A doctor doesn't make a diagnosis on a whim. They open a medical reference book and point to the article they're relying on. That's roughly how the new system KG-TRACE works. It analyzes the genome of the tuberculosis bacterium and cross-checks the found mutations against a knowledge map from the World Health Organization — a sort of atlas of known genetic glitches that make the microbe invulnerable to drugs.

An ordinary neural network can find a statistical link between a DNA segment and antibiotic resistance. But it can't say why that link works. KG-TRACE is built differently. Inside it, there's an "epistemic trust gate." In each ambiguous situation, it weighs whether to rely more on the network's "intuition" or on the hard facts from the WHO database. This approach doesn't chase record accuracy. For the antibiotic isoniazid, it's already high (AUROC 0.9760). But the system offers something more important — transparency.

The system learns not only to detect resistance but also to recognize "gray zones." If data are insufficient, KG-TRACE doesn't guess. It honestly labels the case as "uncertain" and suggests the doctor send the sample for additional lab testing. This is especially valuable for detecting microbes with multidrug resistance, where a mistake can cost lives. Such an "audit trail" turns the neural network's black box into a glass one. It brings clinical trust back where it's needed most — at the patient's bedside.