BIG-TB: A benchmark for evaluating prediction and interpretability of sequence-based machine learning using Mycobacterium tuberculosis genomes

BIG-TB, built from over 17,000 Mycobacterium tuberculosis genomes with measured antibiotic resistance, shows DNA-based foundation models (mean test AUC 0.888) do not yet beat simple ML baselines (0.865).

• Why it matters: Existing foundation-model benchmarks often separate prediction from variant attribution, leaving uncertainty about whether learned sequence representations are both accurate and biologically interpretable.
• What they did: The benchmark defines two tasks—predicting resistance phenotypes and attributing predictions to WHO-catalogued resistance-associated variants—then evaluates DNA- and protein-based foundation models on lineage-informed splits.
• The result: Predictive signal for well-characterized drugs is largely preserved across DNA and protein modalities (mean AUC 0.846–0.963), but top-performing models do not necessarily recover resistance variants best.

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