Does a Genomic Background Proxy Help Relational GNNs Predict Tuberculosis Drug Resistance? An Ablation Study

Graph neural networks (GNNs) that model similarity between clinical isolates can exploit shared genomic context beyond single-SNP effects when predicting *Mycobacterium tuberculosis* (MTB) drug resistance, but it is unclear whether coarse, unsupervised signals about an isolate's broader genomic background add value on top of resistance-associated genotypic similarity alone. We test this with a controlled ablation study on the CRyPTIC compendium of 12,289 global MTB clinical isolates across four drugs (rifampicin, rifabutin, isoniazid, ethambutol). We build a sample graph connecting isolates by resistance-SNP Jaccard similarity, and add a second edge type, a genomic-background-proxy relation derived via unsupervised clustering of genome-wide SNPs, combined through a relational GCN with learned per-relation attention. Across four conditions (full model, structural ablation, runtime masking, and a SNP-co-occurrence reference graph) evaluated over 5 seeds, ROC-AUC ranges from 0.935 to 0.974, and none of 12 paired comparisons against the full model reach significance after correction ($p_{adj} \geq 0.75$).. Masking the background-proxy relation at inference leaves performance unchanged, despite attention weight on it growing during training. These results suggest the background proxy adds no measurable predictive signal beyond resistance-SNP similarity for MTB drug-resistance classification, though this may partly reflect a performance ceiling rather than the proxy carrying no information.