Accurate Identification of Key Groups of Microeukaryotes Using Multimodal Deep Learning: An Integrated Classification Model Combining Morphological and Molecular Data.

Traditional classification of flagellates (i.e., flagellated protists) relies on morphological traits or a single molecular marker, which suffer from subjectivity and limited data sources. This study proposes a multimodal deep learning model, Residual Multi-Feature Attention-50 (ResMFA50), that integrates photomicrographs and small subunit ribosomal RNA (SSU rRNA) gene sequences of flagellates. The dual-branch architecture (DNA sequence and image branches) extracts local and global features, while the Multi-Feature Attention (MFA) mechanism dynamically fuses heterogeneous data. Experiments were conducted on a dataset comprising 296 SSU rRNA gene sequences and 308 standardized photomicrographs, evaluated using 10-fold cross-validation. The results demonstrate that ResMFA50 achieves an accuracy of 92.5% in classifying flagellates at a batch size of eight, which is significantly higher than the accuracies achieved by SVM (82.4%), Random Forest (84.2%), EfficientNet (83.6%), ResNet50 (87.4%), and MMNet (91.3%). Moreover, ablation experiments comparing early, intermediate, and late fusion strategies demonstrate that the proposed late fusion scheme consistently outperforms other fusion timings, achieving improvements of 3.2%-3.8% over early fusion across different batch sizes. This study establishes a methodological foundation for modelling multimodal biological data in complex systems, advancing deep learning applications in integrative taxonomy. This advantage is attributed to the dual-channel global pooling mechanism (Global Average/Max Pooling fusion), which balances the variance-bias trade-off through complementary strategies of spatial statistical smoothing and local salient feature detection, enhancing robustness to data scale expansion.