Small-Sample-Size Trait Imputation Using Deep-Learning Techniques.

Area of Science:

• Ecology
• Evolutionary Biology
• Bioinformatics

Background:

• Trait imputation is crucial for ecological and evolutionary studies.
• Small sample sizes and missing data present significant challenges.
• Existing methods like KNN often struggle with overfitting or underfitting.

Purpose of the Study:

• Introduce Dual-Branch BioTraitNet, a deep-learning model for trait imputation.
• Address data sparsity and leverage both quantitative and qualitative trait data.
• Provide a robust and flexible solution for missing trait data.

Main Methods:

• Developed a dual-branch deep-learning architecture.
• Combined unsupervised and supervised learning strategies.
• Applied the model to lizard and fish trait datasets.

Main Results:

• Achieved high R² values on lizard (body length: 0.862, body weight: 0.67) and fish (body length: 0.876, spawning temp: 0.402, egg diameter: 0.496) datasets.
• Demonstrated strong predictive stability and robustness, avoiding negative R² values.
• Maintained high accuracy without phylogenetic information.

Conclusions:

• Dual-Branch BioTraitNet effectively handles trait imputation in small, data-sparse ecological and biological datasets.
• The model generalizes well across diverse taxa and outperforms conventional methods.
• Offers a reliable framework for ecological and evolutionary research where data may be missing or uncertain.