Integrating Deep Learning Derived Morphological Traits and Molecular Data for Total-Evidence Phylogenetics: Lessons...

Area of Science:

• Evolutionary Biology
• Computational Biology
• Bioinformatics

Background:

• Deep learning (DL) models have demonstrated success in automatically generating morphological traits with phylogenetic significance.
• Integrating diverse data types is crucial for robust phylogenetic inference.

Purpose of the Study:

• To explore the combination of molecular data with DL-derived morphological traits from insect images for total-evidence phylogenetics.
• To identify challenges and optimize methods for integrating DL-generated traits into phylogenetic analyses.

Main Methods:

• Utilized a dataset of rove beetle images to train DL models for morphological trait extraction.
• Compared the performance of DL-derived traits alone versus in combination with molecular data (total-evidence phylogenetics).
• Evaluated the impact of different dataset splits (e.g., cladistic) and deep metric loss functions (e.g., contrastive loss).

Main Results:

• DL-derived morphological traits, while informative, were less effective in isolation than molecular data for phylogenetics.
• Incorporating DL-derived traits into total-evidence analyses improved phylogenetic resolution compared to molecular data alone.
• A cladistic dataset split and contrastive loss function showed a slight preference in performance.
• The optimal combination of genes for phylogenetic inference varied depending on whether genes were analyzed individually or in a total-evidence framework.

Conclusions:

• Combining deep learning-derived morphological traits with molecular data offers a promising approach for total-evidence phylogenetics.
• Challenges remain in maximizing the phylogenetic signal from DL-extracted traits and managing resource-intensive data acquisition.
• Future research should focus on enhancing trait extraction methods and developing disentangled networks for better trait interpretability.