Abstract
<title>Abstract</title> <p>Drug-induced autoimmunity (DIA) is a rare, mechanistically heterogeneous adverse drug reaction whose prediction is constrained by small labelled corpora, severe class imbalance and chemically diverse scaffolds. The recently released InterDIA benchmark established the first dedicated machine-learning baseline, in which an EasyEnsembleClassifier trained on a 65-feature genetic-algorithm-selected RDKit descriptor subset reaches an external-test area under the receiver-operating-characteristic curve (AUC) of 0.8930 from a single molecular view, without out-of-distribution (OOD) evaluation or calibrated uncertainty. We propose a heterogeneous six-stream ensemble that fuses classical RDKit descriptors, Morgan ECFP6 fingerprints, two pretrained chemical language models (ChemBERTa-MTR and MoLFormer-XL), a descriptor-transformer hybrid view, and a 7-billion-parameter Qwen-2.5 large language model fine-tuned with QLoRA. A simple mean blend reaches a five-seed external-test AUC of 0.9324 ± 0.0107, exceeding the baseline at every seed. The central finding is robustness rather than the marginal in-distribution gain: on the 68% of held-out drugs whose Bemis-Murcko scaffolds are absent from training, the ensemble degrades by only 0.054 AUC, whereas individual streams degrade by 0.10 to 0.20. We further report the first cross-corpus external evaluation of a DIA model on 427 drug-induced liver injury compounds, split-conformal prediction sets with 91.7% empirical coverage at the 90% target, and transparent SHAP and chain-of-thought interpretability with explicit hallucination disclosure. Representation-level heterogeneity, not algorithmic novelty, yields a predictor that generalises to novel chemistry and reports calibrated confidence. All artefacts, per-seed predictions and code are released for reproduction.</p>