出更优的测试性能,在其余任务中也保持了相当的性能。进一步搭建了基于深度学习模型的ai4solvents性质预测平台,为不同研发场景的分子筛选与设计提供了高效的智能化工具。
Property prediction models can predict key physicochemical,thermodynamic,and environmental health and safety properties of molecules from structural information,thereby enabling forward or inverse screening and design of functional molecules with targeted properties across vast chemical spaces.Such models play a fundamental role in chemical process design and optimization,and are widely applied in separation,reaction,and product engineering.With the rapid advancement of deep learning methods in natural language processing field,chemical language-based property prediction models show great promise,yet have not been systematically applied to modeling conventional molecular physicochemical properties.In this study,25 key physicochemical properties of conventional molecules are comprehensively selected,including 4 critical properties(critical temperature,critical pressure,critical volume,and acentric factor),4 basic physical properties(normal boiling point,melting point,autoignition temperature,and flash point),5 standard thermodynamic properties(enthalpy of formation,Gibbs energy of formation,enthalpy of fusion,enthalpy of vaporization,and liquid molar volume),5 environment-related properties(LC50,LD50,photochemical oxidation potential,bioconcentration factor,and permissible exposure limit),and 7 solubility-related parameters(acid dissociation constant,water solubility,octanol-water partition coefficient,Hildebrandt solubility parameter,and three Hansen solubility parameters).Using the SMILES molecule as the model input,a rigorous molecular property prediction modeling study was conducted using the Transformer modeling method based on natural language processing.Compared with existing models developed on the same database,our models exhibit better test performance across 14 prediction tasks while maintaining comparable performance on the remaining tasks.Furthermore,we establish a property prediction platform based on deep learning models of this work entitled ai4solvents,which provides an efficient tool for molecule screening and design in diverse research scenarios.