HOCs-BDPred：基于多模态Transformer的卤代有机污染物可生物降解性预测平台

HOCs-BDPred: a multimodal Transformer-based platform for predicting the biodegradability of halogenated organic contaminants

摘要:

卤代有机污染物(halogenated organic contaminants, HOCs)具有高环境持久性与生物累积性，其降解机制与消除路径是环境科学领域研究的焦点。针对实验测定通量低、降解样本稀缺，以及传统预测模型难以处理分子结构复杂性和极度不平衡数据等挑战，本研究构建了一个包含614个HOCs样本的高质量基准数据集(47个易降解，567个难降解)，并开发了基于多模态Transformer架构的生物降解性预测平台——HOCs-BDPred。该平台核心采用双流并行的Transformer架构，实现了简化分子线性输入规范(simplified molecular input line entry system, SMILES)序列的语义特征与分子指纹的高维拓扑信息的深度协同表征。为突破小样本不平衡数据的限制，研究引入了SMILES数据增强技术焦点损失函数(focal loss)，显著增强了模型对少数类(易降解样本)的识别灵敏度与泛化边界。实验结果显示，模型总体准确率达93.48%，平衡准确率为76.99%，马修斯相关系数(Matthews correlation coefficient, MCC)为0.548。此外，同步开发并部署了在线网络预测平台(https://hocs-bdpred.streamlit.app/)并进行了应用实践验证，为HOCs的生物降解性快速筛查及绿色替代品设计开发提供了智能化、高通量的决策支持工具。

Abstract:

Due to their high environmental persistence and bioaccumulation, the degradation mechanisms and elimination pathways of halogenated organic contaminants (HOCs) have become a focal point of research in environmental science. To address challenges such as low experimental throughput, the scarcity of degradation samples, and the limitations of conventional prediction models in handling molecular complexity and highly imbalanced data, we constructed a high-quality benchmark dataset comprising 614 HOC samples (47 readily biodegradable and 567 non-readily biodegradable) and developed a biodegradability prediction platform, HOCs-BDPred, based on a multi-modal Transformer architecture. This platform utilizes a dual-stream parallel Transformer framework to achieve deep synergistic representation of semantic features from SMILES sequences and high-dimensional topological information from molecular fingerprints. To overcome the constraints imposed by small-sample imbalanced data, we introduced SMILES data augmentation techniques and a Focal Loss weighted function, significantly enhancing the model’s recognition sensitivity and generalization boundaries for the minority class. Experimental results demonstrated that the model achieved the overall accuracy of 93.48%, the balanced accuracy of 76.99%, and the Matthews correlation coefficient (MCC) of 0.548. Furthermore, an online Web prediction platform (https://hocs-bdpred.streamlit.app/) was developed and deployed for practical validation, providing an intelligent, high-throughput decision-support tool for the rapid screening of HOC biodegradability and the proactive design of green alternative chemicals.