Hierarchical Disease-State Generators for Neurodegenerative Genomics: A Benchmark Proposal for Intervention-Conditioned Multi-omic Generation

We present a benchmark proposal and evaluation framework for the disease-state generator task: intervention-conditioned generation of transcriptomic and epigenomic cell states, evaluated through mechanism-grounded acceptance criteria rather than generic sample-quality scores. Targeting neurodegeneration (AD/PD) as a biologically demanding test bed, we define (i) a formal task specification for conditional generation under drugs, CRISPR perturbations, and regulatory edits; (ii) an architecture blueprint-multimodal latent encoders coupled to conditional diffusion with hierarchical regulatory priors (enhancer TF gene) ; and (iii) a barrier-and-frontier evaluation suite testing hierarchy fidelity, perturbation prediction, cross-context generalization, and uncertainty-calibrated intervention ranking. The framework also serves as an evaluation surface for DNA foundation models, measuring whether sequence-derived priors improve intervention-conditioned generation. We report proof-of-concept experiments on the Norman 2019 CRISPRa dataset that validate the evaluation protocols, while identifying a key bottleneck- gene-regulatory-network sparsity-that must be resolved before hierarchy-fidelity testing is meaningful. This is a benchmark and evaluation contribution; the architecture is a proposed blueprint, not a fully validated system.