嗅觉在生物体感知外界化学信号方面发挥着关键作用，而气味评估是人们了解生物体嗅觉世界的重要手段。然而，由于评估人员的主观性而导致的气味描述词多样化，使基于计算方法的分子气味属性预测面临巨大的挑战。本研究基于单标签数据，采用 6 种机器学习算法和软投票模型集成策略，针对 5 种高频气味属性，构建了一个多类别的气味属性预测模型。该模型在测试集和外部测试集上的Macro F1 分数均达到 0。7 以上，展现出良好的气味属性预测能力和泛化性能，并且在反直觉结构-气味关系方面具有一定的识别能力，为分子气味属性的有效预测提供了新的可能性。同时，本研究还预测了具有诱蚊效应的化合物分子可能的气味类别，为揭示蚊虫行为与气味嗅觉偏好之间的关系提供了重要线索。
Research on multi-category odor model based on machine learning
Olfaction plays a key role in the perception of external chemical signals by organisms,and odor assessment is an important means for humans to understand the olfactory world of organisms.However,the diversity of odor descriptors caused by evaluator subjectivity presents a significant challenge to the prediction of molecular odor attributes using computational methods.Based on single-label data,this study utilizes six machine learning algorithms and soft voting model integration strategies to construct a multi-category odor attribute prediction model for five high-frequency odor categories.The Macro F1 score of the model on the test set and the external test set are all above 0.7,showing good predictive ability and generalization performance.The model also shows some ability to detect counter-intuitive structure-odor relationship,presenting a new possibility for the effective prediction of molecular odor attributes.Simultaneously,this study also predicted the possible odor categories of molecules with mosquito-trapping effect,providing vital clues for elucidating the relationship between mosquito behavior and odor preference.
国家科技期刊平台
NSTL
万方数据
