Why Diffusion’s Biggest Wins Are in Drug Design, Not Images

Most innovative diffusion research today is happening in 3D structure prediction for drug discovery, not image generation. Genesis Molecular AI claims its Pearl model achieves sub-angstrom accuracy on protein–ligand complexes—a threshold where predictions become useful for real drug design. This mirrors the moment LLMs crossed from interesting to indispensable.

The team applies LLM scaling laws: pre-training on synthetic data generated via physics simulations, post-training fine-tuning, and inference-time scaling through iterative diffusion with physics feedback. They explicitly target the metric of pose RMSD below one angstrom, which ensures hydrogen bonds are correctly positioned, avoiding plausible but wrong structures.

Drug discovery has resisted ML due to scarce data and resolution demands. Gilead and Insitro now use these models to find first-in-class binders and optimize candidates. Yet pose prediction is only one piece—separate models for ADMET and safety are still required.

Engineering leaders should note the pattern: a narrow domain with extreme precision requirements is yielding to AI after years of progress on data engineering, metrics, and physics-hybrid architectures. The lesson is that crossing the utility threshold demands obsessive measurement and the right inductive biases, not just more data or compute.