Minerva: alignment-free genomic interaction mapping

Minerva is a genome language model framework for coevolutionary mining that achieves alignment-free prediction of ncRNA base-pairing and protein contacts directly from sequence. By leveraging categorical Jacobian fingerprinting and interaction heads to generate 2D structural maps, Minerva bypasses traditional homology constraints and can be applied efficiently at scale. Applied to 150 bacterial genomes, it revealed that 84.3% of intergenic base-pairing is unannotated, while characterizing structural extensions in Pseudomonas ncRNAs and sequence-diverse arrays in UG27 prophages. This sequence-direct architecture provides researchers with a scalable, machine-learning-driven workflow to annotate highly divergent regulatory networks and molecular machines across large genomic datasets without relying on sequence alignments.