The Second Brain: Diffusion Models for Realistic Human Microbiome Generation

Microbiome synthesis is a useful testbed for generative biology because exact zeros encode ecological and clinical signal, not merely measurement artifacts. We present a diffusion-based generator for human microbiome profiles with a sparsity-preserving zero-inflated decoder, prevalence-aware initialization, and a hard sparsity loss trained with straight-through estimators. On the American Gut Project (4,827 samples, 500 taxa), the full 15.2M-parameter model reaches 1.4\% sparsity deviation in the main comparison and $2.6\%\pm0.5\%$ across three seeds, while achieving the best prevalence correlation among evaluated methods (0.996). Specialized simulators remain stronger on several ecological metrics, so the claim is not distributional indistinguishability. Rather, explicit zero-inflation makes deep generative microbiome models more suitable for downstream biological ML and future agentic discovery workflows.