Antibiotic AI study puts unseen drug pairs—and its own predictions—to the test
An antibiotic-pair study reports a ranking score of 0.78 for HALO and 0.53 for a variant facing unfamiliar pairs and strains; the models and tests differ.
A study published October 5 presents HALO, a model for ranking antibacterial combinations by whether they reinforce or hinder each other.
Its evaluation keeps entire drug pairs outside training and selects input features using training data alone. That makes a familiar pair harder to mistake for a successful new prediction. [1]
HALO scored 0.78 on ROC-AUC, which measures how well a model separates the two outcomes across possible decision thresholds: 0.5 is chance ranking and 1 is perfect. A permissively evaluated variant scored 0.85; a variant tested on unfamiliar pairs and bacterial strains scored 0.53. Both the variants and testing protocols differ, so this comparison does not isolate the effect of one change. [1]
The useful result is a more demanding benchmark for choosing laboratory experiments. These are retrospective analyses of existing measurements, not prospective demonstrations that newly suggested combinations work. Model ranking, antibacterial activity and successful treatment remain separate questions. [1]
Sources & details 1 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Revised 9 October 2026. Copy edited 9 October to remove commentary about graphic design; reported claims, qualifications and source dates preserved.
Reporting sources
- PLOS Computational Biology · HALO study
Yousefabadi and Mehrmohamadi, October 5. Peer-reviewed article, currently an uncorrected proof. Section 2.2 reports the three scores; sections 2.1–2.3 describe the data and evaluation.
Images & figures
Reported ROC-AUC on one 0–1 scale: permissive M4, 0.85; pair-held-out HALO, 0.78; pair-and-strain-held-out M1, 0.53. The reference at 0.5 means chance ranking. Both models and evaluation protocols differ; these values neither isolate one cause nor measure treatment success.
Euangelion · original explanatory graphic · Source · Usage terms