基于预训练和微调策略的多味道分子预测模型

doi:10.16085/j.issn.1000-6613.2025-1064

摘要/Abstract

摘要：

味觉感知分析在食品科学研究中起着至关重要的作用，它直接影响食品消费、人体营养和健康。传统的风味感官分析方法由于主观性强且研发周期长，难以满足现代食品科学对风味分子快速筛选和优化的需求。因此，本文通过整合多个开源数据库，构建了一个包含17633个味道分子的标准化数据集，涵盖甜、苦、鲜、酸和其他少见味型（如咸味、辛辣、涩味、麻味等）5类味道分子。通过结合Uni-Mol2预训练模型与微调策略，构建了高精度的分子多味道预测模型，并利用积分梯度和原子贡献分析对模型的机制进行解释性分析。结果表明，模型在甜、苦、鲜、酸和其他少见味型上均实现了准确预测，准确率达到95.2%，验证了多数据库融合与精细化微调策略在味觉分子预测中的有效性，为风味分子的快速筛选与功能解析提供了新的技术途径。

关键词: 风味分子预测, 预训练, 多分类, 机器学习, 微调策略

Abstract:

Taste perception analysis plays a vital role in food science, directly affecting food consumption, nutrition, and health. Traditional sensory evaluation methods for flavor are highly subjective and time-consuming, making them inadequate for the rapid screening and optimization of flavor molecules. To overcome this limitation, we integrated multiple open-source databases and constructed a standardized dataset comprising 17633 taste-related molecules, covering five categories: sweet, bitter, umami, sour, and other less common tastes (e.g., salty, spicy, astringent, numbing, etc.). Based on this dataset, we developed a high-accuracy multi-flavor prediction model by combining the Uni-Mol2 pre-trained molecular model with fine-tuning strategies. Furthermore, we employed integrated gradients and atomic contribution analysis to interpret the predictive mechanisms of the model. Experimental results showed that the proposed model achieved accurate predictions across all taste categories, with an overall accuracy of 95.2%, thereby validating the effectiveness of multi-database integration and fine-tuning strategies for taste molecule prediction and providing a new technical pathway for the rapid screening and functional analysis of flavor molecules.

Key words: flavor molecule prediction, pre-training, multi-classification, machine learning, fine-tuning strategies

中图分类号:

TQ656⁺.1

宋英杰, 张磊, 都健. 基于预训练和微调策略的多味道分子预测模型[J]. 化工进展, 2025, 44(S1): 29-37.

SONG Yingjie, ZHANG Lei, DU Jian. Multi-flavor molecule prediction model based on pre-training and fine-tuning strategies[J]. Chemical Industry and Engineering Progress, 2025, 44(S1): 29-37.

图/表 13

图1 基于Uni-Mol2的预训练和微调策略的多味道分子预测模型训练流程

表1 数据集来源汇总

数据集	味道类别	分子数
FlavorDB^[25]	甜、苦、酸、未定义	25595
BitterDB^[26]	苦、非苦	2250
UMP442^[27]	鲜、非鲜	140
ChemTastesDB^[28]	甜、苦、酸、鲜、其他少见味型	2944
SweetnerDB^[29]	甜	316
BitterSweet^[17]	甜、苦	1202
PlantMolecularTasteDB^[30]	甜、苦、酸、鲜、其他少见味型	1527
VirtuousMultiTaste^[14]	甜、苦、鲜、其他少见味型	4717
国际纯粹与应用化学联合会（IUPAC）^[15]	酸	1513

表2 典型分子在多数据库中的风味标注及归一化示例

分子SMILES	FlavorDB	ChemTastesDB	VirtuousMultiTast	本研究
[Ca+2].[Cl-].[Cl-]	未定义	苦、咸	苦	苦、其他
NC(Cc1ccccc1)C( $̿$ O)O	苦、甜、无味	轻微苦、无味	甜	苦、甜、其他
O $̿$ C(O)C(O)C(O)C( $̿$ O)O	微酸	酸、鲜	其他（非甜苦鲜）	酸、鲜
O $̿$ CO	辛辣、酸	酸	甜	甜、酸、其他
CCCCCCCC( $̿$ O)CCC 1 $̿$ CC[ $̿$ C(C $̿$ C1)O]OC	辛辣	辛辣、麻	甜	甜、其他

