Accurate Identification of Key Groups of Microeukaryotes Using Multimodal Deep Learning: An Integrated Classification...

Area of Science:

• Computational Biology and Bioinformatics
• Machine Learning in Taxonomy
• Protistology

Background:

• Traditional flagellate classification relies on subjective morphological traits or single molecular markers, limiting accuracy and data scope.
• Existing methods struggle with the complexity and heterogeneity of biological data, necessitating advanced analytical approaches.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning model for accurate flagellate classification.
• To integrate diverse data types, including photomicrographs and SSU rRNA gene sequences, for enhanced taxonomic resolution.

Main Methods:

• Proposed a Residual Multi-Feature Attention-50 (ResMFA50) deep learning model with dual-branch architecture for feature extraction.
• Employed a Multi-Feature Attention (MFA) mechanism for dynamic fusion of image and genetic data.
• Conducted experiments on a dataset of 296 SSU rRNA sequences and 308 photomicrographs, using 10-fold cross-validation.

Main Results:

• ResMFA50 achieved a classification accuracy of 92.5%, significantly outperforming SVM, Random Forest, EfficientNet, ResNet50, and MMNet.
• Ablation studies confirmed the superiority of the proposed late fusion strategy, improving accuracy by 3.2%-3.8% over early fusion.
• The dual-channel global pooling mechanism enhanced model robustness and performance across varying data scales.

Conclusions:

• ResMFA50 offers a robust and accurate method for flagellate classification by effectively integrating multimodal data.
• This study provides a foundational methodology for applying deep learning to complex multimodal biological datasets in integrative taxonomy.
• The findings advance the application of artificial intelligence in biological classification and data analysis.