Supervised learning on phylogenetically distributed data.

Area of Science:

• Computational Biology
• Machine Learning
• Evolutionary Biology

Background:

• Developing robust machine learning (ML) models is crucial for biology and medicine.
• Non-independent and identically distributed (non-iid) data, common in evolutionary contexts, pose challenges for ML model generalization.
• Phylogenetic relatedness in data, such as antibiotic resistance across bacterial species, requires specialized ML approaches.

Purpose of the Study:

• To introduce DendroNet, a novel neural network training approach designed for evolutionary data.
• To address out-of-distribution generalization problems in ML models trained on phylogenetically related samples.
• To improve the accuracy and robustness of ML models in biological and medical applications.

Main Methods:

• DendroNet explicitly models the relatedness of training and testing data using phylogenetic trees.
• The approach allows neural network parameters to evolve along the branches of the phylogenetic tree.
• Simulated data and real-world biological datasets were used for evaluation.

Main Results:

• DendroNet significantly outperforms non-phylogenetically aware ML approaches on simulated data.
• The method demonstrates superior performance in predicting antibiotic resistance in bacteria.
• DendroNet also achieves high accuracy in predicting trophic levels in fungi.

Conclusions:

• DendroNet offers a powerful new framework for applying ML to evolutionary biology data.
• Explicitly accounting for phylogenetic relatedness enhances model generalization and predictive accuracy.
• The approach has practical implications for fields like infectious disease research and ecology.