Beating AlphaFold3 (genesis.ml)

Genesis unveiled Pearl (Placing Every Atom in the Right Location), a generative 3D foundation model that it says is the first to clearly surpass AlphaFold 3 on protein–ligand cofolding — the central "holy grail" problem for structure-based drug discovery. On public benchmarks Pearl improves relative to AlphaFold 3 by ~15% on Runs N’ Poses and up to ~40% on PoseBusters, and it maintains a decisive >40% advantage on internal crystallographic success thresholds. Crucially, Pearl achieves ultra-high accuracy (<1 Å) far more often than competing models, translating to real downstream utility (e.g., potency prediction) rather than just benchmark wins. Technically, Pearl combines several innovations to reach this performance: training on large amounts of physics-generated synthetic complexes to overcome the scarcity of experimental PDB data and demonstrating scaling laws with that synthetic data; an SO(3)-equivariant diffusion head that encodes 3D geometric priors; a compute-efficient trunk architecture; and NVIDIA cuEquivariance kernels (≈15% faster training, up to 80% faster inference in practice). It also offers scientist-in-the-loop controls (templating, inference-time steering) and two modes — unconditional and conditional cofolding — to mirror discovery workflows. Pearl is being deployed internally and through partnerships via the GEMS platform, aiming to shift medicinal chemistry from trial-and-error to faster, hypothesis-driven design.