X-Admix: An Interpretable Multimodal Cross-Attention Framework for Integrating Genotype, Local Ancestry, and Social Drivers of Health in Admixed African American Populations

Disease risk in admixed human populations is shaped by interactions among genotype, locus-specific ancestry, and the social environment, but predictive frameworks rarely model these three modalities jointly. We introduce X-Admix, an interpretable multimodal framework integrating genotype, local ancestry, and social drivers of health through structured pairwise cross-attention streams, softmax-gated fusion, and a Random Forest classifier. Unlike concatenation-based fusion, these directed streams learn conditional representations in which genotype is contextualized by local ancestry and social drivers of health, and local ancestry by social drivers of health. Leave-one-stream-out ablation decomposes predictive performance and top cross-modal pair candidates. Applied to 240 African American children with severe asthma from the BIG dataset, X-Admix predicted inhaled-corticosteroid response with mean area under the receiver operating characteristic curve 0.771 {+/-} 0.072 across 10-fold cross-validation, whereas ridge logistic baselines performed near chance. This performance pattern replicated in 666 African American adults from the All of Us dataset under matched inclusion criteria. On the BIG dataset, a two-tier consensus pipeline yielded 932 top cross-modal feature pairs whose stream dependencies decomposed into stream-independent (14.7), single-stream-conditional (28.7), multi-stream-dependent (33.5), and all-stream-dependent (23.2). Without the genotype-local-ancestry stream, performance is unchanged, yet the top cross-modal pairs identified change, indicating that a stream's contribution to interpretation and to prediction are separable: a stream redundant for prediction can still define the candidate interactions carried forward for discovery. To our knowledge, X-Admix is the first framework to jointly model the three data modalities via cross-attention in admixed cohorts, yielding an interpretable catalog of candidate interactions underlying inhaled-corticosteroid non-response.