Multi-omic machine learning overcomes transcriptomic blind spots to rescue therapeutic targets in ALS

<title>Abstract</title> <p> Amyotrophic lateral sclerosis (ALS) lacks effective therapies partly because transcriptomic screens cannot distinguish disease-relevant genes from downstream perturbations. Here we integrate a cross-tissue meta-analysis of 13 datasets (1,227 samples; CNS, myeloid, lymphoid, and mesenchymal domains) with an explainable machine learning framework trained on evolutionary constraint, brain expression, interaction-network topology, Mendelian disease annotations, and 12 engineered cross-database interaction terms. SHAP-based decomposition of an ensemble classifier reveals that ALS gene identity is defined by the convergence of neurological disease history, evolutionary constraint, brain-enriched expression, and network centrality, not differential expression magnitude. In silico drug-reversal simulations formalize this as a model-derived hypothesis, the Genomic Architecture Barrier, demonstrating that within the model’s feature space pharmacological correction of modifiable features alone is insufficient to shift predicted disease relevance beyond classification threshold. A nine-axis multi-evidence evaluation framework identifies eleven consensus prioritized candidates, led by <italic>ATP1A1</italic> , <italic>YWHAG</italic> , and <italic>ANLN</italic> , that share the genomic signature of established ALS associated genes and interact physically with ALS proteins across autophagy-proteostasis, RNA-processing, cytoskeletal, and innate-immunity axes. Instead of cataloging expression changes, this framework identifies new therapeutic targets by prioritizing genes whose genomic architecture mirrors that of known ALS-associated genes. </p>