表2 典型分子在多数据库中的风味标注及归一化示例

分子SMILES	FlavorDB	ChemTastesDB	VirtuousMultiTast	本研究
[Ca+2].[Cl-].[Cl-]	未定义	苦、咸	苦	苦、其他
NC(Cc1ccccc1)C( $̿$ O)O	苦、甜、无味	轻微苦、无味	甜	苦、甜、其他
O $̿$ C(O)C(O)C(O)C( $̿$ O)O	微酸	酸、鲜	其他（非甜苦鲜）	酸、鲜
O $̿$ CO	辛辣、酸	酸	甜	甜、酸、其他
CCCCCCCC( $̿$ O)CCC 1 $̿$ CC[ $̿$ C(C $̿$ C1)O]OC	辛辣	辛辣、麻	甜	甜、其他

图2 不同风味分子数据分布与正负样本数量对比

图3 Uni-Mol2预训练框架的整体结构、原子与对偶表示机制及主干网络结构[24]

图4 Uni-Mol2-84m预训练模型微调流程及评估

表3 不同模型在测试集上的预测性能指标对比

模型	准确率	接受者操作特征曲线下面积（AUROC）	平均精度-召回率曲线下的面积（AUPRC）	F₁	召回率
VirtuousMultiTaste^[14]	0.818	0.870	—	0.658	0.713
风味分析和识别（FART）^[15]	0.884	0.969	—	0.771	0.785
Uni-Mol2^[24]	0.945	0.974	0.901	0.831	0.831
Uni-Mol2+微调	0.952	0.975	0.904	0.840	0.833

