基于多模态特征融合的可解释性脂质纳米颗粒转染效率预测模型

打开文本图片集

关键词：多模态特征融合；NNConv；MLP；脂质纳米颗粒；SHAP；模型可解释性中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2026）04-011-1054-07doi：10.19734/j. issn.1001-3695.2025.08.0282

Explainable lipid nanoparticles transfection efficiency prediction model based on multimodal feature fusion

Pang Guojun1，²，Lin Bangjiang ，Zheng Bowen²，Chen Jian 2 ，Xu Bohui² （1.Colegeofeldlercaling，AcuedFesrUiesityzo；Q Equipment Manufacturing Research Center，Haixi Instituteof ChineseAcademyof Sciences，Quanzhou Fujian 3622oo，China）

Abstract： In recent years，research on lipid nanoparticles （LNPs）for drug delivery systems usuall focusesona single featureandneglectstheroleofauxiliarylipids.Thisstudyaimedtoconstructanexplainabledeeplearning modelbasedonmultimodalfeaturefusion，named MolGraphNet，foraccuatepredictionandinterpretabilityanalysisofauxiliarylipidtransfection eficiency.MolGraphNetused neuralnetwork convolution（NNConv）toextract molecularstructural graph featuresandapplied a multilayer perceptron（MLP）to extractphysicochemical numericalfeatures.Amultimodal fusion module integrated structuraland property informationatadeeplevel.The modelalsocombined molecularstructure encoding withSHAPanalysis toperforminterpretabilityandvisualizationofkeyfeatures，revealingtheintrinsicrelationshipbetwenauxiliarylipidstructuresand transfection eficiency.Resultsonsixcel-typetasksfrompublicdatasetsshowthatMolGraphNetachieved predictioneorsof 6%～11% and Pearson and Spearman correlation coefficients between O.75andO.92，which outperformed baseline models. MolGraphNetprovideshighpredictionaccuracyandrobustnessandofersinterpretablevisualizationofkeyfeaturesthrough SHAP-based analysis，giving valuable guidance for auxiliary lipid design and LNP formulation optimization.

Keywords：multimodal feature fusion；neural networkconvolution；multi-layerperceptron；lipid nanoparticles；shapelyad ditive explanations；model interpretability

0 引言

随着基因治疗和疫苗传递在现代医学领域中取得革命性突破，优化并加速开发高效、低成本的脂质载药系统逐渐成为研究热点。（剩余21703字）