Degree bias predicts, outperforms and improves machine learning phage-host interaction predictions

An interaction-naive model using only phage and bacterial network degrees achieves higher global AUROC than five published ML models across ten phage-host datasets, indicating AUROC is dominated by network structure.

• Why it matters: Global AUROC can be inflated by degree bias in structured phage-bacterium networks, risking overestimation of how well ML predicts true interactions for phage therapy prioritization.
• What they did: The authors compared an interaction-naive degree-based predictor against five published ML models on ten datasets, then used randomized degree-conserved networks and degree-stratified evaluation to separate true predictions from baseline infection likelihood.
• The result: Regressing out degree bias from training data improved both global and degree-stratified interaction predictability, showing a practical way to make ML-guided phage selection more reliable.

AI summary of the abstract · may contain errors · not medical advice · read the paper.