图5 Uni-Mol2与Uni-Mol2+微调模型性能对比

图6 五类味道分子分类测试集ROC曲线与精确率-召回率曲线

图7 积分梯度法与余弦相似性计算原子贡献示意图（K为超参数，对零向量至分子级嵌入向量间的线性路径进行离散采样的点）

图8 积分梯度法计算风味分子原子贡献度示例

图9 味道分子官能团对比分析

图10 Mol Tastants分子多味道属性预测Web应用界面

参考文献 31

[1]	QUEIROZ L P, NOGUEIRA I B R, RIBEIRO A M. Flavor Engineering: A comprehensive review of biological foundations, AI integration, industrial development, and socio-cultural dynamics[J]. Food Research International, 2024, 196: 115100.
[2]	SUNG Jeehye, FROST Scott, Joon Hyuk SUH. Progress in flavor research in food: Flavor chemistry in food quality, safety, and sensory properties[J]. Food Chemistry: X, 2025, 25: 102071.
[3]	KEMP Sarah E, COMMITTEE IFST PFSG. Application of sensory evaluation in food research[J]. International Journal of Food Science & Technology, 2008, 43(9): 1507-1511.
[4]	KIANI Sajad, MINAEI Saeid, Mahdi GHASEMI-VARNAMKHASTI. Fusion of artificial senses as a robust approach to food quality assessment[J]. Journal of Food Engineering, 2016, 171: 230-239.
[5]	DI ROSA Ambra Rita, LEONE Francesco, CHELI Federica, et al. Fusion of electronic nose, electronic tongue and computer vision for animal source food authentication and quality assessment—A review[J]. Journal of Food Engineering, 2017, 210: 62-75.
[6]	毛海涛. 混合机器学习与机理建模的香精产品设计[D]. 大连: 大连理工大学, 2020.
	MAO Haitao. Fragrance product design with a hybrid machine learning and mechanistic modeling approach[D]. Dalian: Dalian University of Technology, 2020.
[7]	ZENG Xiangquan, CAO Rui, XI Yu, et al. Food flavor analysis 4.0: A cross-domain application of machine learning[J]. Trends in Food Science & Technology, 2023, 138: 116-125.
[8]	GOEL Mansi, BAGLER Ganesh. Computational gastronomy: A data science approach to food[J]. Journal of Biosciences, 2022, 47(1): 12.
[9]	YANG Xin, Chi-Tang HO, GAO Xiaoyu, et al. Machine learning: An effective tool for monitoring and ensuring food safety, quality, and nutrition[J]. Food Chemistry, 2025, 477: 143391.
[10]	QUEIROZ Luana P, REBELLO Carine M, COSTA Erbet A, et al. Transfer learning approach to develop natural molecules with specific flavor requirements[J]. Industrial & Engineering Chemistry Research, 2023, 62(23): 9062-9076.
[11]	ROJAS Cristian, BALLABIO Davide, CONSONNI Viviana, et al. Quantitative structure-activity relationships to predict sweet and non-sweet tastes[J]. Theoretical Chemistry Accounts, 2016, 135(3): 66.
[12]	BO Weichen, QIN Dongya, ZHENG Xin, et al. Prediction of bitterant and sweetener using structure-taste relationship models based on an artificial neural network[J]. Food Research International, 2022, 153: 110974.
[13]	DUTTA Prantar, JAIN Deepak, GUPTA Rakesh, et al. Classification of tastants: A deep learning based approach[J]. Molecular Informatics, 2023, 42(12): e202300146.
[14]	ANDROUTSOS Lampros, PALLANTE Lorenzo, BOMPOTAS Agorakis, et al. Predicting multiple taste sensations with a multiobjective machine learning method[J]. NPJ Science of Food, 2024, 8: 47.
[15]	ZIMMERMANN Yoel, SIEBEN Leif, SENG Henrik, et al. A chemical language model for molecular taste prediction[J]. NPJ Science of Food, 2025, 9(1): 122.
[16]	KOU Xingran, SHI Peiqin, GAO Chukun, et al. Data-driven elucidation of flavor chemistry[J]. Journal of Agricultural and Food Chemistry, 2023, 71(18): 6789-6802.
[17]	TUWANI Rudraksh, WADHWA Somin, BAGLER Ganesh. BitterSweet: Building machine learning models for predicting the bitter and sweet taste of small molecules[J]. Scientific Reports, 2019, 9(1): 7155.
[18]	ROJAS Cristian, BALLABIO Davide, CONSONNI Viviana, et al. Classification-based machine learning approaches to predict the taste of molecules: A review[J]. Food Research International, 2023, 171: 113036.
[19]	MALAVOLTA Marta, PALLANTE Lorenzo, MAVKOV Bojan, et al. A survey on computational taste predictors[J]. European Food Research and Technology, 2022, 248(9): 2215-2235.
[20]	DUTTA Prantar, GAJULA Kishore, VERMA Nitu, et al. Computational screening of umami tastants using deep learning[J]. Molecular Diversity, 2025, 29(4): 2979-2993.
[21]	BANERJEE Priyanka, PREISSNER Robert. BitterSweetForest: A random forest based binary classifier to predict bitterness and sweetness of chemical compounds[J]. Frontiers in Chemistry, 2018, 6: 93.
[22]	SONG Renxiu, LIU Kaifeng, HE Qizheng, et al. Exploring bitter and sweet: The application of large language models in molecular taste prediction[J]. Journal of Chemical Information and Modeling, 2024, 64(10): 4102-4111.
[23]	QUEIROZ Luana P, REBELLO Carine M, COSTA Erbet A, et al. Generating flavor molecules using scientific machine learning[J]. ACS Omega, 2023, 8(12): 10875-10887.
[24]	JI Xiaohong, WANG Zhen, GAO Zhifeng, et al. Uni-Mol2: Exploring molecular pretraining model at scale[EB/OL]. (2024-07-01)[2025-07-24]. .
[25]	GOEL Mansi, GROVER Nishant, BATRA Devansh, et al. FlavorDB2: An updated database of flavor molecules[J]. Journal of Food Science, 2024, 89(11): 7076-7082.
[26]	ZIAIKIN Evgenii, DAVID Moran, USPENSKAYA Sofya, et al. BitterDB: 2024 update on bitter ligands and taste receptors[J]. Nucleic Acids Research, 2025, 53(D1): D1645-D1650.
[27]	CHAROENKWAN Phasit, YANA Janchai, NANTASENAMAT Chanin, et al. iUmami-SCM: A novel sequence-based predictor for prediction and analysis of umami peptides using a scoring card method with propensity scores of dipeptides[J]. Journal of Chemical Information and Modeling, 2020, 60(12): 6666-6678.
[28]	ROJAS Cristian, BALLABIO Davide, PACHECO SARMIENTO Karen, et al. ChemTastesDB: A curated database of molecular tastants[J]. Food Chemistry: Molecular Sciences, 2022, 4: 100090.
[29]	Jean-Baptiste CHÉRON, CASCIUC Iuri, GOLEBIOWSKI Jérôme, et al. Sweetness prediction of natural compounds[J]. Food Chemistry, 2017, 221: 1421-1425.
[30]	GRADINARU Teodora-Cristiana, PETRAN Madalina, DRAGOS Dorin, et al. PlantMolecularTasteDB: A database of taste active phytochemicals[J]. Frontiers in Pharmacology, 2022, 12: 751712.
[31]	SUNDARARAJAN Mukund, TALY Ankur, YAN Qiqi. Axiomatic attribution for deep networks[EB/OL]. (2017-06-13)[2025-07-24]. .