Data Fusion for Integrative Species Identification Using Deep Learning.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Ecology

Background:

• DNA analysis is crucial for species identification but struggles with closely related species.
• Image data offers an alternative but faces similar limitations.
• An integrated approach combining molecular and image data via machine learning shows promise for enhanced species identification.

Purpose of the Study:

• To systematically assess and compare DNA data preprocessing and feature encoding methods.
• To evaluate different strategies for fusing molecular and image data using machine learning.
• To statistically evaluate the performance of integrated data approaches across diverse eukaryotic datasets.

Main Methods:

• Systematic assessment of DNA preprocessing (aligned, unaligned, SNP-reduced) and encoding (fractional, ordinal).
• Utilized artificial neural networks for feature extraction from molecular and image data.
• Investigated three fusion strategies: direct fusion, post-FC layer fusion, and score fusion.
• Evaluated methods using Leave-One-Out Cross-Validation on plant and animal datasets.

Main Results:

• Aligned DNA sequences with decimal vector encoding achieved highest accuracy.
• Direct fusion of molecular and visual features after extraction performed best for most datasets.
• Combined DNA and image data significantly improved accuracy in three out of four datasets.
• Image data integration notably improved species-level resolution, aiding identification of genetically similar species.

Conclusions:

• Optimized preprocessing and integration of molecular and image data significantly enhance species identification.
• The combined approach is particularly beneficial for genetically similar and morphologically indistinguishable species.
• This study provides practical insights for biologists on integrating multi-modal data for improved automated identification.