🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation"

Drug discovery drives diffusion innovation

The most cutting-edge diffusion research now occurs in 3D molecular structure prediction rather than image generation, establishing drug discovery as the new frontier for generative AI primitives.

GANs failed for molecular applications

While generative adversarial networks showed promise for images in 2017-2018, they proved ineffective for proteins and drug discovery, with diffusion models providing the necessary computational primitive for the space.

Three-stage scaling adapted from LLMs

Genesis applies pre-training scaling, post-training refinement, and inference-time computing to molecular models, mirroring the development path of Llama and other language models.

Synthetic data via physics simulations

Unlike proteins, small molecules can be accurately modeled with physics simulations, allowing Genesis to generate unlimited synthetic training data beyond the limited 200,000 structures in public databases.

Inference-time thinking in structural space

Pearl uses 'thinking tokens' adapted to molecular coordinates, allowing the model to iteratively refine crystal structure representations with physics-based guidance during inference.

Cracking the 'key and lock' problem

Pearl predicts 3D coordinates of protein-small molecule complexes with sufficient accuracy to determine binding potency, overcoming a barrier that previously required expensive, months-long laboratory experiments.

Navigating 10^60 molecular possibilities

The model addresses the vast search space of drug-like molecules by using diffusion to move beyond pattern matching toward generalizable physical predictions for previously undruggable targets.